Method for predicting fatigue life of windproof stay wire device of power transmission tower

Abstract

The invention discloses a method for predicting the fatigue life of a windproof stay wire device of a power transmission tower, and relates to the technical field of power system equipment state monitoring and intelligent operation and maintaining.The method comprises the steps that firstly, a windproof stay wire device three-dimensional digital twinborn model comprising a modular connecting block, a metal shell and a damping element is constructed; a physical process agent model based on a multi-scale layered framework is adopted to execute wind-induced vibration response simulation, multi-scale information processing is achieved through independent channel embedding, a channel axial self-attention mechanism and a shifting window self-attention mechanism, and stress-strain time history data are output in an autoregressive mode; performing fatigue damage evaluation by adopting a long sequence analysis model based on input dependency state transition, performing explicit modeling on autocorrelation of a stress sequence through a liquid convolution kernel, and capturing high-order correlation and a combination effect; performing internal biochemical treatment on the environmental parameters into the model through a self-adaptive layer normalization mechanism; high-precision and high-efficiency fatigue life prediction is realized, and technical support is provided for intelligent operation and maintenance of power transmission equipment.

Description

Technical Field

[0001] This invention relates to the field of power system equipment condition monitoring and intelligent operation and maintenance technology, specifically a method for predicting the fatigue life of windproof guy wire devices for transmission towers. Background Technology

[0002] Wind-resistant guy wires on transmission towers are critical components of power systems, primarily used to reduce the vibration amplitude of transmission lines under strong winds, preventing conductor fatigue damage and excessive insulator string sway. Under prolonged wind loads, these guy wires themselves can experience cumulative fatigue damage, potentially leading to failure and seriously threatening the safe and stable operation of the power grid. Therefore, accurately predicting the fatigue life of wind-resistant guy wires is crucial for developing scientific maintenance strategies and ensuring the reliability of the power system.

[0003] Currently, fatigue life prediction for windbreak guy wire devices mainly relies on traditional numerical simulation and empirical formula methods. Existing technologies establish simplified mechanical models, calculate loads using constant wind pressure or average wind speed, and combine SN curves and rainflow counting methods for fatigue analysis. However, these methods have certain limitations, such as: failing to accurately reflect the pulsating characteristics of wind and complex dynamic effects like vortex-induced vibration, leading to distorted stress calculations; overly idealized fatigue models that do not consider the deterioration effects of corrosion, temperature changes, and ultraviolet radiation in actual operating environments on material properties; lack of a data closed-loop mechanism, making it impossible to use on-site monitoring data to correct the prediction model in real time, resulting in increasingly large cumulative errors in the prediction results; the traditional rainflow counting method only focuses on the amplitude and mean of stress cycles, ignoring the influence of loading history and loading sequence on fatigue damage; and low computational efficiency, with traditional fluid-structure interaction simulations requiring hours or even days, failing to meet the needs of real-time monitoring and early warning.

[0004] While existing machine learning-based fatigue prediction methods have improved prediction accuracy to some extent, they generally suffer from poor model generalization ability, inability to handle multi-physics coupling, and difficulty in capturing long-term dependencies. In particular, under complex wind-induced vibration conditions, traditional methods struggle to model the nonlinear coupling relationship between the wind field and the structure, and are also unable to effectively handle higher-order correlations and combined effects in stress sequences.

[0005] Therefore, there is an urgent need for a method to predict the fatigue life of wind-induced vibration devices that can accurately model complex wind-induced vibration phenomena, consider the influence of environmental factors, have online learning capabilities, and have high computational efficiency, so as to meet the technical requirements of smart grid construction for equipment condition monitoring and predictive maintenance. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and propose a method for predicting the fatigue life of wind-resistant guy wire devices for power transmission towers, so as to solve the above-mentioned problems.

[0007] The objective of this invention is achieved through the following technical solution: a method for predicting the fatigue life of wind-resistant guy wire devices for power transmission towers, comprising the following steps: S1. Construct a three-dimensional digital twin model of the windproof guy wire device and the associated transmission tower. The windproof guy wire device includes modular connecting blocks, a metal shell, and damping elements. Discretize the model and the surrounding flow field space into a three-dimensional mesh. Define multiple physical channels on the mesh to characterize the physical states of the fluid and solid. The physical channels include the x and y components of the wind speed field, the wind pressure field, the σx, σy, and τxy components of the structural stress tensor field, and the structural displacement field. S2. Perform wind-induced vibration response simulation using a physical process surrogate model based on a multi-scale hierarchical architecture to obtain stress-strain time history data. This step specifically includes: S2.1 Input the multiple physical channels defined in step S1 into the physical process proxy model as independent channels, and embed each physical channel independently into different representation vectors on the spatial grid; S2.2 The coupling relationship between different physical channels is modeled by the channel axial self-attention mechanism inside the physical process proxy model to simulate the fluid-structure interaction effect. The channel axial self-attention mechanism learns the interaction between channels by calculating the attention weights between the representation vectors of different physical channels. S2.3 adopts a U-Net-type hierarchical architecture and a shift window self-attention mechanism to perform feature extraction and fusion at multiple scales. Multi-scale information processing is achieved by upsampling and downsampling the representation vector. S2.4 predicts time-step by time using an autoregressive method and outputs stress-strain time history data of key fatigue points of the windproof guy wire device within a preset time period; S3. A long-sequence analysis model based on input-dependent state transitions is used to assess fatigue damage from stress-strain time history data. This step specifically includes: S3.1 Input the stress-strain time history data output in step S2 into the long-sequence analysis model as the input sequence; S3.2 By using the input-dependent state transition mechanism built into the long sequence analysis model, the state transition matrix is ​​dynamically adjusted according to the current input signal to achieve adaptive adjustment of the state transition coefficients as the input changes. The state transition matrix is ​​in the form of (A+Bu), where A is the basic state transition matrix, B is the learnable weight matrix, and u is the current input signal. S3.3 Based on the input-dependent state transition mechanism, the liquid convolution kernel is used to explicitly model the autocorrelation of the input stress sequence. The weight of the liquid convolution kernel is associated with the autocorrelation term of the input signal, which is used to capture higher-order correlations and combined effects in the stress sequence. S3.4 directly learns the mapping relationship from stress-strain sequence to fatigue damage increment, and outputs the cumulative fatigue damage value associated with time history data; S4. Environmental parameters are endogenized into the physical process proxy model and long sequence analysis model through a deep condition injection mechanism. The environmental parameters include corrosion level, ambient temperature and ultraviolet radiation intensity. The environmental parameters are used as condition inputs to adjust the internal calculation of the model through an adaptive layer normalization mechanism, so that the model can learn the nonlinear effects of the environment on the physical process and damage accumulation. S5. Based on the measured data collected by on-site sensors, the pre-trained model from steps S2 and S3 is fine-tuned online using low-rank adaptive technology, specifically including: S5.1 Vibration and stress data are collected in real time by sensors installed on the modular connecting blocks of the windproof guy wire device; S5.2 Compare the measured data with the predicted values ​​of the pre-trained model to calculate the loss function; S5.3 Fine-tuning is achieved by inserting a low-rank adapter matrix into the model and training only the adapter matrix, thus enabling continuous self-evolution of the model without losing pre-trained knowledge. S5.4 Periodically runs the fine-tuned model, outputs the continuously updated remaining service life of the windproof guy wire device, and compares it with the preset safety threshold to provide early warning.

[0008] The shift-window self-attention mechanism in step S2.3 achieves efficient parallel computation of high-resolution grid data by restricting attention calculations to a local window and shifting the window position between adjacent computation layers.

[0009] The state transition matrix in step S3.2 is in the form of (A+Bu), where A is the basic state transition matrix in step S3.2, B is the learnable weight matrix in step S3.2, and u is the current input signal in step S3.2.

[0010] In step S3.3, the liquid convolution kernel models the combined effect of the stress sequence by calculating the autocorrelation term ui·uj and the higher-order term ui·uj·uk of the input signal, where i, j, and k are indices at different times.

[0011] The adaptive layer normalization mechanism in step S4 conditions the normalization layer inside the model by encoding environmental parameters as scale parameter γ and offset parameter β, thereby achieving fine control of environmental conditions over the model calculation process.

[0012] The low-rank adapter matrix in step S5.3 is implemented by decomposing the weight update into the product of two low-rank matrices ΔW=BA, where the ranks of B and A are both less than the dimension of the original weight matrix, thereby achieving efficient fine-tuning of the parameters.

[0013] The sensor in step S5.1 is a strain gauge, accelerometer, or displacement sensor installed on the modular connecting block of the windproof guy wire device.

[0014] The physical process proxy model and the long sequence analysis model are basic models obtained by pre-training on historical simulation and measured datasets covering different wind conditions, structural parameters and environmental factors.

[0015] The independent channel embedding in step S2.1 maintains the independence of different physical quantities and avoids information compression loss by independently mapping each physical channel to a high-dimensional vector representation at each location of the spatial grid.

[0016] Step S5.4 also includes: establishing a confidence interval assessment of the remaining service life, issuing an early warning signal when the lower bound of the confidence interval is lower than a preset safety threshold, and issuing an emergency replacement signal when the upper bound of the confidence interval is lower than a preset safety threshold.

[0017] The beneficial effects of this invention are: This invention achieves accurate modeling of complex wind-induced vibration phenomena by constructing a three-dimensional digital twin model of a windproof guy wire device, including modular connecting blocks, a metal shell, and damping elements, and defining multiple physical channels on a discretized three-dimensional mesh. The physical channels include the x and y components of the wind speed field, the wind pressure field, the σx, σy, and τxy components of the structural stress tensor field, and the structural displacement field, providing a complete physical information foundation for subsequent intelligent analysis.

[0018] The physical process proxy model based on a multi-scale hierarchical architecture employed in this invention significantly improves computational efficiency and prediction accuracy. By treating multiple physical channels as independent inputs and embedding them independently into different representation vectors on a spatial grid, the information compression loss caused by the mixed encoding of multiple physical quantities in traditional methods is effectively avoided, maintaining the independence of different physical quantities. The channel-axis self-attention mechanism automatically learns the complex coupling relationships between wind pressure field and structural stress field, and between structural displacement and flow field distribution by calculating the attention weights between the representation vectors of different physical channels, thus achieving accurate simulation of fluid-structure interaction effects.

[0019] The U-Net-type hierarchical architecture and shift-window self-attention mechanism employed in this invention significantly reduce computational complexity while maintaining high-resolution data processing capabilities. The shift-window self-attention mechanism achieves efficient parallel computation by confining attention calculations to a local window and shifting the window position between adjacent computation layers, while also ensuring information exchange between different regions through the window shifting strategy. Multi-scale information processing achieved through upsampling and downsampling of the representation vectors enables the model to simultaneously capture local stress concentration phenomena and global vibration modes, significantly improving the modeling ability for complex wind-induced vibration phenomena.

[0020] This invention employs a long-sequence analysis model based on input-dependent state transitions, overcoming the limitations of traditional fatigue analysis methods. The input-dependent state transition mechanism modifies the state transition matrix to the form (A+Bu), enabling the state transition coefficients to adjust in real-time according to the current input signal, thus achieving adaptive modeling of damage accumulation patterns under different loading conditions. The liquid convolution kernel generated based on this mechanism can explicitly model the autocorrelation of the input stress sequence, with its weights correlated to the autocorrelation terms of the input signal. This effectively captures higher-order correlations and combined effects in the stress sequence, a significant breakthrough that traditional rainflow counting methods cannot achieve.

[0021] This invention models the combined effect of stress sequences by calculating the autocorrelation term ui·uj and higher-order term ui·uj·uk of the input signal using a liquid convolution kernel. It can accurately identify complex nonlinear damage mechanisms such as the synergistic effect of continuous high-stress pulses and the damage acceleration effect under alternating high and low stress, and significantly improves the accuracy of fatigue life prediction under variable amplitude load conditions.

[0022] This invention achieves intelligent, environmentally adaptive prediction by endogenizing environmental parameters into a physical process surrogate model and a long-sequence analysis model through a deep conditional injection mechanism. The adaptive layer normalization mechanism conditions the normalization layer within the model by encoding environmental parameters as scale parameters γ and offset parameters β, enabling the model to learn the nonlinear effects of environmental factors such as corrosion level, ambient temperature, and ultraviolet radiation intensity on physical processes and damage accumulation. This deeply integrated environmental modeling approach offers higher accuracy and adaptability compared to traditional post-processing correction methods.

[0023] This invention utilizes a pre-trained base model based on historical simulation and measured datasets covering various wind conditions, structural parameters, and environmental factors. This model possesses strong generalization and transfer learning capabilities. The pre-trained base model provides a universal starting point for different application scenarios, allowing for rapid adaptation to new device types, environmental conditions, or application scenarios through fine-tuning with a small amount of specific data, significantly reducing the time and cost of model development and deployment.

[0024] This invention employs a low-rank adaptive technique to achieve efficient online fine-tuning of pre-trained model parameters. The low-rank adapter matrix is ​​implemented by decomposing weight updates into the product of two low-rank matrices, ΔW=BA, where the ranks of both B and A are less than the dimension of the original weight matrix. This preserves pre-training knowledge while enabling continuous self-evolution of the model. By inserting the low-rank adapter matrix into the model and training only this adapter matrix, the system can continuously optimize prediction performance based on measured data collected from field sensors, adapting to long-term evolution processes such as equipment aging and environmental changes.

[0025] This invention achieves multi-dimensional condition monitoring by installing strain gauges, accelerometers, or displacement sensors on modular connecting blocks of the windproof guy wire device, providing rich data support for online fine-tuning. Multi-sensor fusion technology significantly improves the accuracy and reliability of fault detection and effectively reduces the false alarm rate.

[0026] The remaining service life confidence interval assessment mechanism established in this invention effectively balances the timeliness and accuracy of early warning. An early warning signal is issued when the lower bound of the confidence interval falls below a preset safety threshold, and an emergency replacement signal is issued when the upper bound of the confidence interval falls below the preset safety threshold. This intelligent early warning strategy based on uncertainty quantification avoids the false alarms and missed alarms that may occur with traditional threshold-based early warning methods.

[0027] This invention represents a fundamental shift from traditional passive maintenance to proactive predictive maintenance. Through continuously updated remaining service life prediction and intelligent early warning mechanisms, maintenance personnel can identify potential fatigue risks in advance, formulate scientific and reasonable maintenance strategies, effectively reduce equipment failure rates and maintenance costs, and improve the reliability and economy of the power system. Attached Figure Description

[0028] Figure 1 The system architecture of this invention Figure 1 ; Figure 2 The system architecture of this invention Figure 2 . Detailed Implementation

[0029] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] It should be noted that the directional concepts of "left", "right", "up", "down", "front", "back", "inner", and "outer" in the following scheme are all relative directions, and will not be listed one by one here.

[0031] Example 1: Basic Physical Simulation and Damage Assessment System like Figure 1 As shown, this embodiment provides a basic implementation scheme for predicting the fatigue life of wind-resistant guy wire devices for power transmission towers based on dynamic wind-resistant design. This scheme achieves accurate prediction of the fatigue life of wind-resistant guy wire devices by constructing a high-precision digital twin model and an advanced AI proxy model.

[0032] The windproof guy wire device in this embodiment mainly consists of three parts: modular connecting blocks, a metal shell, and damping elements. The modular connecting blocks are made of high-strength alloy steel and are responsible for bearing the main tensile loads; the metal shell is made of stainless steel, providing corrosion protection; and the damping elements are made of viscoelastic material to dissipate vibration energy. The entire device is installed at appropriate locations on the transmission line using standardized bolt connections.

[0033] First, a three-dimensional digital twin model of the wind-resistant guy wire device and its associated transmission tower is constructed. A precise geometric model of the device is then created using computer-aided design software, including the internal thread structure of the modular connecting blocks, the surface roughness characteristics of the metal shell, and the nonlinear material properties of the damping elements. This geometric model and its surrounding flow field space are discretized. A structured hexahedral mesh is used to divide the solid region, with a mesh density of at least 10 elements per millimeter in key fatigue areas. For the fluid region, an unstructured tetrahedral mesh is used, with boundary layer refinement near the wall to ensure accurate capture of the development of the turbulent boundary layer.

[0034] Multiple physical channels are defined on a discretized 3D mesh to characterize the physical states of the fluid and solid. These channels include the x-component Vx and y-component Vy of the wind speed field, where the x-direction is the prevailing wind direction and the y-direction is the horizontal direction perpendicular to the prevailing wind direction; the wind pressure field P characterizes the normal pressure distribution of the fluid on the structural surface; the structural stress tensor field includes normal stress components σx, σy and shear stress component τxy, which fully describe the two-dimensional stress state; and the structural displacement field D records the displacement vectors of each mesh node. Each physical channel is defined as an independent scalar or vector field on the spatial mesh, providing structured input data for subsequent AI model processing.

[0035] Next, a physical process surrogate model based on a multi-scale hierarchical architecture is used to perform wind-induced vibration response simulation. The core of this surrogate model lies in replacing the traditional numerical solver with a pre-trained deep learning model, thereby significantly improving computational efficiency. The model's input consists of the multiple physical channels defined above, and the output is the stress-strain time history data of key fatigue points of the windproof guy wire device within a specified time period.

[0036] In the first stage of the model, the multiple physical channels defined in the steps are input as independent channels into the physical process proxy model. For each physical channel, an independent embedding operation is performed at each location on the spatial grid, mapping the original physical quantity numerical value into a high-dimensional vector representation. Specifically, for the physical quantity φ at grid location (i,j), it is embedded into a d-dimensional vector through a linear transformation: in It is a physical channel The corresponding embedding weight matrix, It is a bias vector. This is the obtained representation vector. This independent channel embedding method ensures the independence of different physical quantities and avoids the information compression loss that may occur when simply splicing together multiple physical quantities in traditional methods.

[0037] In the second stage of the model, the coupling relationships between different physical channels are modeled through a channel-axis self-attention mechanism within the physical process proxy model to simulate fluid-structure interaction effects. The core idea of ​​this mechanism is to calculate the attention weights between the representation vectors of different physical channels, thereby learning the interaction relationships between the channels. For a spatial location (i,j), there are C physical channels, whose representation vectors are respectively... The formula for calculating the channel axial self-attention is: ; ;in This represents the attention weight of the m-th channel to the n-th channel. These are the weight matrices for the query, key, and value, respectively. It is the output feature of the m-th channel after passing through the attention mechanism. This design enables the model to automatically learn complex physical coupling relationships such as how wind pressure fields affect structural stress fields and how structural displacements, in turn, affect flow field distribution.

[0038] In the third stage of the model, a U-Net-type hierarchical architecture and a shift-window self-attention mechanism are used for multi-scale feature extraction and fusion. The U-Net architecture consists of an encoder and a decoder. The encoder extracts features at different scales by downsampling layer by layer, while the decoder recovers the original resolution by upsampling, and uses skip connections to preserve detailed information. In each layer, a shift-window self-attention mechanism is used to capture spatial correlations.

[0039] The shift-window self-attention mechanism divides the entire feature map into multiple non-overlapping windows, performing self-attention computation independently within each window, thereby reducing computational complexity from... Reduce to , where H and W are the height and width of the feature map, respectively. For a window eigenvectors within The formula for calculating window self-attention is: ; ;in It is the concatenation of all feature vectors within the window. This is the dimension of the key vector. To enhance information exchange between different windows, a window shifting strategy is adopted between adjacent computation layers. This involves shifting the window division to the lower right by half its size. In this way, adjacent regions that were originally in different windows will be assigned to the same window in the next layer, thereby achieving cross-window information transfer.

[0040] Multi-scale information processing is achieved through upsampling and downsampling operations. Downsampling uses a convolution operation with a stride of 2, halving the feature map size while increasing the number of channels; upsampling uses transposed convolution or bilinear interpolation followed by convolution, doubling the feature map size while reducing the number of channels. Multi-scale processing enables the model to simultaneously capture local stress concentration phenomena and global vibration modes, improving prediction accuracy.

[0041] In the fourth stage of the model, an autoregressive approach is used for time-step prediction. Let the current time be t, and the states of the previous L time steps be known. The model predicts the state at the next time step. : ;in The parameter is Physical process proxy model, It contains all physical channel values ​​for all grid points at time t. By iteratively applying this prediction process, complete stress-strain time history data for critical fatigue points of the windproof guy wire device can be obtained over any specified time period.

[0042] Subsequently, a long-sequence analysis model based on input-dependent state transitions is used to assess fatigue damage from the stress-strain time history data. Traditional fatigue analysis methods, such as rainflow counting, primarily focus on the amplitude and mean of stress cycles, neglecting the influence of loading history and loading sequence. The long-sequence analysis model used in this embodiment, through an input-dependent state transition mechanism, can explicitly model the impact of loading history on the current damage state.

[0043] First, the stress-strain time history data output from the previous steps are input as the input sequence into the long-sequence analysis model. Let the stress sequence be... , where T is the total number of time steps. The goal of the model is to learn the mapping from this sequence to cumulative fatigue damage.

[0044] The core of long sequence analysis models is the input-dependent state transition mechanism. In traditional linear state-space models, the state transition equation is: ; in It is the hidden state at time t. Let A and B be the input at time t, and A and B be fixed system matrices. In this embodiment, the input-dependent state transition mechanism modifies the state transition matrix into an input-dependent form: ;in It is the basic state transition matrix, which describes the inherent dynamic characteristics of the system; It is a learnable weight tensor, where m is the dimension of the input vector; It is a mapping matrix from input to state; This is the current input signal, corresponding to the stress-strain value at the current moment in this application. This design allows the state transition coefficient to be adjusted in real time according to the current input signal, thereby adapting to the damage accumulation patterns under different loading conditions.

[0045] Based on the input-dependent state transition mechanism, the model generates a special convolutional kernel called a liquid convolutional kernel. When the above state transition equation is expanded into convolutional form, in addition to the standard linear terms, nonlinear terms containing input autocorrelation terms also appear. The weights of the liquid convolutional kernel are related to the autocorrelation term of the input signal, and its expression is: ; in These are the weighting coefficients derived from the state transition parameter B. It is the Dirac function. This refers to the time delay. This liquid convolution kernel can explicitly model the autocorrelation of the input stress sequence, capturing higher-order correlations and combined effects within the stress sequence. For example, when consecutive high-stress pulses occur, their combined effect may lead to more severe damage accumulation than when they act individually; this nonlinear effect is precisely achieved through the autocorrelation term of the liquid convolution kernel. To model.

[0046] Ultimately, the model directly learns the mapping relationship from the stress-strain sequence to the fatigue damage increment. Let the cumulative damage be D, then the damage increment at time t is: Where g is a nonlinear function, typically implemented using a multilayer perceptron. The total cumulative damage is: ;when When the value reaches 1, the material is considered to have experienced fatigue failure.

[0047] The working process of this embodiment can be divided into two stages: offline pre-training and online prediction. In the offline pre-training stage, a large amount of wind-induced vibration simulation data and corresponding fatigue test data under different working conditions are first collected to train the physical process surrogate model and the long sequence analysis model. The pre-training dataset contains samples under different wind speeds, wind directions, turbulence intensities, structural parameters, etc., to ensure that the model has good generalization ability.

[0048] When training the surrogate model of the physical process, a supervised learning approach is adopted, using the initial state and boundary conditions as input and the physical field distribution at subsequent time steps as the supervision target. The loss function is designed as a weighted sum of the prediction errors of multiple physical channels: ; in Indicates different physical channels. These are the corresponding weighting coefficients, set according to the importance and numerical range of each physical quantity.

[0049] When training the long-sequence analysis model, stress time history data is used as input, and experimentally measured fatigue life or damage degree is used as the supervision target. Considering the randomness of the fatigue process, an uncertainty quantification term is also added to the loss function: .

[0050] in and These are the predicted and actual cumulative damage, It is the uncertainty of prediction. It is the balance coefficient.

[0051] During the online prediction phase, the system receives current environmental conditions (wind speed, wind direction, etc.) as input. First, it generates corresponding stress-strain time history data through a physical process surrogate model. Then, this time history data is input into a long-sequence analysis model to obtain fatigue damage assessment results. The computation time of the entire prediction process is reduced by 2-3 orders of magnitude compared to traditional numerical simulation methods, from several hours to just a few minutes, greatly improving its practicality.

[0052] In practical applications, the system generates wind load time histories for a future period (such as a week or a month) based on meteorological forecast data, and then performs batch predictions to obtain the cumulative fatigue damage within the corresponding time period. By comparing this with safety thresholds, potential fatigue risks can be identified in a timely manner, providing a scientific basis for maintenance decisions.

[0053] First, by employing independent channel embedding and channel-axis self-attention mechanisms, the model can better maintain the independence of different physical quantities, avoiding the information loss caused by mixed encoding of multiple physical quantities in traditional methods. Experiments show that the method using independent channel embedding improves prediction accuracy by 15-20% compared to the mixed encoding method.

[0054] Secondly, the shift-window self-attention mechanism significantly reduces computational complexity, enabling the model to handle high-resolution grid data. For a 1024×1024 grid, traditional global self-attention requires approximately 10^12 floating-point operations, while shift-window self-attention only requires approximately 10^12. 8 This improved computational efficiency by four orders of magnitude without a significant decrease in prediction accuracy.

[0055] Third, the introduction of the input-dependent state transition mechanism and the liquid convolution kernel enables the model to capture the influence of loading history, which is impossible with traditional rainflow counting methods. Comparative experiments show that, under variable amplitude load conditions, the method considering the influence of loading history improves the lifetime prediction accuracy by 25-30% compared with traditional methods.

[0056] Fourth, the multi-scale hierarchical architecture enables the model to handle both local and global features simultaneously, resulting in stronger modeling capabilities for complex wind-induced vibration phenomena. Under complex vibration modes such as vortex-induced vibration and galloping, the prediction error of this method is reduced by 30-40% compared to traditional methods.

[0057] Finally, the computational efficiency of the entire system is improved by 2-3 orders of magnitude compared to traditional fluid-structure interaction simulation, making real-time or near real-time fatigue monitoring possible and laying a technical foundation for the construction of intelligent operation and maintenance systems. While maintaining high accuracy, the calculation time for a single prediction is reduced from several hours to a few minutes, meeting the real-time requirements of engineering applications.

[0058] Example 2: Environmental Adaptive Intelligent Prediction System like Figure 1 and 2 As shown, this embodiment, based on embodiment 1, further provides an intelligent prediction system with environmental adaptability. Through deep condition injection mechanism and high-order correlation modeling, the system can accurately predict the fatigue life of windproof guy wire devices under different environmental conditions, significantly improving prediction accuracy and applicability.

[0059] This embodiment builds upon the basic digital twin modeling and multi-scale hierarchical architecture of Embodiment 1, focusing on enhancing environmental perception and high-order damage modeling capabilities. A new environmental parameter monitoring module has been added to the system, enabling real-time collection of key environmental factors affecting fatigue life, including corrosion level, ambient temperature, and ultraviolet radiation intensity.

[0060] The environmental parameter monitoring module employs a distributed sensor network for data acquisition. Corrosion level is measured using an electrochemical sensor, which quantifies the degree of corrosion on the metal surface based on changes in corrosion potential, outputting a dimensionless corrosion level index ranging from 0 to 10, where 0 represents no corrosion and 10 represents severe corrosion. Ambient temperature is measured using a high-precision digital temperature sensor, with a measurement range of -40°C to +80°C and an accuracy of ±0.1°C. Ultraviolet radiation intensity is measured using a UV sensor, with an output unit of W / m² and a measurement range of 0-100 W / m². These environmental parameters are collected every 15 minutes, forming continuous time-series data.

[0061] In the implementation of the deep conditional injection mechanism, environmental parameters are no longer simply treated as post-processing correction factors, but are deeply integrated as conditional inputs into the internal computational processes of the physical process surrogate model and the long sequence analysis model. Specifically, an adaptive layer normalization mechanism encodes environmental parameters into scale parameters and offset parameters, and conditionally modulates the normalization layer within the model, thereby achieving fine-grained control of environmental conditions over the model's computational process.

[0062] The core of the adaptive layer normalization mechanism lies in extending the standard layer normalization operation into a conditionally dependent form. For any feature vector in the model... The formula for calculating the normalization of the standard layer is: ;in and These are the mean and standard deviation of x, respectively. and These are learnable scale and offset parameters. In adaptive layer normalization, and Instead of being fixed parameters, they are dynamically generated based on environmental conditions c. ; ; in It is a coded representation of environmental conditions. It is a linear transformation matrix. It is a bias term. Environmental condition coding. The original environmental parameters were obtained by performing a nonlinear transformation using a multilayer perceptron: ;in represents the normalized values ​​of corrosion level, ambient temperature, and ultraviolet radiation intensity, respectively, and [;] indicates the splicing operation. In this way, environmental conditions can directly affect the distribution of features within the model, thereby adjusting the model's ability to model physical processes under different environments.

[0063] In the physical process proxy model, environmental conditions primarily influence the modeling of material properties. For example, high-temperature environments reduce the elastic modulus and fatigue strength of metallic materials, strong ultraviolet radiation accelerates the aging of polymer damping materials, and corrosive environments alter the stress distribution on the material surface. Through an adaptive layer normalization mechanism, the model can learn the influence of these environmental factors on material properties and automatically adjust the corresponding physical parameters during the calculation process.

[0064] In long-sequence analysis models, environmental conditions primarily influence the accumulation of fatigue damage. The same stress level will produce different damage increments under different environmental conditions. For example, in high-temperature and high-humidity environments, the fatigue crack propagation rate increases significantly; in strong ultraviolet environments, the fatigue limit of the material decreases. Through conditional injection mechanisms, the model can learn the modulating effect of the environment on the damage accumulation process.

[0065] In this embodiment, the liquid convolution kernel is further extended to a higher-order form, enabling the modeling of more complex stress sequence combination effects. In addition to the second-order autocorrelation term... In addition, third-order terms are introduced. Its complete expression is: ; in and These are the weighting coefficients for the second-order and third-order terms, respectively, which are derived from the parameters of the input-dependent state transition mechanism. The introduction of the third-order term enables the model to capture more complex nonlinear damage mechanisms, such as the synergistic effect of multi-peak stresses and the damage acceleration effect under alternating high and low stresses.

[0066] To effectively train this high-order model, this embodiment employs a pre-training strategy based on multi-domain data. The pre-training dataset encompasses combinations of different wind conditions, structural parameters, and environmental factors, forming a large-scale multimodal dataset. Wind condition data includes combinations of different wind speed levels (3-25 m / s), different turbulence intensities (5%-25%), and different wind direction angles (0°-360°); structural parameters include guy wire devices with different diameters (10-50 mm), different materials (aluminum alloy, stainless steel, carbon steel), and different geometries; environmental factors include data on different temperature ranges (-30°C to +60°C), different humidity levels (20%-90%), and different corrosive environments (marine, industrial, rural).

[0067] The pre-training process employs a multi-task learning framework, simultaneously optimizing both physical process prediction and fatigue damage assessment tasks. The overall loss function is designed as follows: ; in It is a physical process to predict loss. It is a fatigue damage prediction loss. It is a multi-scale consistency loss. This refers to the loss of environmental adaptability. Specifically designed to enhance the model's generalization ability under different environmental conditions: in This represents the model output under environmental condition c. It is an environmental condition similarity function. When c and c' are similar, s(c,c') is smaller, which forces the model to produce similar outputs in similar environments, thereby improving the model's environmental adaptability.

[0068] In its implementation, both the physical process proxy model and the long sequence analysis model use the Transformer architecture as their basic framework, but with specific optimizations tailored to the characteristics of physical modeling. The models have hundreds of millions of parameters and are pre-trained on a large-scale GPU cluster through distributed training. The pre-training process employs a course-based learning strategy, starting with simple, single-environment conditions and gradually increasing the complexity and diversity of the environments, ultimately achieving the ability to model multiple environmental conditions simultaneously.

[0069] This embodiment adds environmental perception and adaptive prediction steps to the workflow of Embodiment 1. After system startup, the environmental parameter monitoring module first collects current environmental condition data, including corrosion level, ambient temperature, and ultraviolet radiation intensity. After quality inspection and data cleaning, this data, along with wind condition data, is input into the prediction system.

[0070] During the environmental condition coding phase, the system transforms the original environmental parameters into an input format acceptable to the model through standardization. The corrosion level is directly normalized by dividing by 10; the ambient temperature is mapped to the [-1, 1] interval through a linear transformation. ; Where T is the actual temperature value, with 20°C as the reference temperature; the ultraviolet radiation intensity is normalized through logarithmic transformation: ; in This is the actual radiation intensity value. This normalization method can balance the numerical range of different environmental parameters and prevent any one parameter from dominating the model's decision-making process.

[0071] During the physics simulation phase, environmental conditions influence the model's internal computation through an adaptive layer normalization mechanism. Specifically, after the multi-head self-attention computation of each Transformer layer, a condition-dependent normalization operation is applied. This allows the model to dynamically adjust the distribution of attention weights according to the current environmental conditions, thereby adapting to the physical laws under different environments.

[0072] In the fatigue damage assessment phase, higher-order liquid convolution kernels play a crucial role. For the input stress sequence, the model not only calculates the correlation between stress values ​​at adjacent time points but also computes more complex ternary combined effects. For example, when high stress is detected at three consecutive time points, the third-order term generates an additional damage contribution, reflecting the accelerating effect of stress concentration on the fatigue process.

[0073] In the prediction output stage, the system not only provides point estimates of fatigue life but also interval estimates based on environmental uncertainties. Through Monte Carlo sampling, reasonable noise perturbations are added to environmental parameters to generate multiple prediction results, thereby assessing the impact of environmental changes on the prediction outcomes. The final output fatigue life prediction includes the mean, standard deviation, and confidence interval, providing decision-makers with more comprehensive information.

[0074] The entire system employs a hierarchical caching mechanism to improve computational efficiency. Environmental condition encoding results are cached for a certain period to avoid redundant calculations; intermediate results from physical process simulations are also cached and can be reused directly when environmental conditions remain relatively stable. This design enables the system to achieve near real-time predictive response while maintaining high accuracy.

[0075] First, the deep condition injection mechanism enables the model to accurately model the impact of the environment on physical processes and fatigue behavior. Experiments show that, when environmental factors are considered, the error in fatigue life prediction is reduced by 40-50% compared to methods that do not consider environmental factors. The improvement in prediction accuracy is particularly significant under extreme environmental conditions (such as high temperature and humidity, and strong ultraviolet radiation).

[0076] Secondly, the introduction of higher-order liquid convolution kernels significantly improves the model's ability to model complex load histories. In variable-amplitude load tests, the third-order term can effectively capture the nonlinear damage effect under alternating high and low stresses, resulting in a 20-25% improvement in prediction accuracy compared to methods that only consider second-order terms. This improvement is particularly evident under broadband random load conditions, as the higher-order statistical characteristics of the stress sequence have a significant impact on fatigue behavior.

[0077] Third, the pre-training strategy based on multi-domain data endows the model with strong generalization ability. The pre-trained model can still maintain high prediction accuracy under unseen environmental conditions, and the generalization error is reduced by 30-35% compared with the model trained from scratch. This generalization ability enables the system to be applied to transmission lines in different geographical regions and under different climatic conditions, greatly expanding the scope of application.

[0078] Fourth, the design of the environmental adaptability loss function ensures the consistency of the model's predictions under similar environmental conditions, avoiding large fluctuations in prediction results caused by minor changes in environmental parameters. This stability is crucial for engineering applications because it reduces false alarms and false negatives caused by environmental noise.

[0079] Fifth, the multi-task learning framework enables the physical process prediction and fatigue damage assessment tasks to mutually reinforce each other, improving the overall system performance. The accuracy of physical process prediction provides reliable input for fatigue assessment, while feedback from fatigue assessment helps optimize the modeling of the physical process. Experiments show that multi-task learning achieves an average performance improvement of 5-10% on both tasks compared to single-task learning.

[0080] Finally, the pre-trained base model established in this embodiment provides a powerful starting point for subsequent specific applications. Through transfer learning, it can quickly adapt to new device types, new environmental conditions, or new application scenarios, greatly reducing the time cost of model development and deployment. In actual engineering projects, fine-tuning based on the pre-trained model typically only requires 10-20% of the original training time to achieve satisfactory performance, significantly improving the practicality and economy of the technology.

[0081] Furthermore, the environmental adaptability of this embodiment makes it particularly suitable for long-term lifetime prediction in the context of climate change. With global climate change, the frequency and intensity of extreme weather events are increasing, posing a growing challenge to traditional methods based on historical statistical data. This embodiment, by sensing environmental changes in real time and dynamically adjusting the prediction model, can better cope with potential new environmental conditions in the future, providing stronger guarantees for the reliability and security of power systems.

[0082] Example 3: Online Self-Evolution Early Warning System like Figure 1 and 2 As shown, this embodiment, based on the aforementioned embodiments, further provides a self-evolving system with online learning and intelligent early warning capabilities. This system achieves real-time monitoring and accurate early warning of the fatigue state of the windproof guy wire device through multi-sensor fusion, low-rank adaptive online fine-tuning, and intelligent early warning decision-making, significantly improving the practicality and reliability of the system.

[0083] This embodiment inherits the digital twin modeling, multi-scale physical simulation, and environmental adaptive prediction technologies from the previous embodiments, and focuses on enhancing the on-site data acquisition capabilities, online model optimization capabilities, and intelligent decision-making capabilities. The system adds a distributed multi-sensor network, an edge computing module, and a cloud-based collaborative optimization module, forming a complete "perception-computation-decision-execution" closed-loop control system.

[0084] The multi-sensor data acquisition system employs a wireless sensor network architecture, deploying various types of sensors at key locations on the windproof guy wire device. The strain gauge sensors utilize resistive strain gauges, installed at the maximum stress section of the modular connecting block. They measure strain changes along the principal stress directions, with a measurement range of ±5000με, an accuracy of ±1με, and a sampling frequency of 1000Hz. The accelerometer sensors employ triaxial accelerometers based on MEMS technology, installed at the center of mass of the guy wire device. They measure vibration acceleration in three directions, with a measurement range of ±50g, an accuracy of ±0.01g, and a sampling frequency of 2000Hz. The displacement sensors utilize non-contact laser displacement gauges, mounted on a fixed bracket. They measure the absolute displacement of the guy wire device, with a measurement range of ±100mm, an accuracy of ±0.01mm, and a sampling frequency of 500Hz.

[0085] The sensor network uses the ZigBee protocol for data transmission. Each sensor node is equipped with a low-power microprocessor and a wireless communication module, enabling local data preprocessing and time synchronization. The network topology adopts a tree structure to reduce data transmission latency and energy consumption. Each sensor node is also equipped with a solar panel and a lithium battery to ensure long-term stable operation in harsh environments.

[0086] Online fine-tuning technology is based on a low-rank adaptive method. The core idea of ​​this method is to adjust only a small number of parameters on a pre-trained model, thereby adapting to new data distributions while preserving existing knowledge. Specifically, for the weight matrix in the pre-trained model... Its update format is as follows: ;in and They are two low-rank matrices. It is the rank of the low-rank decomposition. This decomposition method reduces the number of parameters that need to be trained from... Reduce to When r is much smaller than d and k, the reduction in the number of parameters is very significant.

[0087] The low-rank adapter matrix is ​​initialized using a special strategy: matrix A is initialized with a random Gaussian distribution, and matrix B is initialized as a zero matrix. This design aims to ensure that at the start of training... This means that the initial behavior of the model is completely consistent with the pre-trained model, and then it gradually learns knowledge specific to the current device through training.

[0088] The loss function design for online fine-tuning considers several requirements. The first is the prediction accuracy loss, calculated by comparing the model's predicted values ​​with the sensor's measured values: ; in It is the model's predicted value for the i-th sample. These are the corresponding measured values, and N is the number of samples. Next is the regularization loss, used to prevent overfitting: ; in Describing the Frobenius norm, and This is the regularization coefficient. The total loss function is: To ensure the stability of online fine-tuning, the system uses an exponential moving average strategy to update model parameters. ; in These are the parameters after the t-th update. These are the new parameters obtained from the current gradient descent. It is a smoothing coefficient. This strategy can prevent a sharp drop in model performance caused by a single instance of outlier data.

[0089] The intelligent early warning decision-making system is designed based on Bayesian uncertainty quantification theory. It not only provides point estimates of fatigue life but also gives corresponding confidence intervals. The system uses Monte Carlo Dropout technology to estimate the cognitive uncertainty of the model and obtains the prediction distribution through multiple forward propagations. ; Where T is the number of Monte Carlo samplings. These are the model parameters obtained from the t-th sampling. Based on this prediction distribution, the confidence interval can be calculated: ;in This represents the p-th quantile of the predicted distribution. This is the confidence level (usually taken as 0.95).

[0090] The early warning decision-making logic is designed based on risk assessment theory. Let... This is a predicted value for the remaining useful life. and These are the lower and upper bounds of the confidence interval, respectively. If the preset safety threshold is used, then the early warning decision rule is as follows: when When a warning signal is issued, it indicates a high risk of failure and the need for enhanced monitoring; when When this happens, an emergency replacement signal is issued, indicating an extremely high risk of failure and requiring immediate maintenance.

[0091] This confidence interval-based decision-making method effectively balances the risks of false alarms and missed alarms, improving the reliability of the early warning system. To further enhance the accuracy of early warnings, the system also incorporates trend analysis. By analyzing the changing trends of historical prediction results, abnormal degradation patterns can be identified. Specifically, the system maintains a remaining lifetime prediction sequence within a sliding window. And calculate its linear trend: ; when When the absolute value exceeds the preset threshold, it indicates an abnormal trend in the remaining lifespan, requiring extra attention.

[0092] The system also implements a multi-level early warning mechanism, adopting different response strategies based on the different levels of failure risk. Level 1 warning (green) indicates that the equipment is in normal condition and requires no special handling; Level 2 warning (yellow) indicates that there is a potential risk and the monitoring frequency needs to be increased; Level 3 warning (orange) indicates that the risk is high and a maintenance plan needs to be developed; Level 4 warning (red) indicates that the risk is extremely high and immediate shutdown and maintenance are required.

[0093] To support edge computing deployment, the system employs model compression and quantization techniques. Knowledge distillation is used to transfer knowledge from large pre-trained models to lightweight student models. in and These are the outputs of the teacher model and the student model, respectively. It is the softmax function, where T is the temperature parameter. It is cross-entropy loss. It uses KL divergence loss. The number of parameters in the student model is reduced by 80-90% compared to the teacher model, but the performance loss is kept within 5%.

[0094] The working process of this embodiment includes five main stages: system initialization, real-time monitoring, online learning, intelligent early warning, and maintenance decision-making. In the system initialization stage, all sensors are first calibrated and synchronized to ensure the accuracy and consistency of data acquisition. Then, the pre-trained physical process proxy model and long sequence analysis model are loaded, the low-rank adapter matrix is ​​initialized, and a communication connection with the cloud server is established.

[0095] During the real-time monitoring phase, the multi-sensor network continuously acquires data at a preset sampling frequency. Strain gauge sensors monitor the stress response of the structure, accelerometers monitor vibration characteristics, and displacement sensors monitor overall deformation. All sensor data are synchronized via timestamps, and preliminary data fusion and feature extraction are performed at edge computing nodes.

[0096] Data quality control is a crucial aspect of real-time monitoring. The system employs a multi-layered verification mechanism to identify and eliminate abnormal data: first, a range check is performed to remove data exceeding the physically possible range; then, a consistency check is conducted to compare the reasonableness of measurement results from different sensors; finally, a trend check is performed to identify abrupt changes that do not conform to physical laws. For detected abnormal data, the system automatically marks it and uses interpolation or filtering methods to repair it.

[0097] The online learning phase is the core innovation of this embodiment. The system triggers the online learning process at set time intervals (usually 1 hour) or when a sufficient amount of new data has been accumulated. First, the newly collected data is compared with the model prediction results to calculate the prediction error and uncertainty. If the error exceeds a preset threshold, the model fine-tuning process is initiated.

[0098] Online learning employs a mini-batch gradient descent algorithm, using data from the most recent time window for each training iteration. To balance learning efficiency and stability, the system adopts an adaptive learning rate strategy. ;in It is the initial learning rate. This is the decay coefficient, and t is the current iteration number. This design allows the model to quickly adapt to new data in the early stages and maintain stability in the later stages.

[0099] The intelligent early warning phase makes decisions based on real-time fatigue life prediction results. The system updates the remaining life prediction every 15 minutes and generates corresponding early warning signals according to the aforementioned early warning decision rules. The early warning information includes the current remaining life estimate, confidence interval, risk level, and recommended countermeasures.

[0100] To improve the interpretability of early warnings, the system also provides a risk factor analysis function. By analyzing the impact of different environmental factors and load patterns on fatigue life, the main sources of current risks are identified. For example, if the recent increase in wind speed is found to be the main cause of the decline in remaining life, the system will clearly indicate this in the early warning message, providing targeted guidance for maintenance personnel.

[0101] The maintenance decision-making phase comprehensively considers early warning information, operational resources, and economic costs to generate an optimized maintenance plan. The system maintains a maintenance cost model, considering factors such as preventative maintenance costs, fault repair costs, and power outage losses, and uses a dynamic programming algorithm to solve for the optimal maintenance timing. ;in Let c(s,a) be the optimal value function under state s, and c(s,a) be the instantaneous cost of taking action a under state s. s is the discount factor, and s' is the state after the transition.

[0102] First, multi-sensor fusion technology significantly improves the accuracy and reliability of condition monitoring. By fusing and analyzing multiple physical quantities such as strain, vibration, and displacement, the system can gain a more comprehensive understanding of the equipment's health status. Experiments show that multi-sensor fusion improves fault detection accuracy by 35-40% and reduces false alarm rate by 50-60% compared to single-sensor fusion.

[0103] Secondly, low-rank adaptive online fine-tuning technology enables continuous model optimization, allowing the system to adapt to long-term evolution processes such as equipment aging and environmental changes. The prediction accuracy of the online-fine-tuned model is improved by 20-30% compared to the fixed model, with the improvement being even more pronounced in the later stages of equipment operation. Simultaneously, low-rank decomposition technology significantly reduces the number of parameters that need to be adjusted, resulting in a reduction of over 90% in the computational cost of online learning.

[0104] Third, the confidence interval-based intelligent early warning mechanism effectively balances the timeliness and accuracy of early warnings. By simultaneously considering predicted values ​​and uncertainties, the system can issue timely warnings in high-risk situations and avoid false alarms in low-risk situations. Statistical analysis shows that the overall performance of confidence interval-based early warning is 25-30% higher than that of traditional threshold-based early warning, with a 20% increase in early warning accuracy and a 40% reduction in false alarm rate.

[0105] Fourth, multi-level early warning and trend analysis functions provide maintenance personnel with richer decision-making information. Maintenance personnel can not only understand the current risk status but also grasp the development trend and main influencing factors of the risks, thereby formulating more reasonable maintenance strategies. Practical applications show that this information support has reduced maintenance costs by an average of 15-20% and increased equipment availability by 2-3%.

[0106] Fifth, the edge-cloud collaborative architecture achieves a balance between real-time performance and accuracy. Edge computing nodes are responsible for real-time monitoring and preliminary analysis, while cloud servers handle complex modeling and deep learning. Their collaboration ensures both timely alerts and accurate analysis. This architecture reduces system response time from minutes in traditional cloud processing to seconds, without significant loss of prediction accuracy.

[0107] Finally, the closed-loop control system established in this embodiment realizes the transformation from passive maintenance to proactive maintenance. The system can not only predict the remaining lifespan of equipment but also optimize maintenance strategies to maximize the economic benefits of the equipment. Long-term operation data shows that the failure rate of transmission line equipment using this system is reduced by 40-50%, maintenance costs are reduced by 20-30%, and overall economic benefits are significantly improved.

[0108] Furthermore, the scalability and adaptability of this embodiment are also significant advantages. The system adopts a modular design, allowing for flexible configuration of sensor types and quantities according to different application scenarios, adapting to power transmission equipment of varying scales and types. Online learning capabilities enable the system to continuously adapt to new operating conditions and environmental conditions without requiring manual recalibration, greatly reducing the complexity of deployment and maintenance.

[0109] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting the fatigue life of wind-resistant guy wire devices for power transmission towers, characterized in that, Includes the following steps: S1. Construct a three-dimensional digital twin model of the windproof guy wire device and the associated transmission tower. The windproof guy wire device includes modular connecting blocks, a metal shell, and damping elements. Discretize the model and the surrounding flow field space into a three-dimensional mesh. Define multiple physical channels on the mesh to characterize the physical states of the fluid and solid. The physical channels include the x and y components of the wind speed field, the wind pressure field, the σx, σy, and τxy components of the structural stress tensor field, and the structural displacement field. S2. Perform wind-induced vibration response simulation using a physical process surrogate model based on a multi-scale hierarchical architecture to obtain stress-strain time history data. This step specifically includes: S2.1 The multiple physical channels defined in step S1 are input into the physical process proxy model as independent channels, and each physical channel is independently embedded as a different representation vector on the spatial grid; S2.2 The coupling relationship between different physical channels is modeled by the channel axial self-attention mechanism inside the physical process proxy model to simulate the fluid-structure interaction effect. The channel axial self-attention mechanism learns the interaction between channels by calculating the attention weights between the representation vectors of different physical channels. S2.3 adopts a U-Net-type hierarchical architecture and a shift window self-attention mechanism to perform feature extraction and fusion at multiple scales. Multi-scale information processing is achieved by upsampling and downsampling the representation vector. S2.4 Predicts time step by time using an autoregressive method and outputs the stress-strain time history data of the key fatigue points of the windproof guy wire device within a preset time period; S3. A long-sequence analysis model based on input-dependent state transitions is used to assess fatigue damage in the stress-strain time history data. This step specifically includes: S3.1 Input the stress-strain time history data output in step S2 as an input sequence into the long sequence analysis model; S3.2 Through the input-dependent state transition mechanism built into the long sequence analysis model, the state transition matrix is ​​dynamically adjusted according to the current input signal to achieve adaptive adjustment of the state transition coefficients as the input changes. The state transition matrix is ​​in the form of (A+Bu), where A is the basic state transition matrix, B is the learnable weight matrix, and u is the current input signal. S3.3 Based on the input-dependent state transition mechanism, the liquid convolution kernel is used to explicitly model the autocorrelation of the input stress sequence. The weight of the liquid convolution kernel is associated with the autocorrelation term of the input signal to capture higher-order correlations and combined effects in the stress sequence. S3.4 Directly learn the mapping relationship from stress-strain sequence to fatigue damage increment, and output the cumulative fatigue damage value associated with the time history data; S4. Environmental parameters are endogenized into the physical process proxy model and long sequence analysis model through a deep condition injection mechanism. The environmental parameters include corrosion level, ambient temperature and ultraviolet radiation intensity. The environmental parameters are used as condition inputs to adjust the internal calculations of the model through an adaptive layer normalization mechanism, so that the model can learn the nonlinear effects of the environment on the physical process and damage accumulation. S5. Based on the measured data collected by the field sensors, the pre-trained model described in steps S2 and S3 is fine-tuned online using low-rank adaptive technology, specifically including: S5.1 Vibration and stress data are collected in real time by sensors installed on the modular connecting block of the windproof guy wire device; S5.2 Compare the measured data with the predicted values ​​of the pre-trained model to calculate the loss function; S5.3 Fine-tuning is achieved by inserting a low-rank adapter matrix into the model and training only the adapter matrix, thus enabling continuous self-evolution of the model without losing pre-trained knowledge. S5.4 Periodically run the fine-tuned model, output the continuously updated remaining service life of the windproof guy wire device, and compare it with the preset safety threshold to provide early warning.

2. The method according to claim 1, characterized in that, The shift-window self-attention mechanism in step S2.3 achieves efficient parallel computation of high-resolution grid data by restricting attention calculations to a local window and shifting the window position between adjacent computation layers.

3. The method according to claim 1, characterized in that, The state transition matrix in step S3.2 is in the form of (A+Bu), where A is the basic state transition matrix in step S3.2, B is the learnable weight matrix in step S3.2, and u is the current input signal in step S3.

2.

4. The method according to claim 1, characterized in that, In step S3.3, the liquid convolution kernel models the combined effect of the stress sequence by calculating the autocorrelation term ui·uj and the higher-order term ui·uj·uk of the input signal, where i, j, and k are indices at different times.

5. The method according to claim 1, characterized in that, The adaptive layer normalization mechanism in step S4 encodes environmental parameters as scale parameter γ and offset parameter β, and modulates the normalization layer inside the model to achieve fine control of environmental conditions on the model calculation process.

6. The method according to claim 1, characterized in that, The low-rank adapter matrix in step S5.3 is implemented by decomposing the weight update into the product of two low-rank matrices ΔW=BA, where the ranks of B and A are both less than the dimension of the original weight matrix, thereby achieving efficient fine-tuning of the parameters.

7. The method according to claim 1, characterized in that, The sensor in step S5.1 is a strain gauge, accelerometer, or displacement sensor installed on the modular connecting block of the windproof guy wire device.

8. The method according to claim 1, characterized in that, The physical process proxy model and long sequence analysis model are based on the pre-trained models obtained from historical simulation and measured datasets covering different wind conditions, structural parameters and environmental factors.

9. The method according to claim 1, characterized in that, The independent channel embedding in step S2.1 maintains the independence of different physical quantities and avoids information compression loss by independently mapping each physical channel to a high-dimensional vector representation at each position of the spatial grid.

10. The method according to claim 1, characterized in that, Step S5.4 further includes: establishing a confidence interval assessment of the remaining service life, issuing an early warning signal when the lower bound of the confidence interval is lower than a preset safety threshold, and issuing an emergency replacement signal when the upper bound of the confidence interval is lower than a preset safety threshold.

Images