Wind power tower material crack propagation prediction method and system

Abstract

The present application belongs to the technical field of wind power generation, and specifically relates to a wind power tower material crack propagation prediction method and system, which is suitable for fatigue life evaluation of large supporting structures in the wind energy industry. The method comprises: deploying a multi-source heterogeneous sensor network to collect strain, acoustic emission and environmental load data; constructing a digital twin high-fidelity finite element submodel to obtain a dynamic stress intensity factor; constructing a deep neural network, embedding the Paris law of fracture mechanics in the form of residual into the loss function to form a physical information neural network, and realizing crack propagation rate prediction; using the extended finite element method combined with the level set function and the maximum circumferential stress criterion to dynamically track the three-dimensional space expansion path of the crack; combining Monte Carlo simulation and recursive Bayesian update to evaluate the remaining life and trigger graded early warning. The present application solves the problems of insufficient fusion of physical mechanisms and data-driven in the prior art, low prediction accuracy under small samples, and lack of crack space path tracking.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, specifically to a method and system for predicting crack propagation in wind turbine tower materials. Background Technology

[0002] As a key supporting structure for wind turbine generators, wind turbine towers not only bear the enormous static loads of the nacelle and blades, but also operate for extended periods in extremely complex natural environments, continuously enduring the coupled effects of multiple alternating loads, including random gusts, periodic mechanical vibrations, centrifugal forces, and salt spray chemical corrosion. Under the long-term driving force of these complex stress vectors, the high-performance structural steel used in the towers is highly susceptible to fatigue crack initiation at locations of geometrical abrupt changes, weld heat-affected zones, or microstructural defects. If the crack propagation and evolution process cannot be accurately and in real-time predicted, the accumulation of damage within the structure may lead to irreversible brittle fracture in the later stages of service, or even trigger tower collapse, threatening the safe operation and economic benefits of the wind farm.

[0003] Predicting fatigue crack propagation in metallic materials mainly falls into two categories: physically analytical and statistically driven approaches. A representative physically analytical approach typically focuses on constructing complex residual stress distribution functions and combining them with classical fracture mechanics criteria such as the Walker model to assess crack propagation life by calculating the stress intensity factor amplitude. However, its limitations become increasingly apparent when dealing with ultra-large, highly redundant engineering structures like wind turbine towers.

[0004] The residual stress field generated during the manufacturing, transportation and installation of wind turbine towers has strong spatial non-uniformity and time-varying characteristics. Traditional analytical models are often based on idealized quasi-static assumptions, which makes it difficult to capture the nonlinear damage accumulation of the material microstructure under random amplitude loads. This leads to significant fluctuations in the prediction accuracy of the models when dealing with extreme weather conditions or complex multiaxial stress states.

[0005] Existing prediction systems suffer from significant gaps in the deep integration of mechanical mechanisms and real-time monitoring data, making it impossible to simultaneously track the spatiotemporal characteristics of crack evolution under highly uncertain service environments. The lack of physical mechanisms and weak spatial dimension prediction have become core obstacles restricting the development of intelligent and high-precision health monitoring of wind turbine tower structures. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for predicting crack propagation in wind turbine tower materials. This invention solves the technical problems of insufficient prediction accuracy under small sample conditions due to the lack of deep integration of physical and mechanical mechanisms and data-driven approaches in existing wind turbine tower crack prediction technologies, as well as the lack of dynamic tracking capability for crack propagation paths in three-dimensional space, which only focuses on the scalar of propagation rate.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for predicting crack propagation in wind turbine tower materials includes the following steps: Step 1, Multi-source heterogeneous service data acquisition and preprocessing: Deploy a high-precision multi-directional strain monitoring matrix, acoustic emission sensor array and environmental load monitoring unit at the preset key monitoring parts of the wind turbine tower to acquire strain, acoustic emission and environmental load data in real time, and obtain time-series feature dataset after preprocessing. Step 2: Construct a high-fidelity finite element sub-model based on digital twin technology: Based on the original design parameters of the wind turbine tower, the material constitutive relationship, and the measured residual stress distribution data, establish a global finite element model and a local refined crack analysis model for key monitoring parts. Drive the sub-model boundary through the displacement field of the global model to obtain the dynamic stress intensity factor of the local area. Step 3: Construct a physical information neural network crack propagation prediction model: The physical information neural network consists of a deep feedforward neural network and a physical constraint operator. The input vector includes time-series feature data and the number of load cycles, and the output vector is the crack length and propagation rate. The loss function introduces a physical constraint residual term based on Paris's law of fracture mechanics, so that the model can follow the physical evolution law of crack propagation while fitting the measured data. Step 4, simulation of crack spatial propagation path based on extended finite element method: using the crack propagation rate predicted in step 3 as the driving force, an extended finite element operator is introduced into the local refined crack analysis model, the crack geometry is described by the level set function, the crack propagation deflection angle is determined according to the maximum circumferential stress criterion, the crack front coordinates are dynamically updated, and the three-dimensional propagation path of the crack is tracked. Step 5, Structural Integrity Assessment and Remaining Life Prediction: Combine the crack propagation rate predicted in Step 3 with the spatial path simulated in Step 4 to calculate the critical relationship between the stress intensity factor corresponding to the real-time crack size and the material fracture toughness. When the preset threshold is reached, a safety warning is triggered, and a remaining life assessment report is generated.

[0008] Furthermore, in step one, the preset key monitoring areas include the heat-affected zone of the tower weld, the flange connection, and the variable cross-section transition zone; the high-precision multi-directional strain monitoring matrix is ​​realized through fiber optic grating sensors with a wavelength range of 1525nm to 1565nm and a strain resolution of 1με; the sensor spacing is determined according to the stress gradient of the tower's stressed skin, and the density within 50mm of the weld edge is not less than 2 sensors per 100cm²; and the sensor surface is covered with a EPDM rubber protective layer.

[0009] Furthermore, in step one, the acoustic emission sensor array uses a piezoelectric ceramic sensor with a response frequency range of 20kHz to 400kHz; a high-viscosity coupling agent is applied to the sensor mounting location, and it is fixed to the tower surface using stainless steel clamps; the environmental load monitoring unit includes an ultrasonic anemometer installed at the top of the tower and a salt spray concentration sensor installed on the outer wall of the tower, used to provide wind load vector data and corrosion environment correction parameters.

[0010] Furthermore, in step one, the data preprocessing includes preprocessing based on an adaptive wavelet threshold denoising algorithm to remove random environmental noise, and data dimensionality reduction based on principal component analysis: by performing eigenvalue decomposition on the covariance matrix composed of multi-channel strain signals, principal component components with a cumulative contribution rate exceeding 95% are extracted as input feature vectors for subsequent physical information neural networks.

[0011] Furthermore, in step two, the material constitutive relation corresponds to Q355ND high-performance structural steel; the full-scale global finite element model is discretized using four-node reduced integral shell elements; the local refined crack analysis model is meshed using eight-node hexahedral linear reduced integral elements, and singular elements are used to handle stress singularity in the preset crack tip region, and the boundary range of the sub-model is set to more than 10 times the preset crack length.

[0012] Furthermore, in step two, the method for obtaining the measured residual stress distribution data of the material is as follows: X-ray diffraction is used to sample and measure typical welds of the tower to obtain the deep residual stress distribution law, and the measured residual stress field is pre-placed in the local refined crack analysis model through the initial stress field import method.

[0013] Furthermore, in step three, the loss function of the physical information neural network... Defined as: in, The total loss function is dimensionless. The weights of the data fitting terms are dimensionless. The mean square error between the network prediction and the monitoring data is dimensionless. The weighting coefficients for the physical constraint terms are dimensionless. The physical constraint residual is dimensionless; Specifically defined as: In the formula, The number of samples in the current training batch is dimensionless. This is a sample index, dimensionless. The first prediction of the neural network Crack propagation rate of each sample, in mm / cycle; These are parameters related to material fatigue characteristics. For the first The amplitude of the equivalent stress intensity factor for each sample; This is the fatigue property index of the material.

[0014] Furthermore, in step three, the amplitude of the equivalent stress intensity factor... Calculated using Walker correction terms: In the formula, The equivalent stress intensity factor amplitude is expressed in MPa·m^{1 / 2}. The maximum stress intensity factor, in MPa·m^{1 / 2}, is calculated by the local refined crack analysis model constructed in step two. The stress ratio is defined as the ratio of the minimum stress to the maximum stress, and is dimensionless. The material sensitivity coefficient is dimensionless.

[0015] Furthermore, in step three, the physical information neural network adopts a 6-layer hidden layer structure, with each layer containing 128 neurons, and the activation function is the Tanh function with second-order continuous derivative characteristics; the physical constraint operator is constructed by calculating the partial derivatives of the output vector with respect to the physical parameters using automatic differentiation technology under the deep learning framework.

[0016] Furthermore, in step four, the deflection angle The calculation formula is: In the formula, The crack propagation deflection angle is expressed in radians. and These are the Type I and Type II stress intensity factors, respectively, with units of MPa·m^{1 / 2}. This is achieved by introducing an equivalent integral segment method into the locally refined crack analysis model. The integral calculation program was used to separate the coordinates; the coordinates of the crack front were updated using the second-order Runge-Kutta method.

[0017] Furthermore, in step five, the triggering condition for the safety warning is specifically: when the stress intensity factor calculated in real time... Achieving material fracture toughness A safety warning is triggered when the crack length reaches 0.85 times the value of the crack. During the remaining life prediction process, a recursive Bayesian update algorithm is used to dynamically adjust the parameters in Paris's law based on real-time monitored crack length data. and .

[0018] This invention also discloses a crack propagation prediction system for wind turbine tower materials applied to the method described above, comprising: The multi-source sensor data acquisition module is set up on the wind turbine tower site. It includes a high-precision multi-directional strain monitoring matrix, an acoustic emission sensor array, and an environmental load monitoring unit deployed at key monitoring locations. It is used to acquire strain, acoustic emission, and environmental load data in real time and perform data preprocessing based on adaptive wavelet threshold denoising and principal component analysis. The digital twin simulation engine is pre-built with a full-scale global finite element model of the tower and sub-models of key parts, which are used to solve the dynamic stress field and stress intensity factor of key parts of the tower in real time. The physical information fusion prediction module integrates a physical information neural network. Its input end is connected to the output end of the multi-source sensor data acquisition module to receive the time-series feature dataset. At the same time, it is connected to the digital twin simulation engine to obtain mechanical parameters and perform crack propagation rate prediction. The spatial path tracking module is connected to the physical information fusion prediction module. Based on the extended finite element method, it uses the predicted crack propagation rate to drive the geometric evolution of the crack surface and dynamically reconstructs the three-dimensional spatial propagation path of the crack. The early warning decision support module connects the spatial path tracking module and the physical information fusion prediction module. It is used to calculate the critical relationship between real-time crack size and material fracture toughness, output a remaining life prediction report, and perform graded early warnings according to preset thresholds.

[0019] Furthermore, the physical information fusion prediction module adopts a multi-task learning architecture, synchronously outputting the crack propagation rate and the uncertainty range at a 95% confidence level; the multi-source sensor data acquisition module transmits the preprocessed data to the edge computing node through the industrial Ethernet protocol, and the edge computing node and the digital twin simulation engine at the back end of the server conduct bidirectional data communication to realize real-time correction of load conditions.

[0020] Furthermore, the spatial path tracing module also includes an adaptive mesh re-division unit, which dynamically reconstructs the finite element mesh of the local refined crack analysis model during crack propagation based on the posterior error estimation results.

[0021] Furthermore, in step three, when constructing the physical information neural network crack propagation prediction model, a dynamic Bayesian adaptive sub-step for physical parameters is also included: The material fatigue characteristic parameters in Paris's law Sum of Indices Defined as an environmentally corrosive factor With cumulative damage state The prior distributions of the dynamic random variables driven by the system are set as log-normal and normal distributions, respectively. Based on variational Bayesian inference, a Bayesian physical information neural network is constructed. The uncertainty in the time-series feature dataset obtained from the preprocessing in step one is quantized into the probability distribution of the input features using a Gaussian mixture model, thereby learning the dynamic random variable. and The posterior distribution parameters, including the mean ,variance mean ,variance ; Among them, the environmental corrosion factor Salt spray concentration ,temperature and relative humidity It is obtained through nonlinear mapping of a pre-trained corrosion acceleration factor network; The cumulative damage state Energy accumulation value in acoustic emission eigenvectors The integral of the crack propagation rate prediction is updated in real time. By sampling from the posterior distribution Set parameters to obtain crack length With expansion rate The predicted distribution is calculated, and the variance of the predicted distribution is used as the uncertainty interval at a 95% confidence level, and output synchronously.

[0022] Furthermore, in step five, the structural integrity assessment and remaining lifetime prediction further include an uncertainty propagation recursive update sub-step: The crack length prediction distribution for each increment step of the above output. As input to the Monte Carlo simulation, where Indicates the first Group sampling parameters, In each load cycle increment step of the Monte Carlo simulation, a value is randomly sampled based on the probability distribution of the current crack length. Combined with the stress intensity factor distribution calculated in step two, the equivalent stress intensity factor amplitude for the next increment step is calculated. The probability distribution; Based on the recursive Bayesian update algorithm, the measured crack length data obtained from real-time monitoring is... As observed values, the dynamic random variable is updated. and The posterior distribution is updated using the following formula: in, for Material parameters after real-time monitoring data are constantly integrated and The posterior distribution of; From the initial moment to The complete set of monitoring data at any given time; Let be the likelihood function, representing the given material parameters. and The measured crack length was observed at that time. The probability of; The measured crack length data is obtained through real-time monitoring, and the unit is mm; for Posterior distribution of material parameters at time t; The final output of the remaining lifetime prediction result is a result containing the mean. ,variance The probability distribution of quantiles is used to generate maintenance decision reports with risk levels.

[0023] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a dual-drive prediction architecture integrating a physical information neural network (PIN) and the extended finite element method (EPM). The PIN embeds Paris's law of fracture mechanics and its modified model into a deep learning loss function as residual constraints. This ensures that the neural network consistently follows the physical evolution of crack propagation when fitting limited monitoring data, solving the fundamental problems of overfitting and weak extrapolation ability in pure data-driven models under small sample conditions. The EPM describes the crack surface geometry through level set functions and dynamically calculates the crack deflection angle using the maximum circumferential stress criterion, achieving continuous tracking of the crack's three-dimensional spatial propagation path. The coupling of the PIN and EPM creates a closed-loop feedback between crack propagation rate prediction and spatial path simulation, overcoming the limitation of existing technologies that only focus on the scalar propagation rate while neglecting the dimensionality of spatial path evolution.

[0024] This invention simultaneously acquires three-dimensional dynamic stress, crack initiation elastic wave signals, and corrosion environment data. After adaptive wavelet threshold denoising and principal component analysis dimensionality reduction, a time-series feature dataset is formed. The digital twin simulation engine, through the nested driving of a global finite element model and local sub-models, maps macroscopic loads to dynamic stress intensity factors at crack tips in real time, providing accurate mechanical boundary conditions for the physical information neural network. The bidirectional feedback between multi-source monitoring data and the high-fidelity mechanical model enables the prediction system to respond to extreme weather and sudden load changes, improving the real-time performance and robustness of predictions under varying operating conditions.

[0025] This invention introduces a Bayesian physical information neural network and a recursive Bayesian update mechanism. The material fatigue characteristic parameters in Paris's law are defined as dynamic random variables driven by environmental corrosion factors and cumulative damage states. The posterior distribution of the parameters is learned through variational Bayesian inference, enabling the model to adapt to abrupt changes in the corrosion environment and material performance degradation. The recursive Bayesian update algorithm dynamically corrects the parameter distribution using real-time monitored crack length data, achieving adaptive calibration between the model and the actual structural state. The introduction of the Bayesian framework expands the prediction results from a single numerical value to a probability distribution including mean and variance, quantifying the entire chain of uncertainty from input data and physical model to remaining life prediction uncertainty at a 95% confidence level.

[0026] This invention proposes an enhancement technique combining a modified extended finite element method (FEA) with adaptive mesh refactoring. A fracture process zone basis function based on the Gurson-Tvergaard-Needleman model is introduced into the displacement interpolation function, improving the accuracy of the description of the nonlinear stress-strain field at the crack tip in the weld heat-affected zone. An adaptive mesh refactoring strategy based on posterior error estimation dynamically refines the mesh in the crack lead region, solving the problem of decreased computational accuracy caused by mesh distortion during crack propagation. The secondary correction of the micro-stress field in the weld zone using crystal plastic finite element method couples macroscopic loads, welding residual stresses, and microcrystalline structure effects to the stress intensity factor calculation, making the crack propagation driving force closer to the physical essence. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0028] Figure 1 This is an overall flowchart of the method described in this invention.

[0029] Figure 2 This is a simplified diagram of the system architecture described in this invention.

[0030] Figure 3 This is one of the system operation interface diagrams described in this invention.

[0031] Figure 4 This is the second diagram of the system operation interface described in this invention.

[0032] Figure 5 This is the third diagram of the system operation interface described in this invention. Detailed Implementation

[0033] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0034] The following is in conjunction with the appendix Figures 1-5 The embodiments of the present invention will be described in detail below.

[0035] Example 1: This example discloses a crack propagation prediction system for wind turbine tower materials, including: The multi-source sensor data acquisition module is set up on the wind turbine tower site. It includes a high-precision multi-directional strain monitoring matrix, an acoustic emission sensor array, and an environmental load monitoring unit deployed at key monitoring locations. It is used to acquire strain, acoustic emission, and environmental load data in real time and perform data preprocessing based on adaptive wavelet threshold denoising and principal component analysis. The digital twin simulation engine is pre-built with a full-scale global finite element model of the tower and sub-models of key parts, which are used to solve the dynamic stress field and stress intensity factor of key parts of the tower in real time. The physical information fusion prediction module integrates a physical information neural network. Its input end is connected to the output end of the multi-source sensor data acquisition module to receive the time-series feature dataset. At the same time, it is connected to the digital twin simulation engine to obtain mechanical parameters and perform crack propagation rate prediction. The spatial path tracking module is connected to the physical information fusion prediction module. Based on the extended finite element method, it uses the predicted crack propagation rate to drive the geometric evolution of the crack surface and dynamically reconstructs the three-dimensional spatial propagation path of the crack. The early warning decision support module connects the spatial path tracking module and the physical information fusion prediction module. It is used to calculate the critical relationship between real-time crack size and material fracture toughness, output a remaining life prediction report, and perform graded early warnings according to preset thresholds.

[0036] The physical information fusion prediction module adopts a multi-task learning architecture, synchronously outputting the crack propagation rate and the uncertainty range at a 95% confidence level; the multi-source sensor data acquisition module transmits the pre-processed data to the edge computing node through the industrial Ethernet protocol, and the edge computing node and the digital twin simulation engine at the back end of the server conduct bidirectional data communication to realize real-time correction of load conditions.

[0037] The spatial path tracing module also includes an adaptive mesh re-division unit, which dynamically reconstructs the finite element mesh of the local refined crack analysis model during crack propagation based on the posterior error estimation results.

[0038] This embodiment also discloses a method for predicting crack propagation in wind turbine tower materials, the method comprising the following steps: Step 1: Acquisition and preprocessing of multi-source heterogeneous service data.

[0039] High-precision multi-directional strain monitoring matrices, acoustic emission sensor arrays, and environmental load monitoring units are deployed at key monitoring locations on wind turbine towers, including the heat-affected zone of tower welds, flange connections, and variable cross-section transition zones.

[0040] The high-precision multi-directional strain monitoring matrix is ​​implemented using fiber optic grating sensors with a wavelength range of 1525nm to 1565nm and a strain resolution of 1με. These sensors are used to acquire real-time three-dimensional dynamic stress evolution data of the tower surface. The sensor spacing is determined based on the stress gradient of the tower's stressed skin, with a density of no less than 2 sensors per 100cm² within a 50mm radius of the weld edge. The sensor surface is covered with a EPDM rubber protective layer.

[0041] The acoustic emission sensor array employs piezoelectric ceramic sensors with a response frequency range of 20kHz to 400kHz, used to capture elastic wave signals released during crack initiation and propagation. The installation positions of the acoustic emission sensors correspond to the positions of the strain monitoring matrix. A high-viscosity coupling agent is coated between the sensors and the tower surface, and they are fixed with stainless steel clamps.

[0042] The environmental load monitoring unit includes an ultrasonic anemometer installed at the top of the tower and a salt spray concentration sensor installed on the outer wall of the tower, which are used to simultaneously collect data on wind speed, wind direction, ambient temperature and atmospheric salt spray concentration.

[0043] The acquired raw data undergoes preprocessing using an adaptive wavelet threshold denoising algorithm to remove random environmental noise, resulting in a denoised time-series signal. Further, feature parameter analysis is applied to the acoustic emission signal to extract ring count, energy, amplitude, rise time, and duration to form the acoustic emission feature vector. Principal component analysis eigenvalue decomposition is performed on the covariance matrix of the multi-channel strain signals, extracting principal component components with a cumulative contribution rate exceeding 95% as the strain feature vector. The acoustic emission feature vector, strain feature vector, and environmental load data are then normalized to form a time-series feature dataset.

[0044] Step 2: Construct a high-fidelity finite element sub-model based on digital twin technology.

[0045] Based on the original design parameters, material constitutive relations, and measured residual stress distribution data of the wind turbine tower, a full-scale global finite element model of the wind turbine tower was established. The material constitutive relations correspond to Q355ND high-performance structural steel, with an elastic modulus of 206 GPa, a Poisson's ratio of 0.3, and a yield strength of 355 MPa. The full-scale global finite element model was discretized using four-node reduced integral shell elements.

[0046] Based on this, a localized refined crack analysis model is constructed for the key monitoring areas identified in Step 1 using sub-model technology. This localized refined crack analysis model uses eight-node hexahedral linear reduced integral elements for mesh generation, and singular elements are used to handle stress singularities in the preset crack tip region. The boundary range of the sub-model is set to be more than 10 times the preset crack length to eliminate the influence of boundary effects on the stress field distribution.

[0047] The residual stress distribution data measured by the material were obtained by sampling and measuring typical welds of the tower using X-ray diffraction to obtain the deep residual stress distribution law, and then pre-setting the measured residual stress field into the local refined crack analysis model by importing the initial stress field.

[0048] By using the displacement field of the global model as the boundary driving condition of the sub-model, the accurate transfer of global macroscopic loads to local microscopic stress fields is achieved, and the dynamic stress intensity factor of the local region is obtained. Stress intensity factors include Type I stress intensity factors. and Type II stress intensity factor .

[0049] Step 3: Construct a physical information neural network crack propagation prediction model.

[0050] The physical information neural network model consists of a deep feedforward neural network and physical constraint operators. The input vector of the deep feedforward neural network includes time-series strain characteristics, acoustic emission characteristics, and the number of wind load cycles. And environmental corrosion factors, with the output vector being the crack length. and expansion rate .

[0051] In the design of the loss function for the neural network, a physical constraint term based on Paris's law and its modified model in fracture mechanics is introduced. This physical constraint term takes the form of a residual, constraining the difference between the derivative of the crack length predicted by the neural network with respect to the number of cycles and the theoretical propagation rate based on the stress intensity factor amplitude. The total loss function... Defined as: in: The total loss function is dimensionless. The weight coefficients of the data fitting terms are dimensionless. The mean square error between the network prediction and the monitoring data is dimensionless. The weighting coefficients for the physical constraint terms are dimensionless. The physical constraint residual is dimensionless.

[0052] The physical constraint residual Specifically defined as: in: The number of samples in the current training batch is dimensionless. The first prediction of the neural network Crack propagation rate of each sample, in mm / cycle; These are the fatigue characteristic parameters of the material under inert conditions, expressed in mm / cycle / (MPa·m1 / 2)m. For the first The amplitude of the equivalent stress intensity factor for each sample, in MPa·m^{1 / 2}; This is the material fatigue property index, which is dimensionless.

[0053] The amplitude of the equivalent stress intensity factor Calculated using Walker correction terms: in: The equivalent stress intensity factor amplitude is expressed in MPa·m^{1 / 2}. The maximum stress intensity factor, in MPa·m^{1 / 2}, is calculated by the local refined crack analysis model constructed in step two. The stress ratio is defined as the ratio of the minimum stress to the maximum stress, and is dimensionless. The material sensitivity coefficient is dimensionless.

[0054] The physical information neural network employs a 6-layer hidden layer structure, with each layer containing 128 neurons. The activation function chosen is the Tanh function, which has a second-order continuous derivative. The physical constraint operator is constructed by calculating the partial derivatives of the output vector with respect to the physical parameters using automatic differentiation techniques within a deep learning framework.

[0055] The weight coefficients are dynamically adjusted using the backpropagation algorithm and the Adam optimizer. and This allows the model to fit the measured data while predicting the crack propagation rate according to the laws of fracture mechanics, thus achieving high-precision extrapolation prediction under small sample conditions.

[0056] Step 4: Simulation of crack spatial propagation path based on extended finite element method.

[0057] Using the crack propagation rate predicted in step three as the time step driving force, an extended finite element operator is introduced into the locally refined crack analysis model. The geometry of the crack surface and crack lead is described by level set functions, where... The function represents the location of the crack surface. The function represents the location of the crack leading edge.

[0058] Determine the deflection angle for crack propagation based on the maximum circumferential stress criterion. : in: The crack propagation deflection angle is expressed in radians. It is a type I stress intensity factor, with units of MPa·m^{1 / 2}; It is a type II stress intensity factor, with units of MPa·m^{1 / 2}.

[0059] By introducing the equivalent integral segment method into the local refined crack analysis model The integral calculation program is obtained separately. and .

[0060] The coordinate update of the crack front is achieved by using the second-order Runge-Kutta method. By iteratively calculating the coordinates of the crack front in each incremental step, the spatial geometric topology of the crack is dynamically updated, thereby enabling the tracking of the three-dimensional propagation path of the crack in the complex weld structure of the tower.

[0061] Step 5: Structural integrity assessment and remaining life prediction.

[0062] By combining the crack propagation rate predicted in step three with the spatial path simulated in step four, the real-time stress intensity factor corresponding to the real-time crack size is calculated. Fracture toughness of tower materials The critical relationship between them. For Q355ND steel, fracture toughness... Values .

[0063] When the real-time stress intensity factor achieve A safety warning is triggered when the crack length reaches 0.85 times the critical size. Through Monte Carlo simulation of the wind load spectrum under future service conditions, 10,000 evolution simulations are performed under a preset wind load probability density distribution to predict the remaining number of cycles required for the crack to reach its critical size, generating a tower remaining life assessment report. During the remaining life prediction process, a recursive Bayesian update algorithm is used to dynamically correct the parameters in Paris's law based on real-time monitored crack length data. and .

[0064] In practice, a refined sensor deployment is implemented to address the structural characteristics of large wind turbine towers, particularly in areas with significant stress concentration, such as the heat-affected zone of tower welds, flange connections, and variable cross-section transition zones.

[0065] Taking a 120-meter-high tower made of Q355ND high-performance structural steel as an example, its high-precision multi-directional strain monitoring matrix is ​​realized through fiber optic grating sensors. These sensors are arranged within a 50mm range of the weld edge, with a density controlled at 2 sensors per 100cm², to capture local strain fluctuations under stress gradients. The center wavelength of the fiber optic grating is selected between 1525nm and 1565nm, achieving a strain resolution of 1με, ensuring that minute stress redistribution phenomena can be effectively recorded.

[0066] To ensure long-term stability in extreme salt spray environments such as nearshore areas, the sensor surface is covered with a protective layer of EPDM rubber. Meanwhile, the acoustic emission sensor array uses piezoelectric ceramic material with a response frequency range covering 20kHz to 400kHz. By coating the sensor and tower surfaces with a high-viscosity coupling agent and supplementing it with stainless steel clamps, the elastic wave signal released at the moment of crack initiation is captured.

[0067] In the data preprocessing stage, due to the complex high-frequency vibration noise in the wind farm environment, this invention employs an adaptive wavelet threshold denoising algorithm. The system performs five-level wavelet decomposition on the original time-series signal, selects the db4 wavelet basis function, and dynamically adjusts the soft threshold according to the noise level of each level, thereby filtering out environmental background noise while retaining crack-related characteristic signals. For the acoustic emission signal, the feature parameter analysis method is used to extract ring count, energy, amplitude, rise time, and duration from the denoised waveform signal to construct the acoustic emission feature vector.

[0068] Furthermore, principal component analysis (PCA) is used to reduce the dimensionality of the multi-channel strain signals. By eigenvalue decomposition of the covariance matrix, principal component components with a cumulative contribution rate exceeding 95% are extracted. This step transforms the monitoring signals into principal eigenvectors representing the true stress state of the structure, providing input for the subsequent convergence of the physical information neural network.

[0069] Constructing a high-fidelity finite element sub-model based on digital twin technology is crucial for achieving a seamless connection between macroscopic and microscopic mechanics. This implementation first establishes a full-scale global finite element model based on the original CAD model and measured geometric dimensions of the wind turbine tower. The main body of the tower is discretized using four-node reduced integral shell elements. To introduce the initial damage state of the material, this invention utilizes X-ray diffraction to sample and measure typical welds of the tower, obtaining the distribution law of deep residual stress, and importing it as the initial stress field into the finite element model. For key monitoring areas, the system automatically extracts local regions and establishes a three-dimensional solid element model using sub-model technology. This sub-model uses eight-node hexahedral linear reduced integral elements for meshing, ensuring that the boundary range of the sub-model is at least 10 times the preset crack length. In the crack leading edge region, singular elements are arranged using mesh refinement technology to simulate the crack tip. Stress singularity. The dynamic displacement field of the global model is mapped to the boundary nodes of the sub-model in real time, thereby solving for the dynamic stress intensity factor that reflects the actual service load. and .

[0070] A physical information neural network prediction model was constructed. This model embeds physical laws into the loss function of a deep feedforward neural network. The neural network architecture consists of one input layer, six hidden layers (128 neurons each), and one output layer. The input vector integrates strain features after dimensionality reduction via principal component analysis, acoustic emission feature vectors, and the cumulative wind load cycle count. And the environmental corrosion factor measured by the salt spray sensor. The output vector is the crack length. and expansion rate To ensure that the model follows the evolutionary logic of fracture mechanics, this invention introduces Paris's law and its Walker correction term into the loss function in the form of residuals.

[0071] During training, the loss function Defined as a weighted sum of monitored value deviation and physical residual, it is optimized using the Adam optimizer through multi-parameter collaborative optimization. Weighting coefficients The initial setting is 0.5. The initial value is set to 0.5, and it is gradually increased as the number of training rounds increases. The weighting of these weights allows the model to place greater emphasis on the constraints of physical laws during the later stages of training. This ensures that when monitoring data experiences abnormal fluctuations due to interference in actual engineering projects, the model can still provide predictions based on the constraints of physical laws.

[0072] To obtain the three-dimensional propagation trajectory of cracks in complex welded structures of towers, this invention introduces the extended finite element method (EPM). Unlike the traditional finite element method, the EPM adds an enhancement function to the element displacement interpolation function, allowing cracks to freely pass through elements without remeshing. The system utilizes level set functions to dynamically track the crack surface and crack lead.

[0073] in, The function represents the location of the crack surface. The function represents the location of the crack tip. The direction of crack propagation deflection is determined based on the maximum circumferential stress criterion, which utilizes Type I and Type II stress intensity factors (…). and The calculations were performed. After each load increment step, the system updated the crack front coordinates using the second-order Runge-Kutta method to ensure the numerical stability of the path simulation under variable amplitude loading conditions. This process represents a leap from single propagation rate prediction to three-dimensional geometric evolution tracking.

[0074] Finally, the system compares the predicted real-time crack depth with the fracture toughness of the tower material Q355ND. A comparison was made. Monte Carlo simulations were performed 10,000 times under a preset wind load probability density distribution to determine when the crack reached its critical size. The remaining number of cycles is calculated. The final assessment report includes the predicted remaining life of the crack, as well as the uncertainty interval (95% confidence level) based on the multi-task learning architecture, providing a basis for decision-making by operations and maintenance personnel.

[0075] To further illustrate the technical solution of this invention, a simulation comparison experiment was conducted. The root weld of a 5MW wind turbine tower was selected as the research object. The tower height was 120m, the material was Q355ND, the elastic modulus was 206GPa, the Poisson's ratio was 0.3, and the initial pre-existing crack length was 2mm. A global finite element model and local sub-models of the tower were established using Abaqus software. The load spectrum was generated based on the normal turbulence model in the IEC 61400-1 standard, containing 100,000 cycles of loading. The monitoring data from the first 5,000 cycles were used as training samples, and the data from the last 95,000 cycles were used as the prediction verification interval.

[0076] This embodiment employs the fusion prediction method of the physical information neural network and the extended finite element method described in this invention. The physical information neural network uses 6 hidden layers, with 128 neurons in each layer, 500 training epochs, a learning rate of 0.001, and a loss function weight. The initial value was 0.5. The initial value is 0.5, increasing every 100 rounds. 0.1.

[0077] Comparative Example 1: A pure data-driven prediction method based on Long Short-Term Memory (LSTM) networks is used, without introducing physical constraints. The LTM network has a two-layer structure with 128 units per layer, 500 training rounds, and a learning rate of 0.001.

[0078] Comparative Example 2: Using the traditional theoretical analytical method based on Paris's law, the parameters... , Provided by the materials manual, without dynamic correction based on real-time monitoring data.

[0079] The simulation was conducted under a simulated variable-amplitude wind load spectrum. The sample size of the monitoring data was set to a small sample condition (including only the data from the first 5000 cycles). The simulation results are shown in Table 1.

[0080] Table 1: Comparison of performance indicators of various prediction methods under small sample amplitude loads; As can be seen from the data analysis in Table 1, Example 1 provided by this invention outperforms the comparative example in all core indicators. Due to the lack of physical constraints and limited data, Comparative Example 1's Long Short-Term Memory network exhibits significant errors in the extrapolation prediction stage, with an average absolute percentage error reaching 11.85%. While Comparative Example 2 follows physical laws, its static mechanical parameters lead to deviations between the predicted results and the actual evolution trend because it cannot consider environmental corrosion, material inhomogeneity, and complex changes in real-time loads during service. In contrast, this invention, by embedding Paris's law into the neural network in residual form, achieves a prediction error of 2.13% under small sample conditions. Regarding path simulation, the deviation between the crack deflection path obtained by this invention using the extended finite element method and the baseline finite element simulation path is 0.52 mm.

[0081] Further refining the system's hardware composition, the fiber optic grating demodulator in the multi-source sensor data acquisition module employs a parallel synchronous acquisition architecture based on a tunable laser to ensure strict time-domain alignment of signals from each monitoring point. The data transmission layer uses the industrial Ethernet protocol, ensuring real-time transmission of MB-level data traffic to edge computing nodes. The high-fidelity model pre-built into the digital twin simulation engine not only includes the linear elastic stage in its material constitutive relations but also describes the cyclic plastic behavior of the material at the crack tip by introducing the Ramberg-Osgood relation. During training, the physical information fusion prediction module directly calculates the partial derivatives of the neural network output with respect to the input load using automatic differentiation technology implemented in the TensorFlow backend, thereby constructing physical residual terms and avoiding the truncation errors caused by traditional numerical differentiation. The spatial path tracking module utilizes a method based on equivalent integral segments... The integral calculation program separates the stress intensity factors of each type in each increment step.

[0082] In specific engineering applications, for Q355ND high-strength steel towers, their fracture toughness... The measured value is The system's warning trigger logic is set as follows: when the real-time calculated stress intensity factor... Exceed At this time, the early warning decision support module sends a Level 1 orange warning to the operation and maintenance terminal through the industrial cloud. Simultaneously, the system automatically retrieves the load history from the past 24 hours, analyzes whether there are any abnormal gust wind conditions that could accelerate crack propagation, and provides a suggested operating speed for reduced load operation.

[0083] Example 2: Based on Example 1, this example improves the construction and training method of the physical information neural network, aiming to address the problem of accumulated prediction uncertainty caused by the time-varying nature of Paris law parameters under strong noise interference and variable amplitude load.

[0084] The core of this embodiment lies in fusing variational Bayesian inference with physical constraint operators to construct a Bayesian physical information neural network that can output prediction confidence intervals and has parameter adaptive capabilities.

[0085] Step 1: Multi-source heterogeneous data acquisition and uncertainty quantification preprocessing. This step adds a data uncertainty quantification step to the existing Example 1. In addition to the multi-source data acquisition described in Example 1, this example uses an adaptive Gaussian mixture model to fit the probability density of the denoised time-series signal for each sensor channel during the data preprocessing stage.

[0086] The specific operation is as follows: For a time window consisting of 1000 consecutive sampling points for each strain gauge channel, a mixture model containing K Gaussian components is fitted using the expectation-maximization algorithm (K is automatically determined by the Bayesian information criterion). The fitted Gaussian mixture model parameters (mean) are... ,variance Weight This provides the optimal estimate of the strain value at that moment. Its variance The heteroscedasticity resulting from sensor noise, environmental vibration, and measurement uncertainty was quantified. This variance serves as an independent dimension (i.e., the uncertainty of the input features) for subsequent neural network input. This preprocessing transforms the raw signal into a probabilistic feature vector, enabling the subsequent neural network to utilize the uncertainty information of the input data, providing a foundation for Bayesian inference.

[0087] Step 2: Construct a physical information neural network with fused variational Bayesian layers; The core of this embodiment is to construct a Bayesian physical information neural network, whose architecture is similar to the deep feedforward network in Embodiment 1, but a Bayesian layer is introduced in the hidden layers.

[0088] Bayesian layer: This layer assigns weights to each network layer. and bias It can be considered as a random variable that follows a Gaussian distribution, i.e. , The parameters of these distributions are learned through variational inference. , ( ), rather than a fixed value.

[0089] Loss function reconstruction: Total loss function Based on Example 1, a KL divergence term was added to regularize the model complexity: in: is the total loss function of the Bayesian physical information neural network, which is dimensionless; The weighting coefficients for the data likelihood term are dimensionless. It is the data likelihood term in the weighted posterior distribution The expected value is dimensionless. The weighting coefficients for the physical constraint terms are dimensionless. It is the physical constraint residual in the weighted posterior distribution The expected value is dimensionless. is the weighting coefficient of the KL divergence term, which is dimensionless; It is an approximate posterior. with prior The KL divergence between them is dimensionless and serves as a regularization term; This is the approximate posterior distribution of the weights. Its hyperparameters; The prior distribution of the weights is usually taken as the standard normal distribution. .

[0090] Improved Physical Constraints: Walker Modification of Paris's Law and Corrosion Coupling. This paper addresses the corrosion fatigue problem of Q355ND steel in nearshore environments using physical constraint operators. Paris's law is extended to include a corrosion accelerator: in: The crack propagation rate is based on physical constraints and is expressed in mm / cycle. These are fatigue characteristic parameters under corrosive environments, with units of mm / cycle / (MPa·m{1 / 2})m; The equivalent stress intensity factor amplitude is expressed in MPa·m^{1 / 2}. This is the material fatigue property index, which is dimensionless.

[0091] Corrosion accelerator Defined as: in: This represents the Paris law coefficient for materials in an inert environment, with units of mm / cycle / (MPa·m{1 / 2})m; The coefficient of sensitivity of the material to corrosion is dimensionless and is calibrated by laboratory corrosion fatigue tests. For the salt spray concentration (Unit: mg / m³), Temperature (unit: °C) and relative humidity The nonlinear function (dimensionless, expressed as a percentage) is learned through a small, independent neural network that has been pre-trained on laboratory corrosion fatigue data.

[0092] Step 3: Crack propagation prediction and remaining life uncertainty propagation based on Bayesian model averaging; When using a trained Bayesian physical information neural network for prediction, for the input at each time step, the posterior distribution of the weights is used... Medium sampling Group weight parameters ( ),get Different prediction results (crack length) and expansion rate The final predicted value is this. The Monte Carlo average of the results, and the prediction variance constitutes the 95% confidence interval.

[0093] In remaining lifetime prediction, this embodiment uses the crack length prediction distribution (mean and variance) of each increment step output by the Bayesian physical information neural network as input to the Monte Carlo simulation. At each increment step, the Monte Carlo simulation randomly samples a value based on the probability distribution of the current crack length and calculates the stress intensity factor distribution for the next increment step accordingly.

[0094] This process is repeated until the final remaining lifetime prediction is a probability distribution, quantifying the accumulation and propagation of data uncertainty, model uncertainty, and physical evolution uncertainty throughout the entire prediction chain. This embodiment provides full-chain probability prediction, offering a quantitative basis with risk levels for operational decisions (such as setting maintenance thresholds and calculating failure probabilities).

[0095] Example 3: Based on Example 1, this example proposes a method that combines the modified extended finite element method with an adaptive mesh re-division technique based on error estimation to address the issues of tracking accuracy and computational efficiency in crack propagation paths in large and complex welded structures.

[0096] This embodiment aims to address the problem of decreased accuracy or computational divergence in the traditional extended finite element method when cracks propagate near the heat-affected zone of the weld, the interface between different materials, or complex geometric boundaries, due to the failure of the enrichment function or mesh distortion.

[0097] Step 1: Enrichment function of the modified extended finite element method based on the fracture process zone model; the traditional extended finite element method simulates cracks by adding an enhancement function to the standard displacement field. For the weld heat-affected zone of Q355ND steel, there is a fracture process zone at the crack tip, in which the material undergoes nonlinear damage such as plastic deformation and micropore coalescence.

[0098] This embodiment modifies the displacement interpolation function of the extended finite element method by introducing a process region enhancement term based on the Gurson-Tvergaard-Needleman model into the crack tip enrichment function. Displacement approximation function Revised to: in: This is an approximate displacement function, with units in mm; These are standard finite element shape functions, dimensionless; Standard node degrees of freedom, in mm; is the Heaviside step function, used to describe the discontinuous displacements on both sides of the crack surface, and is dimensionless; The crack surface is enriched with degrees of freedom, in mm; It is a dimensionless, nodalized shape function; These are the basis functions for the fracture process region derived based on the Gurson-Tvergaard-Needleman model, used to describe the nonlinear stress-strain field distribution within the fracture process region, and are dimensionless. The process region is enriched with degrees of freedom, in mm; For a standard set of nodes, This is a set of nodes enriched by crack surfaces. This is the set of enriched nodes in the process zone at the crack tip.

[0099] The parameters of the Gurson-Tvergaard-Needleman model were calibrated using uniaxial tensile tests and microfracture analysis of standard specimens of Q355ND steel, specifically including: initial values ​​of void volume fraction. Critical porosity fraction Fracture porosity .

[0100] Step 2: Adaptive mesh re-division based on posterior error estimation; As the crack propagates, especially when the crack deflects, bifurcates, or interacts with other geometric features (such as flange bolt holes) in complex welded structures, the quality of the original extended finite element mesh deteriorates, leading to reduced computational accuracy.

[0101] This embodiment introduces an adaptive mesh repartitioning strategy based on the Zienkiewicz-Zhu error estimator.

[0102] Error estimation: After each expansion step (or every few expansion steps), calculate the posterior error of the strain energy density under the current finite element solution.

[0103] Specifically, the difference between the nodal stresses (discontinuous) obtained from finite element analysis and the nodal stresses (continuous) obtained by extrapolating and smoothing Gaussian point stresses is compared. This difference constitutes a local error index. Defined as: in: For unit The local error index is dimensionless; For unit The integration domain is in mm³; The nodal stress (continuous field) is smoothed and expressed in MPa. This represents the element stress (discontinuous field) obtained from finite element analysis, in MPa.

[0104] Re-partitioning criterion: If the local error within a certain sub-model unit... Exceeding the preset threshold (Set to 1.5 times the global average error), then mark this area for mesh refinement. Re-meshing process: Geometric extraction: based on the current level set function and The geometry of the crack surface and the leading edge of the new crack is extracted.

[0105] Adaptive mesh generation: Using the Delaunay triangulation algorithm, a transition layer mesh is generated around the new crack tip. The element size in this region is determined by the radius of curvature of the crack tip (the mesh is denser where the curvature is greater). In regions far from the crack, the element size remains unchanged or is moderately coarsened.

[0106] Field mapping: Project all state variables (displacement, stress, strain, historical related variables such as equivalent plastic strain, etc.) on the old mesh onto the newly generated mesh through radial basis function interpolation or least squares mapping.

[0107] On the updated mesh, proceed to the next step of crack propagation analysis.

[0108] This method enables dynamic mesh refinement in the areas requiring solution (crack tips, complex geometric junctions) and maintains a coarse mesh in areas where fine solution is not required, thus ensuring the geometric accuracy of crack path tracing while reducing computational overhead.

[0109] Step 3: Sub-model boundary condition correction based on crystal plasticity finite element method; In Example 1, the boundary conditions of the sub-model are directly driven by the displacement field of the global shell model. However, the shell model has difficulty capturing the local stress concentration at the weld caused by differences in crystal orientation.

[0110] This embodiment modifies the boundary conditions of the sub-model: after extracting the displacement field from the global shell model, a separate microscopic representative volume element model based on crystal plastic finite element is used to perform secondary correction on the area near the heat-affected zone of the weld.

[0111] The finite element model of crystal plasticity is as follows: For the weld zone of Q355ND steel, a representative volumetric element model containing ferrite grains and a small amount of pearlite was constructed. The grain size and orientation distribution were determined by electron backscatter diffraction and used as input for the crystal plasticity finite element method.

[0112] The corrected algorithm is as follows: The displacement of the sub-region boundary of the global shell model is used as a macroscopic boundary condition and applied to the crystal plastic finite element-representative volume element model to calculate the microscopic stress field inside the representative volume element. Then, the average stress field correction factor corresponding to the crack surface location of the sub-model is extracted within the representative volume element. This factor reflects the stress amplification effect caused by grain refinement and texture. Ultimately, this is the effective stress intensity factor driving the crack analysis sub-model. for: in: This is the corrected effective stress intensity factor, in MPa·m^{1 / 2}. The stress intensity factor is directly calculated from the global model, and its unit is MPa·m^{1 / 2}. This is a dimensionless stress amplification correction factor obtained from finite element analysis of crystal plasticity.

[0113] This correction couples macroscopic loads, welding residual stresses, and microscopic crystal structure effects together, making the calculation of crack propagation driving force (stress intensity factor) closer to the physical essence.

[0114] Example 4: Building upon Examples 1 and 2, this example addresses the time-varying nature of Paris's law parameters for Q355ND steel under near-shore corrosion and variable-amplitude loads. It fundamentally improves the construction and training method of the physical information neural network, upgrading the physical model parameters from fixed values ​​to dynamic random variables driven by the environment and damage, and achieving full-chain uncertainty propagation. The core of this example lies in constructing a Bayesian physical information neural network capable of outputting prediction confidence intervals and possessing parameter adaptive capabilities.

[0115] (a) Quantification and preprocessing of uncertainty in multi-source data: Based on Example 1, this example adds a quantification step to address data uncertainty. For each strain gauge channel and acoustic emission channel, an adaptive Gaussian mixture model is used to fit the probability density of the denoised time-series signal.

[0116] The specific steps are as follows: For each strain gauge channel, take continuous Each sampling point constitutes a time window Using the expectation-maximization algorithm to fit a graph containing A mixture model with Gaussian components, where The parameters are automatically determined by the Bayesian information criterion. The fitted Gaussian mixture model parameters are: mean... ,variance Weight , The optimal estimate of the strain value at that moment. and its variance The calculation is as follows: in: This is the optimal estimate of strain, in με; For the first The weights of the Gaussian components are dimensionless, and ; For the first The mean of the Gaussian components, expressed in με.

[0117] in: The variance for strain estimation is expressed in (με). ; For the first The variance of each Gaussian component, in units of (με). .

[0118] The variance The heteroscedasticity caused by sensor noise, environmental vibration, and measurement uncertainty was quantified and used as an independent dimension for subsequent neural network input, enabling the model to utilize the uncertainty information of the input data.

[0119] (ii) Constructing a Bayesian physical information neural network; The core of this embodiment lies in constructing a Bayesian physical information neural network. Its architecture is similar to the deep feedforward network in Embodiment 1, but a Bayesian layer is introduced into the hidden layers. The Bayesian layer assigns weights to each network layer. and bias It can be considered as a random variable that follows a Gaussian distribution, i.e. , The parameters of these distributions are learned through variational inference. , ( ), rather than a fixed value.

[0120] 1. The dynamic physical parameters are modeled as follows: The material fatigue characteristic parameters in Paris's law and Defined as an environmentally corrosive factor With cumulative damage state Driven dynamic random variables: The prior distribution is set to a log-normal distribution: To ensure its positive value.

[0121] The prior distribution is set to a normal distribution: .

[0122] Among them, environmental corrosion factors The calculation method is as follows: in: It is an environmental corrosion factor, dimensionless; The input is a pre-trained corrosion accelerator network with 3 hidden layers (32 neurons per layer), and the input is the salt spray concentration. (Unit: mg / m³), Temperature (unit: °C) and relative humidity (Dimensionless, expressed as a percentage), the output is a dimensionless corrosion acceleration factor, used to quantify the nonlinear accelerating effect of environmental factors on crack propagation rate. This network is used in laboratory corrosion fatigue test data (different...) , , Combination The system is pre-trained on a curve to learn the nonlinear acceleration effect of the corrosive environment on the crack propagation rate.

[0123] Cumulative damage state Energy accumulation value in acoustic emission eigenvectors The integral of the crack propagation rate prediction is updated in real time: in: for The cumulative damage state at any given time is dimensionless. The crack propagation rate predicted at the previous moment is expressed in mm / cycle. The critical crack length, in mm, is determined by the fracture toughness in step five. The results were obtained through inversion calculations. The initial crack length is in mm. The weighting coefficient contributing to acoustic emission energy is dimensionless and calibrated to 0.3 based on historical data. for The cumulative acoustic emission energy at any given time is expressed in aJ. The critical cumulative acoustic emission energy is expressed in aJ.

[0124] Dynamic random variables and The posterior distribution parameters (mean) ,variance mean ,variance The mapping relationship is automatically inferred by the Bayesian layer of B-PINN during the learning process and is implicit in the network weight distribution.

[0125] 2. The loss function is reconstructed as follows: Total loss function Based on Example 1, a KL divergence term was added to regularize the model complexity, and constraints on the dynamics of the physical parameters were introduced: in: is the total loss function of the Bayesian physical information neural network, which is dimensionless; is the weighting coefficient for the data likelihood term, dimensionless, with a value of 1.0; It is the data likelihood term in the weighted posterior distribution The expected value is dimensionless. This is the weighting coefficient for the physical constraint term, dimensionless, with a value of 0.8; It is the physical constraint residual in the weighted posterior distribution The expected value is dimensionless. The weighting coefficients for the KL divergence term are dimensionless and take values ​​of [value missing]. ( (Number of training samples) It is an approximate posterior. with prior The KL divergence between them is dimensionless and serves as a regularization term; This is the weighting coefficient for the dynamic parameter constraint term, dimensionless, and has a value of 0.5. These are dynamic parameter constraint terms used for constraint. and The time-varying smoothness is defined as: in: and This is a smoothing coefficient, dimensionless, with a value of 0.1, used to control parameters. and The smoothness of the time-varying process is important to prevent prediction oscillations caused by sudden parameter changes; and They are respectively and time The posterior mean; and They are respectively and time The posterior mean.

[0126] Physical constraint operator Paris's law is extended to include dynamic parameters and corrosion acceleration factors: in: The crack propagation rate is based on physical constraints and is expressed in mm / cycle. The dynamic Paris law coefficients are sampled from the B-PINN posterior distribution, with units of mm / cycle / (MPa·m1 / 2)m_{dyn}. Let be the dynamic Paris law exponent sampled from the B-PINN posterior distribution, which is dimensionless; The equivalent stress intensity factor amplitude is expressed in MPa·m^{1 / 2}. Environmental corrosion factors, dimensionless, are identified through a corrosion accelerator network. Calculated.

[0127] 3. The model training and application process is as follows: Model training employs variational Bayesian inference. This is achieved by minimizing the loss function. The parameters of the posterior distribution of the network weights are learned. After training, for the input at each time step, the parameters are obtained from the posterior distribution of the weights. Medium sampling Group weight parameters, obtain Different prediction results (crack length) and expansion rate The final predicted value is this. The Monte Carlo average of the results, and the prediction variance, constitute the uncertainty interval at the 95% confidence level: in: This represents the predicted mean crack length in mm. Variance of crack length prediction, in mm ; The number of parameter sets sampled from the posterior distribution is dimensionless and takes a value of 1000. For the first The predicted crack length under the group sampling parameters, in mm.

[0128] (III) Uncertainty Propagation and Recursive Bayesian Update; In remaining lifetime prediction, this embodiment predicts the crack length distribution for each incremental step of the B-PINN output. As input to the Monte Carlo simulation, the Monte Carlo simulation randomly samples a value based on the probability distribution of the current crack length at each increment step, and calculates the stress intensity factor distribution for the next increment step accordingly. This process is repeated until the final remaining life prediction is obtained as a probability distribution.

[0129] A recursive Bayesian update algorithm is introduced, utilizing crack length data obtained from real-time monitoring. (Measured by ultrasonic or eddy current sensors), dynamically correcting dynamic random variables. and The posterior distribution: in: for It incorporates the posterior distribution of real-time monitoring data at all times; The likelihood function is constructed based on the difference between the current monitoring data and the B-PINN predicted values, assuming that the observation error follows a Gaussian distribution. , ; for The posterior distribution at time step is provided by the predicted distribution of B-PINN.

[0130] This recursive update process is implemented using the particle filter algorithm. 10 particles, each particle representing a group Parameters. The updated particle set reflects the latest structural state and is used to drive subsequent crack propagation predictions, enabling adaptive calibration of the model.

[0131] To verify the technical effectiveness of this embodiment, a comparative experiment was added based on the simulation experiment of Embodiment 1. The experimental conditions remained the same as in Embodiment 1, but the environmental salt spray concentration was increased from 50,000 cycles. Mutation to The operating conditions were simulated to demonstrate a near-shore wind farm encountering a strong corrosive gas cloud. The predictive performance of each method is shown in Table 2.

[0132] Table 2: Comparison of performance indicators of various prediction methods under corrosive conditions; As can be seen from the data analysis in Table 2, the B-PINN provided in this embodiment has significantly better prediction accuracy than the original scheme after a sudden change in the corrosion environment. The original scheme, due to its fixed Paris parameters, cannot respond to the accelerated crack propagation caused by environmental changes, resulting in a sharp increase in error. In contrast, this embodiment, through dynamic Bayesian adaptation, successfully captures... and The posterior distribution changes ensure that the predicted value closely follows the actual evolution curve. Furthermore, this embodiment provides full-chain probability prediction, with its 95% confidence interval covering 94.2% of the true value. This provides a quantitative risk level basis for operational decisions, addressing the industry pain point of traditional methods that "only provide a number, without knowing its accuracy."

[0133] This embodiment addresses the following industry pain points by introducing a Bayesian physical information neural network: By elevating the parameters of Paris's law from static assumptions to dynamic random variables driven by environment and damage, the fundamental problem of parameter mismatch in traditional fracture mechanics models under corrosion and variable amplitude loads is solved. In the field of wind turbine tower crack prediction, the uncertainty quantification of the entire chain from input data and physical models to the final remaining life is achieved, outputting probabilistic prediction results with confidence levels, providing risk assessment basis for operation and maintenance decisions rather than a single numerical value. Through variational Bayesian inference and recursive updates, the model can quickly adjust the posterior distribution of its internal physical parameters even with only a small amount of new operating condition data, significantly improving the extrapolation prediction capability for unseen operating conditions (such as sudden changes in salt spray concentration). By integrating dynamic Bayesian adaptation of physical parameters with uncertainty propagation technology, an intelligent prediction system that "knows not only what, but also why, and with confidence" is constructed, greatly improving the accuracy, robustness, and engineering practical value of crack propagation prediction in complex service environments, providing core technical support for risk-based condition-based maintenance of wind farms.

[0134] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0135] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting crack propagation in wind turbine tower materials, characterized in that, Includes the following steps: Step 1, Multi-source heterogeneous service data acquisition and preprocessing: Deploy a high-precision multi-directional strain monitoring matrix, acoustic emission sensor array and environmental load monitoring unit at the preset key monitoring parts of the wind turbine tower to acquire strain, acoustic emission and environmental load data in real time, and obtain time-series feature dataset after preprocessing. Step 2: Construct a high-fidelity finite element sub-model based on digital twin technology: Based on the original design parameters of the wind turbine tower, the material constitutive relationship, and the measured residual stress distribution data, establish a global finite element model and a local refined crack analysis model for key monitoring parts. Drive the sub-model boundary through the displacement field of the global model to obtain the dynamic stress intensity factor of the local area. Step 3: Construct a physical information neural network crack propagation prediction model: The physical information neural network consists of a deep feedforward neural network and a physical constraint operator. The input vector includes time-series feature data and the number of load cycles, and the output vector is the crack length and propagation rate. The loss function introduces a physical constraint residual term based on Paris's law of fracture mechanics, so that the model can follow the physical evolution law of crack propagation while fitting the measured data. Step 4, simulation of crack spatial propagation path based on extended finite element method: using the crack propagation rate predicted in step 3 as the driving force, an extended finite element operator is introduced into the local refined crack analysis model, the crack geometry is described by the level set function, the crack propagation deflection angle is determined according to the maximum circumferential stress criterion, the crack front coordinates are dynamically updated, and the three-dimensional propagation path of the crack is tracked. Step 5, Structural Integrity Assessment and Remaining Life Prediction: Combine the crack propagation rate predicted in Step 3 with the spatial path simulated in Step 4 to calculate the critical relationship between the stress intensity factor corresponding to the real-time crack size and the material fracture toughness. When the preset threshold is reached, a safety warning is triggered, and a remaining life assessment report is generated.

2. The method for predicting crack propagation in wind turbine tower materials according to claim 1, characterized in that, In step one, the preset key monitoring areas include the heat-affected zone of the tower weld, the flange connection, and the variable cross-section transition zone; the high-precision multi-directional strain monitoring matrix is ​​realized by fiber optic grating sensors with a wavelength range of 1525nm to 1565nm and a strain resolution of 1με; the sensor spacing is determined according to the stress gradient of the tower's stressed skin, and the density within 50mm of the weld edge is not less than 2 sensors per 100cm²; and the sensor surface is covered with a EPDM rubber protective layer.

3. The method for predicting crack propagation in wind turbine tower materials according to claim 1, characterized in that, In step one, the acoustic emission sensor array uses a piezoelectric ceramic sensor with a response frequency range of 20kHz to 400kHz; a high-viscosity coupling agent is applied to the sensor mounting location, and it is fixed to the tower surface by a stainless steel clamp; the environmental load monitoring unit includes an ultrasonic anemometer installed at the top of the tower and a salt spray concentration sensor installed on the outer wall of the tower, used to provide wind load vector data and corrosion environment correction parameters.

4. The method for predicting crack propagation in wind turbine tower materials according to claim 1, characterized in that, In step one, the data preprocessing includes preprocessing based on an adaptive wavelet threshold denoising algorithm to remove random environmental noise, and data dimensionality reduction based on principal component analysis: by performing eigenvalue decomposition on the covariance matrix composed of multi-channel strain signals, principal component components with a cumulative contribution rate of over 95% are extracted as input feature vectors for subsequent physical information neural networks.

5. The method for predicting crack propagation in wind turbine tower materials according to claim 1, characterized in that, In step two, the material constitutive relation corresponds to Q355ND high-performance structural steel; the full-scale global finite element model is discretized using four-node reduced integral shell elements. The localized refined crack analysis model uses eight-node hexahedral linear reduced integral elements for meshing, and uses singular elements to handle stress singularity in the preset crack tip region. The boundary range of the sub-model is set to be more than 10 times the preset crack length.

6. The method for predicting crack propagation in wind turbine tower materials according to claim 1, characterized in that, In step two, the residual stress distribution data measured by the material is obtained by sampling and measuring typical welds of the tower using X-ray diffraction to obtain the deep residual stress distribution law, and then pre-setting the measured residual stress field in the local refined crack analysis model by importing the initial stress field.

7. The method for predicting crack propagation in wind turbine tower materials according to claim 1, characterized in that, In step three, the loss function of the physical information neural network Defined as: in, The total loss function is dimensionless. The weight coefficients of the data fitting terms are dimensionless. The mean square error between the network prediction and the monitoring data is dimensionless. The weighting coefficients for the physical constraint terms are dimensionless. The physical constraint residual is dimensionless; Specifically defined as: In the formula, The number of samples in the current training batch is dimensionless. This is a sample index, dimensionless; The first prediction of the neural network Crack propagation rate of each sample, in mm / cycle; These are parameters related to material fatigue characteristics. For the first The amplitude of the equivalent stress intensity factor for each sample; This is the fatigue property index of the material.

8. The method for predicting crack propagation in wind turbine tower materials according to claim 7, characterized in that, In step three, the amplitude of the equivalent stress intensity factor Calculated using Walker correction terms: In the formula, The equivalent stress intensity factor amplitude is expressed in MPa·m^{1 / 2}. The maximum stress intensity factor, in MPa·m^{1 / 2}, is calculated by the local refined crack analysis model constructed in step two. The stress ratio is defined as the ratio of the minimum stress to the maximum stress, and is dimensionless. The material sensitivity coefficient is dimensionless.

9. The method for predicting crack propagation in wind turbine tower materials according to claim 1, characterized in that, In step three, the physical information neural network adopts a 6-layer hidden layer structure, with each layer containing 128 neurons. The activation function is the Tanh function, which has the characteristic of second-order continuous derivative. The physical constraint operator is constructed by calculating the partial derivatives of the output vector with respect to the physical parameters using automatic differentiation technology under the deep learning framework.

10. A crack propagation prediction system for wind turbine tower materials applied to the method described in any one of claims 1-9, characterized in that, include: The multi-source sensor data acquisition module is set up on the wind turbine tower site. It includes a high-precision multi-directional strain monitoring matrix, an acoustic emission sensor array, and an environmental load monitoring unit deployed at key monitoring locations. It is used to acquire strain, acoustic emission, and environmental load data in real time and perform data preprocessing based on adaptive wavelet threshold denoising and principal component analysis. The digital twin simulation engine is pre-built with a full-scale global finite element model of the tower and sub-models of key parts, which are used to calculate the dynamic stress field and stress intensity factor of key parts of the tower in real time. The physical information fusion prediction module integrates a physical information neural network. Its input end is connected to the output end of the multi-source sensor data acquisition module to receive the time-series feature dataset. At the same time, it is connected to the digital twin simulation engine to obtain mechanical parameters and perform crack propagation rate prediction. The spatial path tracking module is connected to the physical information fusion prediction module. Based on the extended finite element method, it uses the predicted crack propagation rate to drive the geometric evolution of the crack surface and dynamically reconstructs the three-dimensional spatial propagation path of the crack. The early warning decision support module connects the spatial path tracking module and the physical information fusion prediction module. It is used to calculate the critical relationship between real-time crack size and material fracture toughness, output a remaining life prediction report, and perform graded early warnings according to preset thresholds.

Images