Knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades

Abstract

The present application relates to the technical field of electric digital data processing, and particularly relates to a knowledge-driven offshore wind turbine blade fatigue crack real-time prediction method.The method comprises the following steps: establishing a parameter-conditioned PINN crack forward prediction model; setting a lightweight correction head at the output end of the PINN crack forward prediction model to construct a correction prediction model; constructing a basic loss function of the PINN crack forward prediction model; constructing a DBN parameter online updating model, and calling the PINN crack forward prediction model for forward prediction; constructing a DBN posterior weighted physical residual loss to optimize the loss function of the PINN crack forward prediction model; constructing an online correction loss function and updating the lightweight correction head; updating parameters and performing crack length prediction.The method can maintain high prediction accuracy under the condition of fewer crack observation points and longer detection cycles, and is more stable and reliable for judging the crack propagation trend, residual life and risk state of the offshore wind turbine blade.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades. Background Technology

[0002] Offshore wind power is an important direction for the large-scale development of renewable energy in my country. As the installed capacity of offshore wind power expands and development areas extend to deeper and more remote seas, the importance of long-term safe operation and intelligent maintenance of the units is further enhanced.

[0003] Offshore wind turbine blades are subjected to a variety of factors over long periods, including wind loads, turbulence, pitch loads, gravity loads, start-up and shutdown impacts, salt spray corrosion, and wet cycling. They are critical components of the turbine unit prone to fatigue damage. During long-term service, blades may experience fatigue accumulation, crack initiation and propagation, damage to adhesive bonding areas, localized stiffness degradation, and material performance deterioration. Fatigue cracks or equivalent damage propagation are characterized by difficulty in early detection, continuous evolution, and accelerated propagation in later stages, directly impacting blade structural safety, turbine operational reliability, and maintenance decisions.

[0004] Existing blade health monitoring methods typically rely on strain, fiber Bragg gratings, acoustic emission, vibration, visual inspection, or UAV inspections to obtain status information, and then identify damage through threshold judgment, signal processing, image recognition, modal analysis, or machine learning methods. While these methods can detect blade anomalies to some extent, in actual offshore wind power operation scenarios, crack or damage observation data are often sparse, discontinuous, and noisy, and the detection cost is high, making it difficult to continuously describe the entire propagation process of fatigue cracks or equivalent damage.

[0005] Current crack propagation models typically employ fixed physical formulas and parameters. While these models can reflect the physical laws governing crack propagation, they cannot be dynamically modified based on new crack observation data. When the blade's service environment, load conditions, material damage state, or observation errors change, fixed-parameter models are prone to causing the predicted trajectory to gradually deviate from the actual crack propagation process.

[0006] Finite element method (FEM), extended finite element method (EPM), or other high-fidelity physical simulation methods can accurately describe the blade structure response, crack tip stress field, and crack propagation process, but their modeling, solving, and parameter scanning processes are computationally intensive. If a high-fidelity physical model is re-run after each new observation data is obtained during the long-term operation of offshore wind turbine blades, it will be difficult to meet the requirements for real-time or near-real-time prediction.

[0007] Therefore, predicting fatigue crack propagation in surface-mounted offshore wind turbine blades urgently requires a predictive method that integrates physical mechanisms, simulation data, monitoring data, and online correction mechanisms. This method should be able to continuously correct model parameters under limited observation conditions, ensuring dynamic consistency between the virtual model and the actual damage evolution state of the blade. This would provide technical support for blade condition assessment, crack propagation prediction, fault early warning, and predictive maintenance. Summary of the Invention

[0008] The purpose of this invention is to overcome the above-mentioned defects in the existing technology and propose a knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades. This method can maintain high prediction accuracy even when there are fewer crack observation points and a longer detection cycle, and makes the judgment of crack propagation trend, remaining life and risk status of offshore wind turbine blades more stable and reliable.

[0009] The technical solution of this invention is: a knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades, comprising the following steps: S1. Establish a parametrically conditional PINN crack forward prediction model; S2. Set a lightweight correction head at the output of the PINN crack forward prediction model to construct a modified prediction model; S3. Construct the basic loss function for the PINN crack forward prediction model; S4. Construct an online DBN parameter update model and call the PINN crack forward prediction model for forward prediction; S5. Construct the DBN posterior weighted physical residual loss and optimize the loss function of the PINN crack forward prediction model; S6. Construct the online correction loss function and update the lightweight correction head; S7. Update parameter feedback and predict crack length.

[0010] In step S1 of this invention, the PINN crack forward prediction model is expressed as: ， in, Indicates the number of loop iterations; Indicates the crack propagation coefficient; Indicates the crack propagation index; This represents the backbone network of the PINN crack forward prediction model; This represents the backbone network parameters of the PINN crack forward prediction model; This represents the predicted crack length value from the PINN crack forward prediction model.

[0011] In step S2, the modified prediction model includes the PINN crack forward prediction model and a lightweight correction head; A lightweight correction head is installed at the output of the PINN crack forward prediction model. At this point, the crack prediction result output by the PINN crack forward prediction model is: in, This indicates a lightweight correction head used to output crack length correction. This indicates the parameters of the calibration head.

[0012] In step S3, during the training phase of the PINN crack forward prediction model, a basic joint loss function is constructed using data loss, physical loss, and monotonicity constraints: ， in, Indicates data loss. Indicates physical loss. This represents the monotonicity constraint loss. Weights representing physical losses; The weights represent the monotonicity constraint loss.

[0013] In step S4, the DBN parameter online update model uses crack propagation parameters as the update object: ， in, Indicates the first The parameter state vector of a crack propagation parameter particle; Indicates the first The crack propagation coefficient corresponding to each parameter particle; Indicates the first The crack propagation index corresponding to each parameter particle; Generate multiple sets of candidate parameter particles around the current parameter state: ， in, Indicates the number of particles; Indicates the first At the first update time The parameter state vector of each candidate parameter particle; Indicates the first At the first update time The crack propagation coefficient corresponding to each candidate parameter particle; Indicates the first At the first update time The crack propagation index corresponding to each candidate parameter particle; Indicates the number of the parameter particle; For each set of parameter particles, the PINN crack forward prediction model is called to perform forward prediction, and the predicted crack length value corresponding to that parameter particle is obtained: ， Based on the predicted crack length and the observed crack length Error between Calculate particle weights: ， ， in, Indicates the first The online update time is numbered as follows The unnormalized weights of the candidate parameter particles; This represents the variance of the observed noise level for crack length. And normalize the particle weights: ， From this, we can obtain the posterior parameter estimates: ， ， in, Indicates the first The posterior estimate of the crack propagation coefficient at each online update time; Indicates the first The posterior estimate of the crack propagation index at each online update time.

[0014] In step S5, a lightweight correction head is first used to perform preliminary local correction on the original prediction results of the PINN crack forward prediction model, resulting in the corrected crack length prediction value: ， in, This indicates the number of cycles in the PINN crack forward prediction model. The predicted value of the original crack length is output below; Indicates the number of cycles for the lightweight calibration head. The local correction amount of the output; Parameters indicating a lightweight calibration head; This represents the predicted crack length after correction by the lightweight correction head. The posterior parameter particles obtained by updating DBN , and its posterior particle weights Introducing the physical loss function of the PINN crack forward prediction model, we construct the DBN posterior weighted physical residual loss function: ， in, Indicates the number of online update points; Indicates the first One online update point; Indicates the first The predicted interval endpoint after one online update point; Indicates the first Each online update point corresponds to the number of subsequent loops. Influence weight; Indicates the number of particles in the DBN parameter; Indicates the number of the parameter particle, and ; Indicates the first The number at each online update point is: The posterior weights of the parameter particles; This indicates that the number of particles in the posterior of the DBN is... The crack propagation coefficient; This indicates that the number of particles in the posterior of the DBN is... Crack propagation index; Indicates the corrected crack length The corresponding range of stress intensity factors; At this point, the loss function of the PINN crack forward prediction model is optimized as follows: , in, Indicates data loss; This represents the posterior weighted physical residual loss of DBN; This represents the loss due to monotonicity constraints. The weighting coefficients represent the posterior weighted physical residual loss of DBN; This represents the weighting coefficient of the monotonicity constraint loss.

[0015] In step S6, at the first After several observation points, an online correction loss function is constructed using observation error, DBN posterior weighted physical residual loss, smoothing constraint, and correction magnitude constraint: ， in, Indicates the loss of observation error. This represents the posterior weighted physical residual loss of DBN. This represents the smoothing constraint loss. This indicates the correction amplitude constraint loss; The observation error loss is: ， in, Indicates the first The number of loops or loading steps corresponding to each online update point This indicates the cumulative number of load cycles. The crack length prediction value obtained after inputting the modified prediction model, i.e., the first... Corrected predicted crack length at each online update point; Indicates the first Crack length observations obtained at each online update point; The smoothing constraint loss is: ， in, This indicates the number of sampling points used to calculate the loss within the local correction interval; Indicates the first Each online update point corresponds to a local correction interval or a set of sampling points; Indicates the number of cycles for the lightweight calibration head. The local correction amount output at the location; This indicates that the lightweight calibration head is at the next adjacent sampling point. The local correction amount output at the location; The correction amplitude constraint loss is: The calibration head parameters are trained using an online calibration loss function. The final corrected head parameters obtained after training are: .

[0016] In step S7, The updated parameters are fed back into the PINN crack forward prediction model: ， The crack length is predicted to be: ， in, Indicates the first The cumulative load cycle count corresponding to each online update point This represents the cumulative load cycle count corresponding to the time to be predicted, and .

[0017] The beneficial effects of this invention are: (1) A closed-loop correction process capable of forming a bidirectional PINN-DBN fusion: This application does not use PINN alone for crack propagation prediction, nor does it use DBN alone for parameter updating. Instead, it constructs a closed-loop correction process that integrates PINN and DBN bidirectionally. On the one hand, during the DBN parameter update process, each parameter particle calls PINN to complete the forward prediction and calculates the particle weight based on the error between the predicted value and the observed value; on the other hand, the parameter particles and their weights obtained from the DBN update are further fed into the PINN physical loss function to construct the posterior weighted physical residual loss. Therefore, this application can form a closed-loop online correction process of "PINN forward prediction → DBN parameter update → weighted physical loss construction based on DBN updated parameters → lightweight correction head local update → PINN continued prediction", so that the observation information not only affects the current observation point, but can continue to act on the subsequent crack propagation prediction stage. (2) Achieve efficient online calibration based on a lightweight calibration head: This application sets a lightweight correction head after the output of the PINN crack forward prediction model: In the online stage, after the DBN parameter online update model completes the parameter update, the lightweight correction head is locally optimized using the newly added observation data, the weighted physical loss based on the DBN parameter online update model, the smoothing constraint and the correction amplitude constraint, so that the prediction trajectory after the observation point can be corrected in stages. Therefore, this application can achieve rapid correction of local prediction deviations by using a lightweight correction head while maintaining the backbone prediction capability of the PINN crack forward prediction model, thereby reducing the computational burden of the online correction process and making it more suitable for intelligent twin prediction scenarios of fatigue crack propagation in offshore wind turbine blades. (3) Improved the model's adaptability to noisy observation data: In the process of updating the DBN parameter online update model, this application adopts a posterior estimation method based on particle weights. The weights are calculated according to the error between the predicted and observed values ​​of each parameter particle, thereby updating the crack propagation parameters probabilistically. This method does not simply replace the predicted value with a single observed value, but forms a posterior estimate through the weighted result of multiple candidate parameter particles. Therefore, even when there are measurement errors, sensor noise, visual recognition errors, or manual reading errors in crack observations, this application can still use the observation information to correct the parameters and subsequent prediction trajectories, thereby reducing the impact of a single abnormal observation on the prediction results. (4) Facilitates crack propagation prediction in digital twin and health monitoring scenarios: This application employs a method that trains the PINN crack forward prediction model using simulation data or existing crack propagation data, triggers DBN parameter updates with a small amount of observation data, uses DBN posterior information to participate in the construction of physical loss, and completes local correction with a lightweight correction head. This method does not require a large amount of real-time crack data to be continuously input during the online phase, nor does it require rebuilding a high-fidelity physical model at each observation point. This application is easy to embed into fatigue crack intelligent twins, structural health monitoring platforms, or predictive maintenance systems for crack propagation state prediction, trend assessment, risk warning, and operation and maintenance decision support.

[0018] In summary, this application addresses the challenges of sparse observation data, observation noise interference, uncertainties, and the difficulty in continuously characterizing damage evolution in offshore wind turbine blade crack prediction. It constructs a crack propagation prediction method with physical consistency and dynamic updates based on observation data. This method maintains high prediction accuracy even with limited crack observation points and long detection cycles, reducing the impact of single abnormal observations or measurement errors on subsequent prediction results and mitigating the risk of continuous error accumulation during long-term prediction. This makes the assessment of crack propagation trends, remaining life, and risk status more stable and reliable. Furthermore, this method reduces reliance on large amounts of continuous detection data and repeated high-cost simulations, improving the applicability of offshore wind turbine blade crack prediction methods in practical health monitoring, risk warning, and operation and maintenance decision-making scenarios. It possesses clear engineering application value and industrial promotion significance. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method described in this application; Figure 2 This is a schematic diagram of the structure of the SENT / SE(T) sample; Figure 3 This is a diagram showing the overall prediction results; Figure 4 This is the effect of the first filter update at a local observation point; Figure 5 This is the effect of the second filtering update at the local observation points; Figure 6 This is the effect of the third filter update at the local observation points; Figure 7 This is the effect of the fourth filter update at the local observation point; Figure 8 This is a comparison between the crack propagation curve predicted by the DBN-PINN model and the actual observed curve; Figure 9 This is a comparison chart of RMSE errors; Figure 10 It is R 2 Goodness-of-fit comparison chart. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many ways other than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] This application proposes a knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades. The flowchart of this method is as follows: Figure 1 As shown, this method uses the Physical Information Neural Network (PINN) as the forward prediction model for crack propagation, and uses the Dynamic Bayesian Network (DBN) or particle filtering method as Paris parameters to update the forward prediction model for crack propagation online. It constructs a closed-loop online correction framework of "PINN forward prediction - DBN parameter update - posterior weighted physical residual construction - lightweight correction head local correction - PINN continued prediction".

[0023] The method specifically includes the following steps.

[0024] The first step is to establish a parametrically conditional PINN crack forward prediction model.

[0025] The structural parameters, load parameters, material parameters, crack length observation data, and finite element or extended finite element simulation data of offshore wind turbine blades are obtained and preprocessed to form a dataset for training the PINN crack forward prediction model.

[0026] The data, including the number of cycles, crack length, load parameters, and material parameters, were obtained from blade fatigue tests, structural health monitoring, non-destructive testing, or simulation calculations, and were compiled into data samples corresponding to the crack propagation process.

[0027] Crack propagation parameters are not acquired directly from sensor observation data. Instead, during the model building phase, initial value ranges or initial estimates are determined based on material fatigue crack propagation test data, existing historical crack propagation data, engineering experience parameters, standard crack propagation models, or offline fitting results. These values ​​are then used as conditional inputs to the parameterized PINN crack forward prediction model. During the online prediction phase, crack propagation parameters are used as variables to be updated in the DBN parameter online update model.

[0028] Before model training, the aforementioned data is preprocessed to form training samples in a unified format. This preprocessing includes data filtering, outlier removal, missing data handling, unit and format standardization, crack propagation process matching, input / output variable construction, and data scale standardization. The preprocessed data is used to train the parameter-conditional PINN crack forward prediction model, enabling the model to predict the change in blade crack length with the number of cycles under different parameter conditions.

[0029] A forward prediction model for PINN cracks is established. This differs from models that rely solely on the number of cycles. As a standard PINN input, this application also incorporates crack propagation parameters into the model input, enabling the PINN crack forward prediction model to output crack length prediction results under different parameter conditions. The PINN crack forward prediction model can be expressed as: ， in, Indicates the number of loop iterations; and Represents the crack propagation parameters, where It represents the crack propagation coefficient, used to characterize the influence of material, structural state, and service environment on the overall level of crack propagation rate; The crack propagation index is used to characterize the sensitivity of the crack propagation rate to changes in the stress intensity factor amplitude or the equivalent crack driving force. This represents the backbone network of the PINN crack forward prediction model; This represents the backbone network parameters of the PINN crack forward prediction model; This represents the predicted crack length value from the PINN crack forward prediction model.

[0030] The PINN crack forward prediction model proposed in this application no longer learns only a single crack propagation curve, but instead learns a crack propagation mapping controlled by crack propagation parameters. After subsequent DBN updates yield new crack propagation parameters, these parameters can be fed back into the PINN crack forward prediction model, enabling it to continue prediction under the updated parameter conditions.

[0031] The second step is to set a lightweight correction head at the output of the PINN crack forward prediction model to construct a modified prediction model, which includes the PINN crack forward prediction model and the lightweight correction head.

[0032] To avoid retraining the entire PINN crack forward prediction model at each observation point during the online phase, this application sets a lightweight correction head after the output of the PINN crack forward prediction model. .

[0033] The improved final prediction form of this application is as follows: ， in, Indicates a lightweight calibration head; Indicates the calibration head parameters; Used to output crack length correction amount ,Right now: ， Therefore, the final prediction result for: ， Right now: ， During the online phase, the backbone network parameters of the PINN crack forward prediction model Keep the model frozen and do not retrain the entire PINN crack forward prediction model; only adjust the lightweight correction head parameters. Perform a small number of iterative updates. This allows us to use newly added observation data to correct local prediction biases while avoiding the high computational cost of fully retraining the PINN crack forward prediction model.

[0034] The third step is to construct the basic loss function for the PINN crack forward prediction model.

[0035] During the training phase of the PINN crack forward prediction model, a basic joint loss function is constructed using data loss, physical loss, and monotonicity constraints: ， in, Indicates data loss. Indicates physical loss. This represents the monotonicity constraint loss. Weights representing physical losses; The weights represent the monotonicity constraint loss.

[0036] The fourth step is to establish an online DBN parameter update model and call the PINN crack forward prediction model for forward prediction.

[0037] Crack length observations are obtained at preset observation points or actual detection points. Specifically, crack images of the surface of the structure under test can be acquired using an image acquisition device, and the crack length can be calculated using image processing methods such as image segmentation, edge detection, or crack skeleton extraction; alternatively, crack length can be obtained through laser ranging, ultrasonic testing, eddy current testing, or manual measurement.

[0038] The DBN parameter online update model uses crack propagation parameters as the update object: ， in, Indicates the first The parameter state vector of a crack propagation parameter particle; Indicates the first The crack propagation coefficient corresponding to each parameter particle is used to characterize the overall level of crack propagation rate. Indicates the first The crack propagation index corresponding to each parameter particle is used to characterize the sensitivity of the crack propagation rate to changes in the crack driving force.

[0039] Generate multiple sets of candidate parameter particles around the current parameter state: ， in, Indicates the number of particles; Indicates the first At the first update time The parameter state vector of each candidate parameter particle; Indicates the first At the first update time The crack propagation coefficient corresponding to each candidate parameter particle; Indicates the first At the first update time The crack propagation index corresponding to each candidate parameter particle; Indicates the number of the parameter particle.

[0040] For each set of parameter particles, the PINN crack forward prediction model is called to perform forward prediction, and the predicted crack length value corresponding to that parameter particle is obtained: ， Then, based on the error between the predicted and observed values... Calculate particle weights: ， ， in, Indicates the first The online update time is numbered as follows The unnormalized weights of the candidate parameter particles are used to characterize the degree of matching between the predicted result corresponding to the candidate parameter particle and the observed crack length. This represents the noise variance of crack length observations, used to characterize the uncertainty of crack length observation data.

[0041] And normalize the particle weights: ， From this, we can obtain the posterior parameter estimates: ， ， in, Indicates the first The posterior estimate of the crack propagation coefficient at each online update time is used to characterize the overall level of the crack propagation rate after being updated with the current observation data; Indicates the first The posterior estimate of the crack propagation index at each online update time is used to characterize the sensitivity of the updated crack propagation rate to changes in the crack driving force, based on the current observation data.

[0042] This step embodies the first layer of fusion, namely: the PINN crack forward prediction model provides a forward crack length prediction for each DBN parameter particle, which is used to calculate particle weights and posterior parameters.

[0043] The fifth step is to construct the DBN posterior weighted physical residual loss.

[0044] The online update model for DBN parameters has completed its first step. After updating the parameters at each online update point, the DBN posterior parameter particles are obtained. , and its posterior particle weights To ensure that the posterior parameter information obtained from the DBN update continues to be used in the subsequent crack propagation prediction process, this application first uses a lightweight correction head to perform preliminary local correction on the original prediction results of the PINN crack forward prediction model, obtaining the corrected crack length prediction value: ， in, This indicates the number of cycles in the PINN crack forward prediction model. The predicted value of the original crack length is output below; Indicates the number of cycles for the lightweight calibration head. The local correction amount of the output; Parameters indicating a lightweight calibration head; This represents the predicted crack length after correction by the lightweight correction head.

[0045] After obtaining the corrected crack length prediction, the posterior parameter particles and their weights obtained from the DBN update are introduced into the physical loss function of the PINN crack forward prediction model to construct the DBN posterior weighted physical residual loss function: ， in, Indicates the number of online update points; Indicates the first One online update point; Indicates the first The predicted interval endpoint after one online update point; Indicates the first Each online update point corresponds to the number of subsequent loops. Influence weight; Indicates the number of particles in the DBN parameter; Indicates the number of the parameter particle, and ; Indicates the first The number at each online update point is: The posterior weights of the parameter particles; This indicates that the number of particles in the posterior of the DBN is... The crack propagation coefficient; This indicates that the number of particles in the posterior of the DBN is... The crack propagation index.

[0046] Indicates the corrected crack length The corresponding stress intensity factor range can be obtained based on blade load parameters, structural geometric parameters, and crack size, using existing fracture mechanics calculation formulas, finite element methods, extended finite element methods, or pre-established crack length-stress intensity factor mapping relationships.

[0047] The physical residual loss proposed in this application is based on the posterior parameter particle updated by DBN. , and its posterior weights Co-construction. In other words, the physical constraints of the PINN crack forward prediction model are no longer determined by a single fixed parameter, but by a weighted distribution of DBN posterior parameters.

[0048] Therefore, the physical residual loss of the PINN crack forward prediction model proposed in this application is transformed from "fixed parameter Paris physical residual" to "DBN posterior weighted Paris physical residual". The DBN update result is not only written back to the PINN crack forward prediction model as new parameter values, but also directly enters the physical loss function of the PINN crack forward prediction model to participate in the construction of physical residuals, so that the newly added observation information can continuously affect the subsequent crack propagation prediction through the posterior parameter distribution.

[0049] Based on this, the loss function of the PINN crack forward prediction model is optimized as follows: , in, Indicates data loss; This represents the posterior weighted physical residual loss of DBN; This represents the loss due to monotonicity constraints. The weighting coefficients represent the posterior weighted physical residual loss of DBN; This represents the weighting coefficient of the monotonicity constraint loss.

[0050] Step 6: Construct the online correction loss function and locally update the correction head.

[0051] In the After several observation points, an online correction loss function is constructed using observation error, DBN posterior weighted physical residual loss, smoothing constraint, and correction magnitude constraint: ， in, Indicates the loss of observation error. This represents the posterior weighted physical residual loss of DBN. This represents the smoothing constraint loss. This indicates the correction amplitude constraint loss.

[0052] The observation error loss is: ， in, Indicates the first The cumulative number of load cycles corresponding to each online observation time; This indicates the cumulative number of load cycles. The crack length prediction value obtained after inputting the modified prediction model, i.e., the first... Corrected predicted crack length at each online update point; Indicates the first Crack length observations obtained at each online update point.

[0053] Smoothing constraint loss is used to avoid abrupt changes in the prediction curve caused by the correction head: ， in, This indicates the number of sampling points used to calculate the loss within the local correction interval; Indicates the first Each online update point corresponds to a local correction interval or a set of sampling points; Indicates the number of cycles for the lightweight calibration head. The local correction amount output at the location; This indicates that the lightweight calibration head is at the next adjacent sampling point. The local correction amount output at that point.

[0054] The correction amplitude constraint loss is used to limit the output amplitude of the correction head to prevent overcorrection. During online optimization, only the calibration head parameters are updated. Do not update the backbone network parameters of the PINN crack forward prediction model .

[0055] The calibration head parameters are trained using an online calibration loss function. The final corrected head parameters obtained after training with the online corrected loss function are: .

[0056] Step 7: Update parameter feedback and continue prediction.

[0057] After completing the online update of DBN parameters and the optimization of the correction head parameters, the updated parameters are fed back into the PINN crack forward prediction model: ， Complete the first After the first online update, the update point is used as the new prediction starting point to predict the subsequent crack propagation process. Let... For the first The cumulative load cycle count corresponding to each online update point, i.e., the total number of cycles the structure has undergone from the start of monitoring or service to the observation time; let... Let be the cumulative load cycle number corresponding to the time to be predicted, and Then, in the cumulative number of loops... The predicted crack length at the location is expressed as: ， If no new observation data is obtained, the updated posterior parameters and the current correction head are used to continue prediction. If new crack length observations are obtained again, the above steps are repeated.

[0058] Specifically, firstly, the PINN crack forward prediction model is based on the number of cycles and the Paris parameter. , The system outputs a predicted crack length and provides forward prediction results for different parameter particles during the update process of the DBN parameter online update model. The DBN parameter online update model updates the parameter particle weights based on the error between the predicted and observed values, obtaining the posterior distribution of the Paris parameters. Subsequently, this application does not simply write back the posterior mean to the PINN crack forward prediction model, but introduces the posterior parameter particles and their weights obtained from the DBN parameter online update model into the PINN physical loss function to construct the DBN posterior weighted physical residual loss function. At the same time, a lightweight correction head is set at the PINN output to correct local prediction biases after the observation point online while freezing the PINN backbone network.

[0059] Through the above process, this application forms a PINN-DBN bidirectional fusion closed loop: First, PINN→DBN: The PINN crack forward prediction model provides forward predictions for the parameter updates of the DBN parameter online update model. Second, DBN→PINN: The updated parameters, particle weights, and observation point influence range of the DBN parameter online update model are fed into the PINN physical loss function, driving the lightweight correction head to perform parameter optimization and local correction. This ensures that the observation information can continuously influence the subsequent crack propagation prediction process, rather than merely remaining at the parameter feedback level.

[0060] To verify the effectiveness of the method described in this application, fatigue crack propagation data of a single-sided notched tensile specimen with an edge crack, i.e., a SENT / SE(T) specimen, was used as the actual fatigue test data in this embodiment. The specimen material was S355 structural steel, with a thickness of 5 mm, a length of 240 mm, a width of 35 mm, and an initial crack length of 5 mm. During the test, the left end of the specimen was fixed, and a cyclic tensile load was applied to the right end, with a stress range of 150 MPa and a stress ratio of 0. The fatigue test and the SENT / SE(T) specimen are shown in the attached figure. Figure 2 As shown.

[0061] Fatigue crack propagation tests were conducted to obtain different numbers of cycles. The corresponding crack length This process generates realistic crack length-cycle count data. The complete experimental curve serves as the prediction and evaluation benchmark, while the crack length at preset observation points is used as the online observation value to trigger online updates of the DBN parameter model, the construction of the DBN posterior weighted physical residual loss, and the local update of the lightweight correction head. Simultaneously, a two-dimensional planar equivalent model consistent with the experimental conditions is established to generate crack propagation simulation data for training the PINN crack forward prediction model; the real experimental data is primarily used for online observation input and prediction result evaluation.

[0062] like Figures 3 to 7 As shown, the method described in this application can effectively track the crack propagation trend in real fatigue tests. At the preset observation point, DBN updates the Paris parameters based on the observed crack length. and Furthermore, the posterior parameter particles and their weights are incorporated into the PINN physical loss function, while a lightweight correction head corrects local prediction biases. Figure 6 The magnified view shows that the prediction curve undergoes a phased correction after receiving the observation information, indicating that the present invention can continuously apply the observation information to the subsequent crack propagation prediction through a closed-loop process of "PINN forward prediction - DBN parameter update - posterior weighted Paris physical residual construction - lightweight correction head local correction - PINN continued prediction".

[0063] Furthermore, comparative experiments with different observation densities and noise levels were conducted based on real fatigue test data to simulate conditions such as sparse engineering inspection points and measurement errors. For example... Figures 8 to 10 As shown, under conditions of dense observations, sparse observations, and different noise levels, the method described in this application can maintain a relatively stable prediction trend with small changes in RMSE. The results show that this application can still achieve parameter updates, dynamic adjustment of physical constraints, and staged corrections under limited, sparse, and noisy observation conditions, thereby improving the stability and physical consistency of crack propagation prediction.

[0064] The knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades provided by this invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from the principles of this invention, and these improvements and modifications also fall within the protection scope of the claims of this invention. The above description of the disclosed embodiments enables those skilled in the art to implement or use this invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of this invention. Therefore, this invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades, characterized in that, Includes the following steps: S1. Establish a parametrically conditional PINN crack forward prediction model; S2. Set a lightweight correction head at the output of the PINN crack forward prediction model to construct a modified prediction model; S3. Construct the basic loss function for the PINN crack forward prediction model; S4. Construct an online DBN parameter update model and call the PINN crack forward prediction model for forward prediction; S5. Construct the DBN posterior weighted physical residual loss and optimize the loss function of the PINN crack forward prediction model; S6. Construct the online correction loss function and update the lightweight correction head; S7. Update parameter feedback and predict crack length; In step S5, a lightweight correction head is first used to perform preliminary local correction on the original prediction results of the PINN crack forward prediction model, resulting in the corrected crack length prediction value: ， in, This indicates the number of cycles in the PINN crack forward prediction model. The predicted value of the original crack length is output below; Indicates the number of cycles for the lightweight calibration head. The local correction amount of the output; Parameters indicating a lightweight calibration head; This represents the predicted crack length after correction by the lightweight correction head. The posterior parameter particles obtained by updating DBN , and its posterior particle weights Introducing the physical loss function of the PINN crack forward prediction model, we construct the DBN posterior weighted physical residual loss function: ， in, Indicates the number of online update points; Indicates the first One online update point; Indicates the first The predicted interval endpoint after one online update point; Indicates the first Each online update point corresponds to the number of subsequent loops. Influence weight; Indicates the number of particles in the DBN parameter; Indicates the number of the parameter particle, and ; Indicates the first The number of each online update point is: The posterior weights of the parameter particles; This indicates that the number of particles in the posterior of the DBN is... The crack propagation coefficient; This indicates that the number of particles in the posterior of the DBN is... Crack propagation index; Indicates the corrected crack length The corresponding range of stress intensity factors; At this point, the loss function of the PINN crack forward prediction model is optimized as follows: ， in, Indicates data loss; This represents the posterior weighted physical residual loss of DBN; This represents the loss due to monotonicity constraints. The weighting coefficients represent the posterior weighted physical residual loss of DBN; This represents the weighting coefficient of the monotonicity constraint loss.

2. The knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades according to claim 1, characterized in that, In step S1, the PINN crack forward prediction model is expressed as: ， in, Indicates the number of loop iterations; Indicates the crack propagation coefficient; Indicates the crack propagation index; This represents the backbone network of the PINN crack forward prediction model; This represents the backbone network parameters of the PINN crack forward prediction model; This represents the crack length prediction value of the PINN crack forward prediction model.

3. The knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades according to claim 1, characterized in that, In step S2, the modified prediction model includes the PINN crack forward prediction model and a lightweight correction head; A lightweight correction head is installed at the output of the PINN crack forward prediction model. At this point, the crack prediction result output by the PINN crack forward prediction model is: in, This indicates a lightweight correction head used to output crack length correction amounts; This indicates the parameters of the calibration head.

4. The knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades according to claim 1, characterized in that, In step S3, during the training phase of the PINN crack forward prediction model, a basic joint loss function is constructed using data loss, physical loss, and monotonicity constraints: ， in, Indicates data loss. Indicates physical loss. This represents the monotonicity constraint loss. Weights representing physical losses; The weights represent the monotonicity constraint loss.

5. The knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades according to claim 1, characterized in that, In step S4, the DBN parameter online update model uses crack propagation parameters as the update object: ， in, Indicates the first The parameter state vector of a crack propagation parameter particle; Indicates the first Crack propagation coefficients corresponding to each parameter particle; Indicates the first The crack propagation index corresponding to each parameter particle; Generate multiple sets of candidate parameter particles around the current parameter state: ， in, Indicates the number of particles; Indicates the first At the first update time The parameter state vector of each candidate parameter particle; Indicates the first At the first update time The crack propagation coefficient corresponding to each candidate parameter particle; Indicates the first At the first update time The crack propagation index corresponding to each candidate parameter particle; Indicates the number of the parameter particle; For each set of parameter particles, the PINN crack forward prediction model is called to perform forward prediction, and the predicted crack length value corresponding to that parameter particle is obtained: ， Based on the predicted crack length and the observed crack length Error between Calculate particle weights: ， ， in, Indicates the first The online update time is numbered as follows The unnormalized weights of the candidate parameter particles; This represents the variance of the observed noise level for crack length. And normalize the particle weights: ， From this, we can obtain the posterior parameter estimates: ， ， in, Indicates the first The posterior estimate of the crack propagation coefficient at each online update time; Indicates the first The posterior estimate of the crack propagation index at each online update time.

6. The knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades according to claim 1, characterized in that, In step S6, at the first After several observation points, an online correction loss function is constructed using observation error, DBN posterior weighted physical residual loss, smoothing constraint, and correction magnitude constraint: ， in, Indicates the loss of observation error. This represents the posterior weighted physical residual loss of DBN. This represents the smoothing constraint loss. This indicates the correction amplitude constraint loss; The observation error loss is: ， in, Indicates the first The number of loops or loading steps corresponding to each online update point This indicates the cumulative number of load cycles. The crack length prediction value obtained after inputting the modified prediction model, i.e., the first... Corrected predicted crack length at each online update point; Indicates the first Crack length observations obtained at each online update point; The smoothing constraint loss is: ， in, This indicates the number of sampling points used to calculate the loss within the local correction interval; Indicates the first Each online update point corresponds to a local correction interval or a set of sampling points; Indicates the number of cycles for the lightweight calibration head. The local correction amount output at the location; This indicates that the lightweight calibration head is at the next adjacent sampling point. The local correction amount output at the location; The correction amplitude constraint loss is: The calibration head parameters are trained using an online calibration loss function. The final corrected head parameters obtained after training are: .

7. The knowledge-driven real-time prediction method for fatigue cracks in offshore wind turbine blades according to claim 6, characterized in that, In step S7, The updated parameters are fed back into the PINN crack forward prediction model: ， The crack length is predicted to be: ， in, Indicates the first The cumulative load cycle count corresponding to each online update point This represents the cumulative load cycle count corresponding to the time to be predicted, and .

Images