ETFE membrane structure life prediction method and system based on machine learning

Abstract

The invention relates to the technical field of ETFE film testing, and discloses an ETFE film structure life prediction method and system based on machine learning, and the method comprises the steps: carrying out the standardization preprocessing of the cyclic load fatigue data of an ETFE film structure in a biaxial stress state, and obtaining a standard feature vector; performing membrane structure geometric weight distribution on the standard feature vector to obtain a weighted feature vector; performing multi-time scale damage feature extraction based on the weighted feature vector to obtain a fused damage feature; inputting the fused damage features into a fatigue damage evolution physical constraint embedder for constraint calculation to obtain physical constraint damage features; bayesian probability life prediction is carried out on the physical constraint damage features, and a fatigue life prediction value of the ETFE film structure is obtained.According to the method, key damage parameters of different deformation stages can be recognized in a self-adaptive mode, the characterization precision of the fatigue behavior of the film structure is remarkably improved, and the life prediction accuracy of the ETFE film structure is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ETFE membrane testing, and in particular to a method and system for predicting the life of an ETFE membrane structure based on machine learning. Background Art

[0002] As a new type of lightweight building material, ETFE (ethylene-tetrafluoroethylene copolymer) membrane structure is widely used in large-span buildings, stadiums, exhibition centers and other projects due to its excellent light transmittance, weather resistance and mechanical properties.

[0003] However, ETFE membrane structures are subject to long-term cyclic loads such as wind and snow during their service life, and their fatigue performance is directly related to the safety and durability of the structure. Due to the complex characteristics of membrane structures, such as large deformation, geometric nonlinearity, and biaxial stress states, traditional life prediction methods based on the linear superposition principle and uniaxial fatigue theory cannot accurately describe their fatigue behavior. Their prediction accuracy is limited and they lack quantitative assessment of uncertainty. Summary of the Invention

[0004] The present invention provides a method and system for predicting the life of ETFE membrane structures based on machine learning. The present invention can adaptively identify key damage parameters in different deformation stages, significantly enhance the accuracy of characterizing the fatigue behavior of membrane structures, and improve the accuracy of ETFE membrane structure life prediction.

[0005] A first aspect of the present invention provides an ETFE membrane structure life prediction method based on machine learning, the ETFE membrane structure life prediction method based on machine learning comprising: The cyclic load fatigue data of ETFE membrane structure under biaxial stress state is standardized and preprocessed to obtain the standard eigenvector; Performing membrane structure geometry weight distribution on the standard eigenvector to obtain a weighted eigenvector; Extracting multi-time-scale damage features based on the weighted feature vector to obtain a fused damage feature; Inputting the fused damage characteristics into a fatigue damage evolution physical constraint embedder for constraint calculation to obtain physical constraint damage characteristics; A Bayesian probability life prediction is performed on the physical constraint damage characteristics to obtain a fatigue life prediction value of the ETFE membrane structure.

[0006] In combination with the first aspect, in a first implementation of the first aspect of the present invention, the cyclic load fatigue data of the ETFE membrane structure under a biaxial stress state is subjected to standardized preprocessing to obtain a standard eigenvector, including: The cyclic loading fatigue data of the ETFE membrane structure under biaxial stress state was collected to obtain the original fatigue data including the principal stress component, principal strain component and number of loading cycles under biaxial stress state; Based on the original fatigue data, geometric parameters and damage-sensitive parameters are constructed to obtain a geometric parameter set including a membrane structure thickness variation parameter, a curvature radius variation parameter, and a principal stretch ratio parameter, and a damage-sensitive parameter set including a stress intensity factor amplitude, a crack growth rate, a maximum stress, a minimum stress, and a stress ratio; Stress and strain relationship operations and vector mapping are performed on the geometric parameter set and the damage sensitive parameter set to obtain a standard eigenvector.

[0007] In combination with the first aspect, in a second implementation of the first aspect of the present invention, performing membrane structure geometric weight distribution on the standard eigenvector to obtain a weighted eigenvector includes: Calculate the attention weight of the standard feature vector to obtain an initial weight distribution matrix; Calculating a geometric nonlinear factor based on the initial weight distribution matrix to obtain a geometric nonlinear factor; Adaptively adjust the weight of the damage sensitive parameter set according to the geometric nonlinear factor to obtain a weight distribution coefficient; Performing an element-by-element product operation on the standard feature vector and the weight distribution coefficient to obtain a weighted feature vector.

[0008] In combination with the first aspect, in a third implementation of the first aspect of the present invention, the calculating of the geometric nonlinear factor based on the initial weight distribution matrix to obtain the geometric nonlinear factor includes: extracting a first principal stretch ratio parameter and a second principal stretch ratio parameter from the initial weight distribution matrix; Calculating the thickness change ratio and the curvature radius change ratio based on the geometric parameter set to obtain the membrane structure thickness change ratio and the membrane structure curvature radius change ratio; Performing geometric nonlinear calculation based on the first principal stretching ratio parameter, the second principal stretching ratio parameter, the membrane structure thickness change ratio, and the membrane structure curvature radius change ratio to obtain an initial nonlinear factor of the ETFE membrane in a large deformation state; The initial nonlinear factor is corrected for its membrane structure deformation history correlation according to the number of load cycles to obtain a geometric nonlinear factor.

[0009] In combination with the first aspect, in a fourth implementation of the first aspect of the present invention, extracting multi-time-scale damage features based on the weighted feature vector to obtain a fused damage feature includes: Inputting the weighted eigenvectors into the instantaneous damage increment branch, the cumulative damage trend branch and the critical damage threshold approximation branch respectively; The instantaneous damage increment branch performs convolution operation and activation processing on the weighted feature vector to obtain an instantaneous damage increment feature; the cumulative damage trend branch performs state update processing on the weighted feature vector to obtain a cumulative damage trend feature; the critical damage threshold approximation branch performs residual connection calculation on the weighted feature vector to obtain a critical damage threshold approximation feature; The instantaneous damage increment feature, the cumulative damage trend feature and the critical damage threshold approximation feature are fused in time scale to obtain a fused damage feature.

[0010] In combination with the first aspect, in a fifth implementation of the first aspect of the present invention, inputting the fused damage feature into a fatigue damage evolution physical constraint embedder for constraint calculation to obtain the physical constraint damage feature includes: Inputting the fused damage features into a fatigue damage evolution physical constraint embedder respectively, wherein the fatigue damage evolution physical constraint embedder includes a Chaboche damage evolution constraint module, a modified Paris formula constraint module, and a viscoelastic constitutive relation constraint module; Performing a continuous damage mechanics equation calculation on the fusion damage feature through the Chaboche damage evolution constraint module to obtain a first constraint feature; Performing a biaxial stress state crack propagation calculation on the fusion damage feature using the modified Paris formula constraint module to obtain a second constraint feature; Performing a Maxwell model calculation on the fusion damage feature through the viscoelastic constitutive relationship constraint module to obtain a third constraint feature; Constraint violation inspection and hard constraint fusion are performed on the first constraint feature, the second constraint feature, and the third constraint feature to obtain a physical constraint damage feature.

[0011] In combination with the first aspect, in a sixth implementation of the first aspect of the present invention, the Chaboche damage evolution constraint module performs a continuous damage mechanics equation calculation on the fusion damage feature to obtain a first constraint feature, including: extracting an equivalent stress value and a current damage value from the fusion damage feature; The Chaboche damage evolution constraint module calculates the Chaboche damage evolution equation based on the equivalent stress value and preset ETFE membrane material parameters to obtain a theoretical damage evolution rate, wherein the preset ETFE membrane material parameters include a damage evolution coefficient, a stress exponent, and a damage exponent; Inputting the theoretical damage evolution rate and the current damage value into the physical constraint layer to perform residual calculation to obtain a damage evolution constraint residual value; The fusion damage feature is corrected by continuous damage mechanics law according to the damage evolution constraint residual value to obtain a first constraint feature.

[0012] In combination with the first aspect, in a seventh implementation of the first aspect of the present invention, performing Bayesian probabilistic life prediction on the physical constraint damage characteristics to obtain a fatigue life prediction value of the ETFE membrane structure includes: Inputting the physical constraint damage feature into a Bayesian neural network to calculate probability distribution parameters to obtain first weight probability distribution parameters and first probability distribution parameters, wherein the first weight probability distribution parameters include a weight mean and a weight variance, and the first probability distribution parameters include a bias mean and a bias variance; performing a variational inference calculation based on the first weighted probability distribution parameter and the first probability distribution parameter to obtain a mean and a variance of fatigue life prediction; Calculating the confidence interval based on the mean and variance of the fatigue life prediction to obtain an upper limit and a lower limit of the confidence interval for the fatigue life corresponding to a 95% confidence level; The fatigue life prediction value of the ETFE membrane structure is obtained by performing probabilistic life integration on the mean value, the variance, the upper limit value of the confidence interval and the lower limit value of the confidence interval of the fatigue life prediction.

[0013] In combination with the first aspect, in an eighth implementation of the first aspect of the present invention, performing a variational inference calculation based on the first weight probability distribution parameter and the first probability distribution parameter to obtain a mean and variance of fatigue life prediction includes: Constructing a variational posterior distribution based on the first weight probability distribution parameter and the first probability distribution parameter to obtain a weighted variational posterior distribution and a biased variational posterior distribution; Inputting the weighted variational posterior distribution and the biased variational posterior distribution into the evidence lower bound loss function to perform ELBO calculation to obtain an evidence lower bound loss value including a data fitting term and a KL divergence regularization term; Performing gradient descent optimization according to the lower bound loss value of the evidence to obtain a second weight probability distribution parameter and a second probability distribution parameter; Fatigue life probability distribution calculation is performed based on the second weight probability distribution parameter and the second probability distribution parameter to obtain a mean and variance of fatigue life prediction.

[0014] A second aspect of the present invention provides an ETFE membrane structure life prediction system based on machine learning, the ETFE membrane structure life prediction system based on machine learning comprising: A preprocessing module is used to perform standardized preprocessing on the cyclic load fatigue data of the ETFE membrane structure under biaxial stress state to obtain a standard eigenvector; A weight distribution module is used to distribute the geometric weight of the membrane structure to the standard eigenvector to obtain a weighted eigenvector; A feature extraction module, configured to extract multi-time-scale damage features based on the weighted feature vector to obtain a fused damage feature; a constraint calculation module, configured to input the fused damage feature into a fatigue damage evolution physical constraint embedder for constraint calculation to obtain a physical constraint damage feature; The life prediction module is used to perform Bayesian probability life prediction on the physical constraint damage characteristics to obtain a fatigue life prediction value of the ETFE membrane structure.

[0015] Compared with the existing technology, the present invention has the following advantages: By establishing a data preprocessing method specifically for the biaxial stress state of ETFE membrane structures, a standard feature vector containing principal stress components, principal strain components, a set of geometric parameters, and a set of damage-sensitive parameters is constructed. Compared with traditional uniaxial fatigue data processing methods, this method can more comprehensively reflect the complex stress state and large-deformation geometric nonlinear behavior of membrane structures. By introducing a membrane structure geometric nonlinear adaptive weight module, the weight distribution of damage-sensitive parameters is dynamically adjusted according to the real-time deformation state of the ETFE membrane through an attention mechanism. Compared with the fixed weight method of traditional neural networks, this method can adaptively identify key damage parameters at different deformation stages, significantly enhancing the characterization accuracy of membrane structure fatigue behavior. A multi-branch parallel architecture is used to simultaneously extract instantaneous damage increment features, cumulative damage trend features, and critical damage threshold approximation features. Compared with the single time-scale damage analysis methods in the existing technology, this method can comprehensively capture the different stage characteristics of microcrack initiation, macrocrack propagation, and overall structural failure during the ETFE membrane fatigue process, achieving a refined description of the fatigue damage evolution process. By embedding the Chaboche damage evolution equation, the modified Paris formula, and the viscoelastic constitutive relation as physical laws into a neural network through a hard constraint mechanism, this approach ensures that the prediction results strictly adhere to the laws of continuous damage mechanics, compared to the existing approach of treating physical laws as soft constraints. This addresses the key issue of traditional data-driven prediction methods lacking physical plausibility. A Bayesian neural network is used to quantify the uncertainty of fatigue life prediction. A variational inference method is used to output the mean, variance, and confidence interval of fatigue life. Compared to traditional deterministic prediction methods that only provide point estimates, this approach can provide a basis for risk assessment and predictive reliability evaluation for engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] The structures, proportions, sizes, etc. depicted in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not intended to limit the conditions under which the present invention can be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportional relationships, or adjustments in size should still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and objectives that can be achieved by the present invention.

[0018] Figure 1 1 is a flow chart of a method for predicting the life of an ETFE membrane structure based on machine learning provided by an embodiment of the present invention; Figure 2 It is a structural schematic block diagram of the ETFE membrane structure life prediction system based on machine learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined, or partially merged, so the actual execution order may vary depending on the actual situation.

[0021] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the present invention. As used in the specification and appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0022] It should be further understood that the term "and / or" used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1In one embodiment of the present invention, an embodiment of the ETFE membrane structure life prediction method based on machine learning includes: Step 100: performing standardization preprocessing on cyclic load fatigue data of the ETFE membrane structure under a biaxial stress state to obtain a standard eigenvector; It is understood that the execution subject of the present invention can be the ETFE membrane structure life prediction system based on machine learning, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0023] Specifically, fatigue data of an ETFE membrane structure under biaxial stress was collected. This process involved setting a standard loading path and combining it with precision measurement equipment, such as a biaxial loading platform and a high-precision strain measurement system, to obtain raw fatigue data. The collected data included the principal stress components σ1 and σ2, as well as the principal strain components ε1 and ε2, corresponding to different numbers of cycles N. These data reflect the stress and strain evolution of the membrane material under actual loading conditions, particularly the material degradation caused by changes in stress amplitude and strain during fatigue accumulation. Based on the raw fatigue data, a set of geometric parameters and damage-sensitive parameters related to the fatigue characteristics of the ETFE membrane structure was constructed. This set of geometric parameters aims to capture the morphological evolution of the membrane material under large deformations, specifically the variation of the membrane thickness t(N), the structural curvature radius R(N), and the dynamic evolution of the principal stretch ratios λ1 and λ2 as the load cycles progress. By tracking these parameters, localized thinning, increased deformation, or stress concentration in the membrane material due to accumulated damage can be detected. In terms of damage-sensitive parameter construction, highly indicative indicators of fatigue life are extracted, such as the stress intensity factor amplitude, which describes the stress field strength at the crack tip; the crack growth rate da / dN, which reflects the speed of crack growth and its acceleration trend; the maximum stress σmax and minimum stress σmin, which reveal the ultimate load range of the material during loading; and the stress ratio R, defined as the ratio of σmin to σmax, which reflects the loading characteristics and cyclic asymmetry during fatigue cycling. Stress and strain relationship calculations and vector mapping are performed on the geometric parameter set and the damage-sensitive parameter set. The original principal stress and strain data are normalized to eliminate scale inconsistencies caused by dimensional differences and ensure that each feature has a balanced influence in the subsequent machine learning process. The normalized principal stresses, principal strains, film thickness changes, curvature radius changes, principal stretch ratios, and various fatigue damage-sensitive parameters are vectorized to form a unified high-dimensional feature vector set. The cycle number N is introduced as a time marker in the vector mapping stage, ensuring that the feature vector not only describes the current state of the material but also contains historical information, resulting in a standardized feature vector.

[0024] Step 200: Distribute the geometric weight of the membrane structure to the standard eigenvector to obtain a weighted eigenvector; Specifically, attention weights are calculated for the standard eigenvectors. Based on the attention mechanism in deep learning, a trainable attention network is constructed. The standard eigenvectors are input into the attention weight calculation module to obtain an initial weight distribution matrix. This matrix provides a preliminary assessment of the importance of different features, dynamically learning the contribution of each feature to life prediction under the current fatigue state, and thus reflecting the dominant effect of various physical properties of the ETFE membrane structure during a specific load cycle. A geometric nonlinearity factor is calculated based on the initial weight distribution matrix to enhance the weighting mechanism's responsiveness to the large deformation characteristics of the membrane structure. The geometric nonlinearity factor is constructed based on the changing trends of the membrane thickness change rate t(N) / t0, the curvature radius change rate R(N) / R0, and the principal stretch ratios λ1 and λ2. This comprehensive dimensionless quantity describes both the degree of local geometric distortion caused by fatigue accumulation and the overall nonlinear evolution of the structure. Based on the geometric nonlinearity factor, the set of damage-sensitive parameters in the standard eigenvectors is adaptively weighted to obtain a dynamically changing weight distribution coefficient. The specific adjustment strategy follows the principle of nonlinear factor dominance. When the geometric nonlinear factor is low, it indicates that the membrane structure is still in a stage close to linear elasticity. At this time, the weight distribution is more biased towards linear characteristic parameters such as stress and strain. When the geometric nonlinear factor increases, it indicates that the membrane structure has entered a stage of large deformation or fatigue acceleration. The weight gradually tilts towards nonlinear damage evolution parameters such as crack growth rate and stress intensity factor. Through the adaptive weight adjustment mechanism, the model can flexibly capture the key characteristic changes in each stage of the membrane structure fatigue process, from small deformation to extreme failure, thereby effectively improving the accuracy and robustness of the feature expression. The adaptively adjusted weight distribution coefficient is element-wise multiplied with the standard eigenvector to obtain the final weighted eigenvector.

[0025] Parameters related to the membrane structure's stress state are extracted from the initial weight distribution matrix: the first principal stretch ratio parameter λ1 and the second principal stretch ratio parameter λ2. These two parameters, representing the strain amplification factors in the principal directions of the membrane material under biaxial stress, effectively describe the characteristics of the ETFE membrane's isotropic tensile deformation under external loads and serve as important indicators of the membrane structure's strain anisotropy. Based on the extracted principal stretch ratio parameters, the thickness change ratio and curvature radius change ratio are calculated, combined with the thickness change and curvature change information contained in the geometric parameter set. The thickness change ratio of the membrane structure, defined as the ratio of the current thickness t(N) to the initial thickness t0, characterizes local thinning caused by material flow and microdamage accumulation during fatigue loading. The curvature radius change ratio, determined by the ratio of the current curvature radius R(N) to the initial curvature radius R0, reflects the evolution of the membrane surface's overall geometric shape under load. A comprehensive calculation of the geometric nonlinearity factor is performed based on the first principal stretch ratio λ1, the second principal stretch ratio λ2, the membrane structure thickness variation ratio t(N) / t0, and the curvature radius variation ratio R(N) / R0. This calculation constructs a dimensionless integrated quantity that takes into account the isotropic tensile deformation along with the thickness and curvature variations, resulting in an initial nonlinearity factor that characterizes the large deformation state of the ETFE membrane. This initial nonlinearity factor captures the deformation extent and nonlinear intensity of the membrane structure under complex loading conditions and quantifies the structure's gradual transition from the linear elastic stage to the geometric nonlinear stage. By introducing the number of load cycles N as a time variable and combining it with the hysteresis characteristics of fatigue damage evolution in the membrane material, the initial nonlinearity factor is corrected for deformation history. This correction utilizes an exponential decay function to gradually accumulate the effects of historical cyclic loads on the current nonlinearity. The corrected geometric nonlinearity factor comprehensively accounts for the deformation trajectory of the membrane structure from initial loading to the current state, effectively reflecting the cumulative effects of material stiffness degradation, fatigue softening, and geometric distortion.

[0026] Step 300: extract multi-time-scale damage features based on the weighted feature vector to obtain fused damage features; It should be noted that the weighted feature vectors are input into the instantaneous damage increment branch, the cumulative damage trend branch, and the critical damage threshold approximation branch, respectively. Each branch is responsible for feature extraction at a different time scale, collaboratively modeling the short-term changes, long-term evolution, and near-failure behavior of membrane structures during fatigue. In the instantaneous damage increment branch, a convolution operation is performed on the input weighted feature vector, selecting a convolution kernel suitable for a small time window to extract local variation patterns within a single or a few cyclic loading cycles. The convolution operation effectively captures fine-grained mutational characteristics during damage evolution, such as microcrack initiation and localized stress concentration. After convolution, a nonlinear activation function, ReLU activation, is introduced to enhance the nonlinearity of feature representation and prevent the vanishing gradient problem. This results in an instantaneous damage increment feature that reflects the instantaneous fatigue damage changes of the ETFE membrane structure. Simultaneously, the weighted feature vector is input into the cumulative damage trend branch, which is based on a long short-term memory (LSTM) architecture and employs a gating mechanism to update the state of time series data. By storing and updating cyclic states, the LSTM effectively captures the cumulative trend of fatigue damage in membrane structures under long-term loading. Its internal cell state dynamically adjusts with each cyclic data input, enabling it to memorize long-range dependencies. This makes it particularly well-suited for modeling the progressive evolution of membrane materials, such as microscopic damage accumulation, material hardening, or localized plastic deformation, caused by long-term stress. It outputs cumulative damage trend features, which reveal the evolution of fatigue cracks from initiation to propagation and ultimately to macroscopic failure. The weighted feature vectors are fed into the critical damage threshold approximation branch. A residual connection architecture is employed, directly superimposing the input with the intermediate layer output via skip connections to better capture the gradual approach of damage features to the critical threshold as fatigue progresses. Residual connections help alleviate the vanishing gradient problem during deep network training and preserve the continuity between original and higher-order features. This makes it particularly well-suited for modeling the critical behavior of membrane structures, where damage evolution rapidly transitions from a stable phase to a failure phase. After residual computation, the resulting critical damage threshold approximation features reflect the accelerated degradation trend of membrane materials nearing failure. The instantaneous damage increment features, cumulative damage trend features, and critical damage threshold approximation features are temporally fused. The fusion operation is completed through a weighted summation or attention weighted mechanism, which dynamically adjusts the contribution ratio of each time scale feature to finally obtain the fused damage feature.

[0027] Step 400: Input the fused damage feature into the fatigue damage evolution physical constraint embedder to perform constraint calculation to obtain the physical constraint damage feature; Specifically, the fused damage signature is fed into three independent physical constraint modules within the fatigue damage evolution physical constraint embedder: the Chaboche damage evolution constraint module, the modified Paris formula constraint module, and the viscoelastic constitutive relation constraint module. This modular design allows for independent modeling and collaborative constraints for different physical laws. In the Chaboche damage evolution constraint module, the fused damage signature is used to calculate the continuum damage mechanics equation. Based on the continuum damage mechanics theory, the fatigue damage evolution process is modeled as a nonlinear growth of the damage variable with the number of cycles. Within the module, the damage evolution equation is applied based on the membrane material parameters to calculate the damage evolution rate corresponding to the fused damage signature. This output is the first constraint signature, which reflects the stiffness degradation, damage accumulation rate, and fatigue life consumption trends of the membrane structure during fatigue loading. Because ETFE membrane exhibits significant fatigue softening, the Chaboche model effectively captures the regularity of the damage variable's gradual growth with increasing loading cycles. The fused damage signature is fed into the modified Paris formula constraint module, where the relationship between the crack tip stress intensity factor and the crack growth rate is calculated based on the crack growth mechanism under biaxial stress. The modified Paris formula incorporates a biaxial stress correction factor based on the traditional uniaxial stress crack growth formula. This accurately accounts for the crack growth behavior of membrane structures under complex loading conditions, particularly the influence of the principal stress ratio on the crack growth rate. By calculating the stress intensity factor and crack growth rate of the fused damage signature, a second constraint signature is derived. This signature effectively characterizes the evolution of microcrack initiation and growth rates within the membrane material, and is of great significance for fatigue failure prediction. Simultaneously, the fused damage signature is input into the viscoelastic constitutive relation constraint module. Based on the Maxwell viscoelastic model, this module utilizes the coupled relationship between stress and strain rate to calculate the deformation response of the membrane structure during fatigue loading. By incorporating viscous damping effects and elastic modulus parameters, the module simulates the stress relaxation and hysteresis characteristics of ETFE membrane materials at various loading rates and cyclic frequencies, resulting in a third constraint signature. Constraint violation tests are performed on the first, second, and third constraint signatures. By calculating the residuals of each physical equation, the physical compliance of the fused damage signature is assessed. When the residual error of a physical equation exceeds a preset threshold, indicating that the corresponding feature violates physical laws, the hard constraint mechanism is triggered. By optimizing and adjusting the model parameters, the model is forced back to a feature space that satisfies the physical constraints. After constraint violation verification and hard constraint fusion, the physical constraint damage feature is obtained.

[0028] The equivalent stress value and current damage value are extracted from the fused damage signature. The equivalent stress value, as a scalar representation of the membrane structure's comprehensive stress state under multiaxial loading, effectively reflects the dominant stress level of local fatigue accumulation, while the current damage value describes the degree of damage accumulation in the material under the current fatigue cycle. Based on the extracted equivalent stress value and pre-set ETFE membrane material parameters, the continuous damage mechanics equation is solved in the Chaboche damage evolution constraint module. The Chaboche damage evolution equation utilizes the nonlinear relationship between equivalent stress and damage variables and defines an expression for the theoretical damage evolution rate based on the material's inherent fatigue properties. The required ETFE membrane material parameters primarily include the damage evolution coefficient, stress exponent, and damage exponent. The damage evolution coefficient controls the overall rate of damage evolution, the stress exponent describes the sensitivity of the damage rate to stress level, and the damage exponent reflects the modulation effect of the damage variable on its own accumulation rate. These parameters, obtained through experimental testing, accurately reflect the damage evolution patterns of membrane materials under different fatigue loading conditions. On this basis, the Chaboche equation can deduce the corresponding theoretical damage evolution rate based on the current equivalent stress level and predict the ideal damage growth trend of the membrane material under the current loading conditions. The theoretical damage evolution rate and the current damage value extracted from the fused damage feature are input into the physical constraint layer to perform residual calculations to measure the degree of deviation between the theoretical damage evolution rate and the actual damage evolution rate. By solving the error norm between the two, the residual value of the damage evolution constraint is obtained. If the residual value is small, it means that the fused feature conforms well to the law of continuous damage mechanics; if the residual value is too large, it indicates that there is a physical violation and the fused feature needs to be further adjusted to be closer to the physical real process. Based on the calculated damage evolution constraint residual value, a constraint correction mechanism is used to correct the fused damage feature according to the law of continuous damage mechanics. During the correction process, a gradient-based optimization method was used, with the damage evolution constraint residual as the objective function. Backpropagation was used to adjust the sensitive components of the fused features, gradually reducing the damage evolution residual until it fell below a preset threshold. This ensured that the corrected features not only conformed to the fatigue evolution trend in numerical fitting but also met the fatigue damage evolution laws of ETFE membrane materials in terms of physical mechanism. After the correction process, the first constraint feature was obtained.

[0029] Step 500: Perform Bayesian probabilistic life prediction on the physical constraint damage characteristics to obtain a fatigue life prediction value of the ETFE membrane structure.

[0030] Specifically, the physical constraint damage signature is input into a Bayesian neural network to calculate the probability distribution parameters. Bayesian neural networks incorporate probabilistic modeling in the setting of network weights and bias parameters, treating each network weight as a random variable obeying a certain probability distribution. At each layer of the neural network, the weight parameters are modeled using a parameterized probability distribution, employing a Gaussian distribution. Thus, each weight is assigned a mean and variance, forming the first weight probability distribution parameters. Simultaneously, each bias parameter is also assumed to have a Gaussian distribution, resulting in a bias mean and bias variance, forming the first probability distribution parameters. Variational inference is performed based on the first weight probability distribution parameters and the first probability distribution parameters. Variational inference, an approximate Bayesian inference method, optimizes a set of parameterized distributions to make them as close as possible to the true posterior distribution, thereby avoiding the expensive posterior distribution integral calculation required in traditional Bayesian inference. In its implementation, the physical constraint damage signature is input into the network using an optimization method based on an evidence lower bound. Through multiple sampling and forward propagation, the mean and variance of the fatigue life prediction are obtained. Confidence intervals are calculated based on the statistical properties of the standard normal distribution. Based on the conventional 95% confidence level, the upper and lower limits of the fatigue life confidence interval are calculated by adding or subtracting 1.96 times the standard deviation from the mean, thereby statistically defining a reasonable range for the life prediction results. The mean, variance, and upper and lower limits of the fatigue life prediction are integrated to form the probabilistic life prediction results.

[0031] Based on the parameters of the first weight probability distribution and the first probability distribution, corresponding variational posterior distributions are constructed for the weights and biases in the neural network, respectively. These posterior distributions are assumed to be Gaussian distributions with learnable mean and variance. The weight variational posterior distribution is parameterized by the weight mean and weight variance, while the bias variational posterior distribution is described by the bias mean and bias variance. This approach directly incorporates model uncertainty into the distributional modeling of the network parameters. The constructed weight variational posterior distributions and bias variational posterior distributions are then fed into the Evidence Lower Bound (ELBO) loss function for computation. The ELBO, as the core optimization objective in variational inference, maximizes the ELBO to make the current variational posterior distribution as close as possible to the true Bayesian posterior distribution. The ELBO consists of two components: the first is a data fitting term, which measures the fitting error between the current model output and the true labeled data. This term is implemented through the expected calculation of the log-likelihood. The second is a Kullback-Leibler (KL) divergence regularization term, which measures the difference between the current variational posterior distribution and the prior distribution, thereby preventing overfitting and enhancing the model's generalization ability. By calculating the ELBO value of these two components and comprehensively considering the model's fitting accuracy and complexity control, the model achieves both accurate prediction capabilities and a low risk of overfitting when faced with complex fatigue life data. Based on the calculated evidence lower bound loss, the network parameters are optimized using gradient descent. By backpropagating the gradient of the ELBO with respect to the weight and bias distribution parameters, the weight mean, weight variance, bias mean, and bias variance are updated to obtain the second weight probability distribution parameters and the second probability distribution parameters. This process is essentially an iterative update of the variational posterior distribution parameters. As training progresses, the variational distribution gradually approaches the true posterior distribution, enabling the model to find the optimal uncertainty representation in parameter space. The fatigue life probability distribution is calculated based on the second weight probability distribution parameters and the second probability distribution parameters. Monte Carlo sampling is performed on the variational distribution of weights and biases, and forward inference is performed in conjunction with physical constraint damage characteristics to obtain the predicted fatigue life distribution. The fatigue life prediction results are accumulated through multiple sampling cycles, and the mean and variance of the fatigue life predictions are statistically calculated.

[0032] In an embodiment of the present invention, a data preprocessing method specifically tailored to the biaxial stress state of ETFE membrane structures is established, constructing a standard feature vector containing principal stress components, principal strain components, a set of geometric parameters, and a set of damage-sensitive parameters. Compared to traditional uniaxial fatigue data processing methods, this method can more comprehensively reflect the complex stress state and large-deformation geometric nonlinear behavior of membrane structures. By introducing a membrane structure geometric nonlinear adaptive weight module, the weight distribution of damage-sensitive parameters is dynamically adjusted according to the real-time deformation state of the ETFE membrane through an attention mechanism. Compared to the fixed weight approach of traditional neural networks, this method can adaptively identify key damage parameters at different deformation stages, significantly enhancing the accuracy of characterizing the fatigue behavior of membrane structures. A multi-branch parallel architecture is employed to simultaneously extract instantaneous damage increment features, cumulative damage trend features, and critical damage threshold approximation features. Compared to existing single-time-scale damage analysis methods, this method can comprehensively capture the characteristics of the different stages of ETFE membrane fatigue, including microcrack initiation, macrocrack propagation, and overall structural failure, achieving a refined description of the fatigue damage evolution process. By embedding the Chaboche damage evolution equation, the modified Paris formula, and the viscoelastic constitutive relation as physical laws into a neural network through a hard constraint mechanism, this approach ensures that the prediction results strictly adhere to the laws of continuous damage mechanics, compared to the existing approach of treating physical laws as soft constraints. This addresses the key issue of traditional data-driven prediction methods lacking physical plausibility. A Bayesian neural network is used to quantify the uncertainty of fatigue life prediction. A variational inference method is used to output the mean, variance, and confidence interval of fatigue life. Compared to traditional deterministic prediction methods that only provide point estimates, this approach can provide a basis for risk assessment and predictive reliability evaluation for engineering applications.

[0033] In a specific embodiment, the process of executing step 100 may specifically include the following steps: The cyclic loading fatigue data of the ETFE membrane structure under biaxial stress state was collected to obtain the original fatigue data including the principal stress component, principal strain component and number of loading cycles under biaxial stress state; Based on the original fatigue data, geometric parameters and damage-sensitive parameters are constructed to obtain a set of geometric parameters including membrane structure thickness variation parameters, curvature radius variation parameters, and principal stretch ratio parameters, and a set of damage-sensitive parameters including stress intensity factor amplitude, crack growth rate, maximum stress, minimum stress, and stress ratio. Stress and strain relationship operations and vector mapping are performed on the geometric parameter set and the damage sensitive parameter set to obtain the standard eigenvector.

[0034] Specifically, based on the physical properties of ETFE membrane, which is highly flexible and susceptible to biaxial stretching, a fatigue testing platform suitable for biaxial loading was constructed. Using a cross-tense tensile machine or a modified planar biaxial loading device, two independently controlled primary loading directions were set up to ensure that the membrane accurately simulated the biaxial tensile stress state experienced in actual engineering applications during testing. Initially, the membrane underwent pre-treatment, including ensuring initial thickness, initial flatness, and the absence of initial defects. Precision stress and strain measurement systems, such as strain gauge arrays and non-contact optical measurement systems, were also installed to enable high-precision, real-time acquisition of the principal stress components σ1 and σ2 and the principal strain components ε1 and ε2. During the loading process, the stress and strain response data of the membrane material were gradually recorded after multiple fatigue cycles, using various load amplitudes and number of cycles. The collected raw data covered the entire fatigue evolution process, from initial loading, crack initiation, crack propagation, to final failure, ensuring comprehensiveness and continuity. The number of load cycles, N, was also recorded. By simultaneously recording the principal stresses, principal strains, and N values ​​within each load cycle, a raw fatigue dataset containing the three fundamental physical quantities was generated. Geometric parameters and damage-sensitive parameters were constructed based on the original fatigue data. To comprehensively characterize the geometric changes of the ETFE membrane structure during fatigue, a set of geometric parameters was extracted from the stress-strain data. These parameters include the membrane thickness change parameter t(N). By measuring the thickness change of the membrane material after different load cycles, the thickness change rate t(N) / t0 was calculated relative to the initial thickness t0, reflecting the local thinning trend of the membrane material under fatigue accumulation. The curvature radius change parameter R(N) was calculated based on the membrane surface deformation measurement results. The local curvature change was normalized with the initial curvature R0 to obtain the curvature radius change ratio R(N) / R0, reflecting the overall geometric evolution of the membrane structure. The principal stretch ratio parameters λ1 and λ2 were calculated from the principal strain components, λ1=1+ε1 and λ2=1+ε2, respectively. They describe the strain amplification degree of each principal axis of the membrane under biaxial tension, thereby accurately characterizing the non-uniform stretching state under multi-axial stress. Furthermore, a set of damage-sensitive parameters was constructed based on the original fatigue data to capture the crack evolution and stress response characteristics of the ETFE membrane structure during fatigue. The stress intensity factor amplitude is calculated by stress field theory under biaxial loading, which can describe the stress concentration effect at the crack tip and is an important control parameter for the crack growth rate. The crack growth rate da / dN is calculated by regularly measuring the crack size and recording the number of cycles to calculate the average rate of crack growth, reflecting the accumulation rate of microscopic damage. The maximum stress σ max and minimum stress σ min The stress peak and valley data directly derived from each cycle describe the amplitude and extreme values ​​of the loading cycle; the stress ratio R, defined as σ min / σ max, reflecting the symmetry of loading and stress cycle characteristics. Stress and strain relationship operations and vector mapping are performed on the geometric parameter set and the damage-sensitive parameter set. Through normalization processing, all parameters are standardized to a unified scale, and mean normalization or maximum and minimum value normalization are used to eliminate the influence of dimensional differences between different physical quantities. After normalization, the stress-strain characteristics of each parameter are calculated according to the physical relationship. For example, the evolution of the local stress field is deduced based on the principal stretching ratio and the change of the film thickness, the stiffness degradation trend of the membrane surface is calculated based on the curvature change and the strain distribution, and the fatigue crack growth driving force is calculated based on the crack propagation rate and the stress intensity factor amplitude. The above-mentioned processed parameters are vectorized and mapped in the set order to construct a standard eigenvector.

[0035] In a specific embodiment, the process of executing step 200 may specifically include the following steps: Calculate the attention weights of the standard feature vectors to obtain the initial weight distribution matrix; Calculate the geometric nonlinear factor based on the initial weight distribution matrix to obtain the geometric nonlinear factor; Adaptively adjust the weight of the damage sensitive parameter set according to the geometric nonlinear factor to obtain the weight distribution coefficient; Perform element-wise product operation on the standard eigenvector and the weight distribution coefficient to obtain the weighted eigenvector.

[0036] Specifically, attention weights are calculated for the standard eigenvector. A trainable attention network module is constructed, taking the standard eigenvector as input. By setting a set of weight matrices and bias vectors, the features are linearly transformed and nonlinear activation functions are introduced. Softmax normalization is then used to convert the weights of different feature dimensions into probability distributions. This process results in an initial weight assignment matrix that assigns different levels of attention to each component in the standard eigenvector. High weights indicate that the corresponding feature contributes significantly to fatigue life prediction, while low weights indicate that the feature has little influence under the current state. The geometric nonlinearity factor is calculated based on the initial weight assignment matrix. The geometric nonlinearity factor aims to quantify the large deformation characteristics of ETFE membrane structures during fatigue loading. In particular, under biaxial stress conditions, the membrane exhibits complex geometric nonlinear behaviors such as localized large strains, thickness reduction, and curvature changes under load. Specifically, the eigenvalues ​​related to the principal stretch ratios λ1 and λ2, the thickness change ratio t(N) / t0, and the curvature radius change ratio R(N) / R0 are extracted from the initial weight assignment matrix. A dimensionless comprehensive factor is constructed based on these physical quantities. A quantitative index characterizing the current geometric nonlinear state of the membrane structure is derived by taking into account the actual weights and physical meanings of each component in the form of a weighted average or normalized product. A larger geometric nonlinearity factor indicates that the current stress state of the membrane structure deviates more from the linear elastic assumption, resulting in more severe local geometric distortion and faster fatigue damage evolution. Based on the geometric nonlinearity factor, a set of damage-sensitive parameters is adaptively weighted to obtain a weight distribution coefficient. The set of damage-sensitive parameters includes stress intensity factor amplitude, crack growth rate, maximum stress, minimum stress, and stress ratio. To dynamically adjust the importance of each damage parameter based on the real-time geometric state of the membrane structure, a weight control function based on adaptive adjustment of the nonlinear factor is introduced. When the geometric nonlinearity factor is low, the membrane is in the small deformation linear elastic range, and more attention is paid to stress levels and cyclic loading parameters, thereby assigning higher weights to parameters such as maximum stress and stress ratio. When the geometric nonlinearity factor is high, the membrane structure has entered the large deformation stage, and more attention is paid to parameters such as crack growth rate and stress intensity factor amplitude that directly reflect the damage accumulation rate. The weight control function is set as a nonlinear mapping, such as a Sigmoid or Tanh function, to ensure smooth transitions and physical rationality in weight adjustment. The resulting weight allocation coefficients, after adaptive adjustment, can reflect the changing importance of different damage-sensitive parameters at different fatigue stages in real time, enhancing the expressive power and predictive performance of the eigenvector at each fatigue evolution stage. The weighted eigenvector is generated by performing an element-by-element product operation on the standard eigenvector and the weight allocation coefficients.

[0037] In a specific embodiment, the step of executing the geometric nonlinear factor calculation based on the initial weight distribution matrix, and the process of obtaining the geometric nonlinear factor may specifically include the following steps: extracting a first principal stretch ratio parameter and a second principal stretch ratio parameter from the initial weight distribution matrix; The thickness change ratio and the curvature radius change ratio are calculated based on the geometric parameter set to obtain the thickness change ratio and the curvature radius change ratio of the membrane structure; Based on the first principal stretching ratio parameter, the second principal stretching ratio parameter, the membrane structure thickness change ratio and the membrane structure curvature radius change ratio, the geometric nonlinear calculation is performed to obtain the initial nonlinear factor of the ETFE membrane in the large deformation state; The geometric nonlinear factor is obtained by correcting the initial nonlinear factor with the correlation of membrane structure deformation history according to the number of load cycles.

[0038] Specifically, the first and second principal stretch ratio parameters, most relevant to the membrane structure's deformation state, are extracted from the initial weight distribution matrix. Each dimension in the initial weight distribution matrix corresponds to a physical parameter component in the standard eigenvector, with the first and second principal stretch ratios λ1 and λ2 representing the stretch magnifications in the two principal strain directions of the membrane material under biaxial stress, respectively. By definition, the principal stretch ratios can be derived from the principal strains ε1 and ε2 via the relationships λ1=1+ε1 and λ2=1+ε2. During extraction, the selected parameters are ensured to have high corresponding weights in the weight matrix, indicating that the network prioritizes these two parameters during the current fatigue phase. This ensures that the subsequent geometric nonlinear factor calculation process focuses on the key control variables that truly reflect the large deformation characteristics of the membrane material. After extracting the principal stretch ratio parameters, the thickness change ratio and the curvature radius change ratio are calculated based on the geometric parameter set. The thickness change ratio t(N) / t0, defined as the ratio of the current membrane thickness t(N) after N fatigue cycles to the initial thickness t0, reflects the local plastic flow, thinning, and microdamage accumulation effects of the material during fatigue. The curvature radius change ratio, R(N) / R0, is obtained by measuring the curvature change of the membrane surface under load. The ratio of the current curvature radius R(N) to the initial curvature radius R0 effectively characterizes the overall geometric distortion trend of the membrane. A geometric nonlinearity calculation is performed based on the first and second principal stretch ratio parameters, the membrane structure thickness change ratio, and the membrane structure curvature radius change ratio to obtain an initial nonlinear factor that characterizes the large deformation state of the ETFE membrane. The specific calculation comprehensively considers the contribution of each parameter to the geometric nonlinear behavior and constructs a dimensionless nonlinear factor expression. This function uses a weighted product or exponential mapping. For example, a normalized weighted product is used to normalize the principal stretch ratio, thickness change ratio, and curvature radius change ratio to a uniform scale. A nonlinear combination is then performed to highlight the coupled influence of these physical quantities on the geometric nonlinear behavior during fatigue evolution. The larger the principal stretch ratio, the smaller the thickness change ratio, the more significant the curvature radius change, and the larger the overall nonlinear factor value, indicating that the membrane structure is more inclined towards the limit state of large deformation, local buckling, or material degradation. The initial nonlinear factor calculated in this way can characterize the geometric complexity and nonlinearity of the membrane material at the current fatigue stage. The initial nonlinear factor is corrected for the deformation history of the membrane structure in combination with the number of load cycles N. The fatigue process has obvious historical dependence. The accumulation of small deformations in the early stage will not immediately cause obvious geometric nonlinear changes. However, as the number of cycles increases, microscopic damage gradually accumulates, resulting in increased macroscopic geometric distortion. Therefore, the correction process introduces a historical memory effect. Specifically, a weight function related to the number of cycles is set, such as an exponential decay or power-law growth correction function, to ensure that the historical effect gradually strengthens as N increases. The historical weight function is multiplied by the initial nonlinear factor to obtain the corrected geometric nonlinear factor.

[0039] In a specific embodiment, the process of executing step 300 may specifically include the following steps: The weighted eigenvectors are input into the instantaneous damage increment branch, the cumulative damage trend branch and the critical damage threshold approximation branch respectively; The instantaneous damage increment branch performs convolution and activation processing on the weighted feature vector to obtain the instantaneous damage increment feature. The cumulative damage trend branch performs state update processing on the weighted feature vector to obtain the cumulative damage trend feature. The critical damage threshold approximation branch performs residual connection calculation on the weighted feature vector to obtain the critical damage threshold approximation feature. The instantaneous damage increment feature, cumulative damage trend feature and critical damage threshold approximation feature are fused on a time scale to obtain the fused damage feature.

[0040] Specifically, the weighted eigenvectors are input into the instantaneous damage increment branch, the cumulative damage trend branch, and the critical damage threshold approximation branch, respectively. These three branches are designed to address the different timescales of changes during the fatigue process, enabling multi-dimensional capture of fatigue damage behavior from three perspectives: instantaneous, long-term cumulative, and critical failure. In the instantaneous damage increment branch, a convolutional neural network is used to process the weighted eigenvectors. A one-dimensional convolutional layer is configured, and the convolution kernel length is designed based on the typical stress-strain variation period within a fatigue cycle, ensuring effective capture of local damage increment characteristics within a single or a small number of cycles. The convolution operation extracts high-order characteristic variation patterns within the local neighborhood and identifies small-scale damage signals at the initial stages of fatigue crack initiation. Nonlinear activation functions, such as ReLU or LeakyReLU, are used to enhance the nonlinear representation of the eigenvector, prevent gradient vanishing, and strengthen the network's sensitivity to abnormal local damage increments. After convolution and activation processing, the resulting instantaneous damage increment features accurately characterize the dynamic changes in subtle damage accumulation in membrane materials during specific fatigue cycles, reflecting the short-term evolution of the material under single-cycle loading. Simultaneously, the weighted feature vector is input into the cumulative damage trend branch, which utilizes a long short-term memory (LSTM) architecture for processing. As a typical recurrent neural network variant, LSTM possesses excellent time series learning capabilities and is suitable for modeling dependencies in long time series. Through a gating mechanism, LSTM effectively filters and retains important historical information, suppresses interference from irrelevant redundant data, and ensures network stability and accuracy in modeling long-term fatigue evolution trends. During processing, each LSTM unit sequentially receives the weighted feature vector, updates its internal state, and outputs a cumulative damage trend signature. This signature reflects the evolution of damage variables in the ETFE film during long-term loading, revealing the complete process from microcrack initiation to crack propagation and macroscopic failure. The weighted feature vector is then input into the critical damage threshold approximation branch. This branch utilizes a residual connection architecture, combining the powerful representation capabilities of deep neural networks with a residual learning strategy, to effectively model the rapid evolution of fatigue damage near the critical failure point. In implementation, skip connections are implemented to directly superimpose the input weighted feature vector with the deep network output, maintaining continuity in feature transfer and gradient flow stability, thus avoiding degradation issues during deep network training. Residual connections help the model more quickly learn the characteristic patterns of critical turning points in the fatigue evolution process, particularly the sudden change in behavior when the material transitions from the stable crack growth stage to the rapid failure stage. The critical damage threshold approximation feature output by this branch effectively captures the accelerated damage accumulation signal of the membrane material before fracture. The instantaneous damage increment feature, cumulative damage trend feature, and critical damage threshold approximation feature are fused over time scales. The fusion operation uses a weighted feature fusion or attention-based weighted fusion strategy, dynamically adjusting the fusion coefficient based on the importance of features at different time scales, or learning the optimal fusion weights through end-to-end training.The fusion layer, designed as a fully connected layer or self-attention module, adaptively adjusts the contribution ratio of features at each scale to achieve the coordinated integration of damage features at different time scales. The resulting fused damage features have excellent temporal representation capabilities and the ability to characterize the entire fatigue evolution cycle.

[0041] In a specific embodiment, the process of executing step 400 may specifically include the following steps: The fused damage features are input into the fatigue damage evolution physical constraint embedder, which includes the Chaboche damage evolution constraint module, the modified Paris formula constraint module and the viscoelastic constitutive relation constraint module. The first constraint feature is obtained by calculating the continuous damage mechanics equation for the fusion damage feature through the Chaboche damage evolution constraint module; The second constraint feature is obtained by performing crack propagation calculation of the biaxial stress state on the fusion damage feature through the modified Paris formula constraint module; The viscoelastic constitutive relationship constraint module is used to calculate the fusion damage characteristics using the Maxwell model to obtain the third constraint characteristics; The first constraint feature, the second constraint feature and the third constraint feature are subjected to constraint violation test and hard constraint fusion to obtain the physical constraint damage feature.

[0042] Specifically, the fused damage signature is fed into three modules within the physical constraint embedder: the Chaboche damage evolution constraint module, the modified Paris formula constraint module, and the viscoelastic constitutive relation constraint module. Each module physically models and constrains the fatigue evolution process from different perspectives. In the Chaboche damage evolution constraint module, the fused damage signature is used to solve the continuum damage mechanics equation. Based on the damage evolution model proposed by Chaboche, this module derives a theoretical damage evolution rate based on the current equivalent stress level, accumulated damage value, and material parameters such as the damage evolution coefficient, stress exponent, and damage exponent. By inputting the equivalent stress extracted from the fused signature and the current damage into the Chaboche equation, the damage increment per fatigue cycle step is calculated to obtain the first constraint signature. This signature reflects the damage accumulation rate and trend of the ETFE film material during fatigue loading, accurately characterizing the entire process from initial micro-damage evolution to later damage acceleration before failure. The fused damage signature is then fed into the modified Paris formula constraint module, which models the fatigue crack growth process based on the fundamental principles of fracture mechanics. Based on the traditional Paris equation, a principal stress ratio correction function is introduced to consider the crack growth characteristics of ETFE membrane materials under biaxial stress. The stress components in both principal stress directions are combined into an equivalent stress intensity factor amplitude. This modified crack growth equation more realistically describes the crack tip growth behavior of membrane materials under complex loading conditions. By inputting parameters such as the stress intensity factor amplitude, maximum and minimum stresses, and stress ratio from the fused feature into the modified Paris model, the increment of fatigue crack growth per cycle is calculated, resulting in a second constraint feature. This feature quantifies the crack growth rate of the membrane material and reflects the dominant fracture mechanism during fatigue evolution. Simultaneously, the fused damage feature is input into the viscoelastic constitutive relation constraint module for processing. This module, based on the Maxwell viscoelastic model, establishes a dynamic relationship between stress and strain rate, reflecting the viscoelastic response of the membrane material during fatigue. By introducing a relaxation time parameter and elastic modulus, the current stress state is linked to the strain rate, simulating the energy dissipation and internal damping behavior of the material during loading and unloading cycles. In specific implementation, the strain rate and stress data from the fused features are fed into the Maxwell model to solve for the viscoelastic stress response, thereby generating the third constraint feature. This feature captures the energy loss characteristics of membrane materials caused by internal molecular chain motion under high-cycle fatigue conditions. Constraint violation tests are performed on the first, second, and third constraint features to ensure that the features output by each physical constraint module strictly conform to their corresponding physical laws.By calculating the residuals between the predicted values ​​of the fusion features and the theoretical values ​​of the physical equations, such as the Chaboche damage evolution residual, the modified Paris crack extension residual, and the Maxwell viscoelastic response residual, the degree of physical violation of the model under the current parameter state is determined. When the residual exceeds the preset threshold, it indicates that the model has a physical violation, which is corrected by introducing a hard constraint mechanism. The hard constraint mechanism minimizes the physical residuals by forcibly adjusting the internal parameters of the network, ensuring that the output features strictly follow the laws of physics while satisfying data fitting. The constraint fusion operation uses a weighted summation method to weight the various constraint residuals into a total physical consistency loss function. The parameters are updated through backpropagation, so that the final output physical constraint damage feature achieves an optimal balance between data-driven and physical constraints.

[0043] In a specific embodiment, the execution step of calculating the continuous damage mechanics equation for the fusion damage feature using the Chaboche damage evolution constraint module to obtain the first constraint feature may specifically include the following steps: Extract equivalent stress value and current damage value from fusion damage features; The Chaboche damage evolution constraint module calculates the theoretical damage evolution rate based on the Chaboche damage evolution equation based on the equivalent stress value and the preset ETFE membrane material parameters, where the preset ETFE membrane material parameters include the damage evolution coefficient, stress exponent, and damage exponent. The theoretical damage evolution rate and the current damage value are input into the physical constraint layer for residual calculation to obtain the damage evolution constraint residual value; The fused damage feature is corrected by the continuous damage mechanics law according to the residual value of the damage evolution constraint to obtain the first constraint feature.

[0044] Specifically, the equivalent stress value and current damage value are extracted from the fused damage signature. The equivalent stress value simplifies the stress information of the membrane structure under biaxial loading into a single scalar quantity. It comprehensively reflects the local stress field intensity and plays a dominant role in the fatigue accumulation process. The current damage value, on the other hand, characterizes the degree of damage evolution within the material. Its value ranges from zero to one, representing the two extreme states of intactness and complete failure, respectively. The equivalent stress and current damage values ​​are input into the Chaboche damage evolution constraint module. Based on the theory of continuum damage mechanics, this module uses preset ETFE membrane material parameters to model the damage evolution rate. These preset parameters include the damage evolution coefficient, which controls the overall rate of damage evolution; the stress exponent, which describes the sensitivity of the damage evolution process to changes in stress levels; and the damage exponent, which controls the influence of the damage variable on its own growth rate. These parameters are pre-calibrated through standard fatigue experiments and reflect the fatigue response characteristics of the ETFE membrane material under specific stress and cyclic conditions. Based on these parameters, the module derives a theoretical damage evolution rate based on the currently extracted equivalent stress and damage values, describing the damage accumulation amplitude per cycle under the current loading and damage conditions. To assess whether the fused features adhere to physical evolution laws, the theoretical damage evolution rate and the current damage value in the actual features are input into the physical constraint layer, and residuals are calculated. The residual reflects the deviation between the theoretical prediction and the actual model output. A larger value indicates a greater deviation from physical laws, while a smaller value indicates a greater physical plausibility of the feature prediction. The difference between the current damage evolution rate and the theoretical rate is calculated to determine the model's physical consistency during fatigue evolution. Based on the residual value of the damage evolution constraint, the fused damage features are corrected using the continuous damage mechanics law. By introducing the physical residual as one of the objectives of neural network optimization and combining it with the traditional data fitting loss, a comprehensive loss function is constructed. During the optimization process, a backpropagation algorithm is used to dynamically adjust network parameters based on the residual, gradually narrowing the gap between theoretical and predicted values. To avoid a trade-off between physical constraints and data fitting, a physical consistency weight coefficient is introduced to balance the contributions of the two during training, ensuring adherence to physical laws while maintaining high prediction accuracy. As the training iterations progress, the damage evolution residual gradually decreases, and the fusion features are continuously optimized towards physical consistency, ultimately obtaining the first constraint feature that conforms to the law of continuous damage mechanics.

[0045] In this embodiment, a modified Paris formula constraint module is used to perform crack propagation calculation in a biaxial stress state on the fusion damage feature to obtain a second constraint feature, including: extracting the stress intensity factor amplitude, crack propagation rate, maximum stress and minimum stress from the fusion damage feature to obtain a crack propagation calculation parameter group; performing a biaxial stress state equivalent stress intensity factor calculation based on the crack propagation calculation parameter group to obtain an equivalent stress intensity factor amplitude considering the interaction of principal stress components; inputting the equivalent stress intensity factor amplitude into the modified Paris formula to perform ETFE membrane material constant correction and crack propagation rate iterative calculation to obtain a modified crack propagation rate under the biaxial stress state; performing cumulative damage integral and Paris formula physical constraint verification on the modified crack propagation rate according to the number of load cycles to obtain the second constraint feature.

[0046] In this embodiment, the fusion damage feature is calculated by Maxwell model through the viscoelastic constitutive relationship constraint module to obtain the third constraint feature, including: extracting elastic modulus parameters, viscous modulus parameters and relaxation time parameters from the fusion damage feature to obtain the Maxwell model constitutive parameter combination; performing elastic component calculation and viscous component calculation based on the Maxwell model constitutive parameter combination to obtain the transient elastic response component and the time-dependent viscous response component; inputting the transient elastic response component and the time-dependent viscous response component into the exponential relaxation function to perform time domain convolution operation to obtain the viscoelastic stress response result of the ETFE membrane material; performing constitutive relationship constraint test and residual calculation on the viscoelastic stress response result according to the number of load cycles to obtain the third constraint feature.

[0047] In a specific embodiment, the process of executing step 500 may specifically include the following steps: Inputting the physical constraint damage feature into the Bayesian neural network to calculate the probability distribution parameters, obtaining a first weight probability distribution parameter and a first probability distribution parameter, wherein the first weight probability distribution parameter includes a weight mean and a weight variance, and the first probability distribution parameter includes a bias mean and a bias variance; performing a variational inference calculation based on the first weighted probability distribution parameter and the first probability distribution parameter to obtain a mean and a variance of the fatigue life prediction; Calculate the confidence interval based on the mean and variance of fatigue life prediction to obtain the upper and lower limits of the confidence interval for fatigue life corresponding to a 95% confidence level. The fatigue life prediction value of ETFE membrane structure is obtained by integrating the mean, variance, upper limit of confidence interval and lower limit of confidence interval of fatigue life prediction into probabilistic life.

[0048] Specifically, the physical constraint damage characteristics are input into a Bayesian neural network. Unlike traditional neural networks, Bayesian neural networks no longer treat weights and biases as fixed, deterministic parameters. Instead, they are modeled as probability distributions with uncertainty, typically assumed to be Gaussian distributions. The parameters of these distributions are learned. During the model initialization phase, the Bayesian neural network parameterizes the weights and biases of each layer. The first weight probability distribution parameters include the weight mean and weight variance, which describe the distribution of weights at each layer; while the first probability distribution parameters include the bias mean and bias variance, which describe the uncertainty characteristics of the bias terms at each layer. In this way, during training, the network not only captures the inherent patterns of the data features but also simultaneously quantifies the prediction uncertainty caused by data incompleteness, environmental perturbations, or model complexity. Based on the first weight probability distribution parameters and the first probability distribution parameters, variational inference calculations are performed. The core concept of variational inference is to approximate the difficult-to-calculate posterior distribution using a set of controllable parameterized distributions. Through optimization methods, the approximated distribution is continuously adjusted to match the true posterior distribution as closely as possible. In the specific implementation, a variational distribution is defined to approximate the true parameter posterior distribution. The difference between this variational distribution and the true distribution is minimized, using the distance between the two distributions as a metric. This approach effectively circumvents the difficulties of directly computing high-dimensional integrals and quickly obtains a good estimate of parameter uncertainty. As training progresses, the mean and variance of the network weights and biases gradually converge, ultimately yielding the mean and variance of the fatigue life prediction. The life prediction mean represents the most likely fatigue life value predicted by the network under the current physical characteristics, while the variance quantifies the uncertainty of the prediction, reflecting the model's confidence in the life prediction. After obtaining the fatigue life prediction mean and variance, a confidence interval is calculated based on statistical inference. To ensure the operability and reliability of the prediction results in engineering applications, a clear confidence interval is provided to reflect the uncertainty range of the fatigue life prediction. Under standard statistical settings, a 95% confidence level is selected. Based on the properties of the normal distribution, the upper and lower limits of the confidence interval are calculated by adding or subtracting multiples of the standard deviation from the fatigue life prediction mean. The lower limit of the confidence interval indicates that at a 95% confidence level, the fatigue life will not be lower than this value, while the upper limit of the confidence interval indicates that at the same confidence level, the life will not be higher than this value. The mean and variance of the fatigue life prediction and the corresponding upper and lower limits of the confidence interval are integrated to form a complete probabilistic life prediction result.

[0049] In a specific embodiment, the step of performing variational inference calculation based on the first weighted probability distribution parameter and the first probability distribution parameter to obtain the mean and variance of fatigue life prediction may specifically include the following steps: Constructing a variational posterior distribution based on the first weight probability distribution parameter and the first probability distribution parameter to obtain a weighted variational posterior distribution and a biased variational posterior distribution; The weighted variational posterior distribution and the biased variational posterior distribution are input into the evidence lower bound loss function for ELBO calculation to obtain the evidence lower bound loss value including the data fitting term and the KL divergence regularization term; Perform gradient descent optimization according to the lower bound loss value of the evidence to obtain the second weight probability distribution parameter and the second probability distribution parameter; The fatigue life probability distribution is calculated based on the second weight probability distribution parameter and the second probability distribution parameter to obtain the mean and variance of the fatigue life prediction.

[0050] Specifically, a variational posterior distribution is constructed based on the first weight probability distribution parameters and the first probability distribution parameters. The first weight probability distribution parameters include the mean and variance of each network connection weight, and the first probability distribution parameters include the mean and variance of each layer bias. These parameters are not fixed but are used to describe the distribution characteristics of each neuron's connection strength and bias under uncertainty. To achieve approximate inference of the posterior distribution, these means and variances are further organized into a standard variational distribution, using a diagonal Gaussian distribution to simplify calculations and assuming independence of different parameters. Each weight and bias has a probability distribution defined by its mean and variance, forming the overall weight variational posterior distribution and bias variational posterior distribution. After constructing the variational posterior distribution, the weight variational posterior distribution and bias variational posterior distribution are input into the evidence lower bound (ELBO) loss function for calculation. The ELBO loss function consists of two components: a data fitting term, which measures the model's ability to explain the training data under the current parameter distribution and is calculated based on the expected log-likelihood. The other component is a regularization term, based on the distance between the variational posterior distribution and the prior distribution, quantified by the Kullback-Leibler divergence. This term aims to prevent the model from overfitting due to excessive reliance on limited training data. This design maximizes data fitting while controlling model complexity, ensuring that the network maintains appropriate generalization capabilities. The variational posterior distribution is used as input, and the data fitting term and the KL divergence regularization term are calculated separately to obtain a combined ELBO loss value. Based on the calculated evidence lower bound loss value, the network parameters are updated using a gradient descent optimization algorithm. The goal of gradient descent optimization is to minimize the negative ELBO value, that is, to maximize the evidence lower bound, thereby ensuring data fitting accuracy while maintaining reasonable parameter uncertainty. Gradients are calculated for the weight mean, weight variance, bias mean, and bias variance, guiding parameter adjustments towards optimal results. To ensure the stability and convergence of the training process, an optimizer based on adaptive learning rate adjustment, such as the Adam optimizer, is used to automatically adjust the update step size of different parameters to avoid falling into local optimality or oscillatory non-convergence in complex parameter spaces. As training iterations proceed, the network's internal weight variational distribution and bias variational distribution gradually adjust and converge, resulting in the second-stage second-weight probability distribution parameters and second-stage second-weight probability distribution parameters. Based on these updated second-weight probability distribution parameters and second-stage second-probability distribution parameters, the fatigue life probability distribution is calculated. Multiple parameter sampling within the new variational distribution is combined with forward inference based on physical constraint damage characteristics to generate multiple fatigue life prediction samples. These samples reflect the predictive distribution characteristics of the model under the current physical state. Statistical analysis of these samples is used to calculate the mean and variance of the fatigue life prediction.The mean reflects the model's optimal prediction result for fatigue life, and the variance quantifies the model's confidence in the result in the current prediction task, that is, the degree of uncertainty.

[0051] The above describes the ETFE membrane structure life prediction method based on machine learning in the embodiment of the present invention. The following describes the ETFE membrane structure life prediction system based on machine learning in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, an ETFE membrane structure life prediction system based on machine learning includes: A preprocessing module 11 is used to perform standardization preprocessing on the cyclic load fatigue data of the ETFE membrane structure under the biaxial stress state to obtain a standard eigenvector; The weight distribution module 12 is used to distribute the geometric weight of the membrane structure to the standard eigenvector to obtain a weighted eigenvector; A feature extraction module 13 is used to extract multi-time scale damage features based on weighted feature vectors to obtain fused damage features; The constraint calculation module 14 is used to input the fused damage characteristics into the fatigue damage evolution physical constraint embedder to perform constraint calculation and obtain the physical constraint damage characteristics; The life prediction module 15 is used to perform Bayesian probability life prediction on the physical constraint damage characteristics to obtain the fatigue life prediction value of the ETFE membrane structure.

[0052] Through the collaborative efforts of these components and the establishment of a data preprocessing method specifically tailored to the biaxial stress state of ETFE membrane structures, a standard feature vector consisting of principal stress components, principal strain components, a set of geometric parameters, and a set of damage-sensitive parameters was constructed. Compared to traditional uniaxial fatigue data processing methods, this method can more comprehensively reflect the complex stress state and large-deformation geometric nonlinear behavior of membrane structures. By introducing a membrane structure geometric nonlinear adaptive weight module, the weight distribution of damage-sensitive parameters is dynamically adjusted according to the real-time deformation state of the ETFE membrane through an attention mechanism. Compared to the fixed weight approach of traditional neural networks, this method can adaptively identify key damage parameters at different deformation stages, significantly enhancing the accuracy of fatigue behavior characterization of membrane structures. A multi-branch parallel architecture is employed to simultaneously extract instantaneous damage increment features, cumulative damage trend features, and critical damage threshold approximation features. Compared to existing single-timescale damage analysis methods, this method can comprehensively capture the characteristics of the different stages of ETFE membrane fatigue, including microcrack initiation, macrocrack propagation, and overall structural failure, achieving a refined description of the fatigue damage evolution process. By embedding the Chaboche damage evolution equation, the modified Paris formula, and the viscoelastic constitutive relation as physical laws into a neural network through a hard constraint mechanism, this approach ensures that the prediction results strictly adhere to the laws of continuous damage mechanics, compared to the existing approach of treating physical laws as soft constraints. This addresses the key issue of traditional data-driven prediction methods lacking physical plausibility. A Bayesian neural network is used to quantify the uncertainty of fatigue life prediction. A variational inference method is used to output the mean, variance, and confidence interval of fatigue life. Compared to traditional deterministic prediction methods that only provide point estimates, this approach can provide a basis for risk assessment and predictive reliability evaluation for engineering applications.

[0053] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0054] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0055] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the life of an ETFE membrane structure based on machine learning, characterized in that: include: The cyclic load fatigue data of ETFE membrane structure under biaxial stress state is standardized and preprocessed to obtain the standard eigenvector; Performing membrane structure geometry weight distribution on the standard eigenvector to obtain a weighted eigenvector; Extracting multi-time-scale damage features based on the weighted feature vector to obtain a fused damage feature; Inputting the fused damage characteristics into a fatigue damage evolution physical constraint embedder for constraint calculation to obtain physical constraint damage characteristics; A Bayesian probability life prediction is performed on the physical constraint damage characteristics to obtain a fatigue life prediction value of the ETFE membrane structure.

2. The ETFE membrane structure life prediction method based on machine learning according to claim 1, wherein The cyclic load fatigue data of the ETFE membrane structure under the biaxial stress state is subjected to standardized preprocessing to obtain a standard eigenvector, including: The cyclic loading fatigue data of the ETFE membrane structure under biaxial stress state was collected to obtain the original fatigue data including the principal stress component, principal strain component and number of loading cycles under biaxial stress state; Based on the original fatigue data, geometric parameters and damage-sensitive parameters are constructed to obtain a geometric parameter set including a membrane structure thickness variation parameter, a curvature radius variation parameter, and a principal stretch ratio parameter, and a damage-sensitive parameter set including a stress intensity factor amplitude, a crack growth rate, a maximum stress, a minimum stress, and a stress ratio; Stress and strain relationship operations and vector mapping are performed on the geometric parameter set and the damage sensitive parameter set to obtain a standard eigenvector.

3. The ETFE membrane structure life prediction method based on machine learning according to claim 2, wherein The step of assigning a membrane structure geometric weight to the standard eigenvector to obtain a weighted eigenvector includes: Calculate the attention weight of the standard feature vector to obtain an initial weight distribution matrix; Calculating a geometric nonlinear factor based on the initial weight distribution matrix to obtain a geometric nonlinear factor; Adaptively adjust the weight of the damage sensitive parameter set according to the geometric nonlinear factor to obtain a weight distribution coefficient; Performing an element-by-element product operation on the standard feature vector and the weight distribution coefficient to obtain a weighted feature vector.

4. The ETFE membrane structure life prediction method based on machine learning according to claim 3, wherein The calculating of the geometric nonlinear factor based on the initial weight distribution matrix to obtain the geometric nonlinear factor includes: extracting a first principal stretch ratio parameter and a second principal stretch ratio parameter from the initial weight distribution matrix; Calculating the thickness change ratio and the curvature radius change ratio based on the geometric parameter set to obtain the membrane structure thickness change ratio and the membrane structure curvature radius change ratio; Performing geometric nonlinear calculation based on the first principal stretching ratio parameter, the second principal stretching ratio parameter, the membrane structure thickness change ratio, and the membrane structure curvature radius change ratio to obtain an initial nonlinear factor of the ETFE membrane in a large deformation state; The initial nonlinear factor is corrected for its membrane structure deformation history correlation according to the number of load cycles to obtain a geometric nonlinear factor.

5. The ETFE membrane structure life prediction method based on machine learning according to claim 1, wherein The extracting of multi-time-scale damage features based on the weighted feature vector to obtain a fused damage feature includes: Inputting the weighted eigenvectors into the instantaneous damage increment branch, the cumulative damage trend branch and the critical damage threshold approximation branch respectively; The instantaneous damage increment branch performs convolution operation and activation processing on the weighted feature vector to obtain an instantaneous damage increment feature; the cumulative damage trend branch performs state update processing on the weighted feature vector to obtain a cumulative damage trend feature; the critical damage threshold approximation branch performs residual connection calculation on the weighted feature vector to obtain a critical damage threshold approximation feature; The instantaneous damage increment feature, the cumulative damage trend feature and the critical damage threshold approximation feature are fused in time scale to obtain a fused damage feature.

6. The ETFE membrane structure life prediction method based on machine learning according to claim 1, wherein The step of inputting the fused damage feature into a fatigue damage evolution physical constraint embedder for constraint calculation to obtain the physical constraint damage feature includes: Inputting the fused damage features into a fatigue damage evolution physical constraint embedder respectively, wherein the fatigue damage evolution physical constraint embedder includes a Chaboche damage evolution constraint module, a modified Paris formula constraint module, and a viscoelastic constitutive relation constraint module; Performing a continuous damage mechanics equation calculation on the fusion damage feature through the Chaboche damage evolution constraint module to obtain a first constraint feature; Performing a biaxial stress state crack propagation calculation on the fusion damage feature using the modified Paris formula constraint module to obtain a second constraint feature; Performing a Maxwell model calculation on the fusion damage feature through the viscoelastic constitutive relationship constraint module to obtain a third constraint feature; Constraint violation inspection and hard constraint fusion are performed on the first constraint feature, the second constraint feature, and the third constraint feature to obtain a physical constraint damage feature.

7. The ETFE membrane structure life prediction method based on machine learning according to claim 6, wherein: The Chaboche damage evolution constraint module is used to calculate the fusion damage feature through a continuous damage mechanics equation to obtain a first constraint feature, including: extracting an equivalent stress value and a current damage value from the fusion damage feature; The Chaboche damage evolution constraint module calculates the Chaboche damage evolution equation based on the equivalent stress value and preset ETFE membrane material parameters to obtain a theoretical damage evolution rate, wherein the preset ETFE membrane material parameters include a damage evolution coefficient, a stress exponent, and a damage exponent; Inputting the theoretical damage evolution rate and the current damage value into the physical constraint layer to perform residual calculation to obtain a damage evolution constraint residual value; The fusion damage feature is corrected by continuous damage mechanics law according to the damage evolution constraint residual value to obtain a first constraint feature.

8. The ETFE membrane structure life prediction method based on machine learning according to claim 1, wherein The Bayesian probabilistic life prediction of the physical constraint damage characteristics to obtain a fatigue life prediction value of the ETFE membrane structure includes: Inputting the physical constraint damage feature into a Bayesian neural network to calculate probability distribution parameters to obtain first weight probability distribution parameters and first probability distribution parameters, wherein the first weight probability distribution parameters include a weight mean and a weight variance, and the first probability distribution parameters include a bias mean and a bias variance; performing a variational inference calculation based on the first weighted probability distribution parameter and the first probability distribution parameter to obtain a mean and a variance of fatigue life prediction; Calculating the confidence interval based on the mean and variance of the fatigue life prediction to obtain an upper limit and a lower limit of the confidence interval for the fatigue life corresponding to a 95% confidence level; The fatigue life prediction value of the ETFE membrane structure is obtained by performing probabilistic life integration on the mean value, the variance, the upper limit value of the confidence interval and the lower limit value of the confidence interval of the fatigue life prediction.

9. The ETFE membrane structure life prediction method based on machine learning according to claim 8, wherein: The performing variational inference calculation based on the first weight probability distribution parameter and the first probability distribution parameter to obtain the mean and variance of fatigue life prediction includes: Constructing a variational posterior distribution based on the first weight probability distribution parameter and the first probability distribution parameter to obtain a weighted variational posterior distribution and a biased variational posterior distribution; Inputting the weighted variational posterior distribution and the biased variational posterior distribution into the evidence lower bound loss function to perform ELBO calculation to obtain an evidence lower bound loss value including a data fitting term and a KL divergence regularization term; Performing gradient descent optimization according to the lower bound loss value of the evidence to obtain a second weight probability distribution parameter and a second probability distribution parameter; Fatigue life probability distribution calculation is performed based on the second weight probability distribution parameter and the second probability distribution parameter to obtain a mean and variance of fatigue life prediction.

10. A machine learning-based ETFE membrane structure life prediction system, characterized in that: Used to execute the ETFE membrane structure life prediction method based on machine learning according to any one of claims 1 to 9, the ETFE membrane structure life prediction system based on machine learning comprises: A preprocessing module is used to perform standardized preprocessing on the cyclic load fatigue data of the ETFE membrane structure under biaxial stress state to obtain a standard eigenvector; A weight distribution module is used to distribute the geometric weight of the membrane structure to the standard eigenvector to obtain a weighted eigenvector; A feature extraction module, configured to extract multi-time-scale damage features based on the weighted feature vector to obtain a fused damage feature; a constraint calculation module, configured to input the fused damage feature into a fatigue damage evolution physical constraint embedder for constraint calculation to obtain a physical constraint damage feature; The life prediction module is used to perform Bayesian probability life prediction on the physical constraint damage characteristics to obtain a fatigue life prediction value of the ETFE membrane structure.

Images