A neural network-based concrete main tower anti-cracking risk grading and classification method

Abstract

The present application relates to the technical field of bridge structure health monitoring, and particularly relates to a concrete main tower anti-cracking risk grading and classification method based on a neural network, which specifically comprises the following steps: obtaining parameters for evaluating the anti-cracking risk of the concrete main tower in an actual working condition; performing normalization processing on the parameters based on physical dimensions and abnormal robustness to obtain a normalized parameter vector; based on concrete fracture mechanics and creep theory, constructing a physical enhancement feature according to the parameters, and splicing the normalized parameter vector and the physical enhancement feature to form an initial feature vector; and inputting the initial feature vector into a pre-trained anti-cracking risk grading and classification model, and outputting an anti-cracking risk grade classification result after model processing. The present application can accurately and efficiently complete the anti-cracking risk grading and classification of the concrete main tower through a neural network guided by physical information.

Description

Technical Field

[0001] This invention relates to the field of bridge structural health monitoring technology, and in particular to a method for classifying and grading the crack resistance risk of concrete main towers based on neural networks. Background Technology

[0002] As modern bridge engineering develops towards longer spans, higher piers, and more complex structures, the concrete main tower, as the core load-bearing component of long-span bridges such as cable-stayed bridges and suspension bridges, directly affects the safety, durability, and service life of the bridge structure due to its crack resistance. In actual engineering projects, concrete main towers face complex stress-environment coupling effects during both the construction and operation phases. Their cracking risk is influenced by multiple factors, including geometric parameters, material performance parameters, construction process parameters, and environmental condition parameters. These parameters exhibit strong nonlinear coupling relationships, and their dimensions and physical meanings differ significantly. While traditional crack risk analysis methods based on finite element simulation can accurately calculate the stress field and crack width under specific working conditions, they suffer from high computational costs, long analysis cycles, and high requirements for the professional skills of engineers, making it difficult to meet the practical engineering needs for rapid assessment, real-time early warning, and batch working condition analysis. In recent years, with the rapid development of artificial intelligence technology, engineering risk prediction methods based on neural networks have gradually attracted attention. However, most existing technologies directly use finite element simulation results as training data, and simply normalize the original input parameters before feeding them into a fully connected network for training. This approach has obvious shortcomings.

[0003] Current methods for assessing the crack resistance risk of concrete structures largely rely on finite element simulations or empirical formulas. These methods suffer from limited data sources and narrow coverage of operating conditions, making it difficult to generate a complete training sample set encompassing extreme conditions and boundary parameters. This results in insufficient generalization ability of the models when facing high-risk scenarios such as high stress ratios, large dimensions, and complex temperature and humidity environments, hindering accurate identification of critical risk states. Traditional data preprocessing methods employ uniform standardization or normalization, ignoring the inherent physical meaning and dimensional differences of each parameter, easily disrupting key physical ratios such as stress-strength and temperature / humidity-creep. Furthermore, conventional methods lack effective robust handling mechanisms for outliers appearing in finite element simulations, which can easily cause gradient oscillations in the early stages of training, affecting model convergence stability. Existing neural network models directly stack original parameters in their input feature construction, failing to demonstrate... The model relies on a large number of samples to implicitly learn physical coupling relationships by encoding prior physical knowledge such as concrete fracture mechanics, creep theory, and size effect. During feature extraction, fully connected layers or single-scale convolutions struggle to simultaneously capture multi-scale information, including macroscopic geometry, mesoscopic material structure, and microscopic response, limiting the model's ability to express complex crack resistance mechanisms. Existing classification models rely solely on label-driven constraints such as cross-entropy in their loss function design, lacking explicit embedding of engineering physics laws. When the stress ratio significantly exceeds the cracking threshold, the model may still output a low-risk prediction, violating common sense in concrete fracture mechanics. Furthermore, the network initialization method does not consider the distribution pattern (skewness, kurtosis) of input features and prior risk categories, easily leading to response lag or gradient instability in the early training stages when there is class imbalance or non-Gaussian feature distribution.

[0004] Therefore, this invention proposes a neural network-based method for classifying and grading the crack resistance risk of concrete main towers to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a neural network-based method for classifying and grading the crack resistance risk of concrete main towers. This invention uses a neural network guided by physical information to accurately and efficiently classify and grade the crack resistance risk of concrete main towers.

[0006] The technical solution of this invention to solve the technical problem is a neural network-based method for classifying and grading the crack resistance risk of concrete main towers, comprising:

[0007] Obtain parameters used to assess the crack resistance risk of the concrete main tower under actual working conditions;

[0008] The parameters are normalized based on physical dimensions and abnormal robustness to obtain a normalized parameter vector; based on concrete fracture mechanics and creep theory, physical enhancement features are constructed according to the parameters, and the normalized parameter vector and physical enhancement features are concatenated to form an initial feature vector.

[0009] The initial feature vector is input into the pre-trained crack resistance risk classification model, and the crack resistance risk level classification result is output after the model is processed.

[0010] In actual working conditions, the parameters used to assess the crack resistance risk of concrete main towers are obtained by the finite element model. Based on the crack resistance risk assessment parameters of concrete main towers in actual engineering, a finite element model is constructed to generate a sample dataset covering multiple working conditions. The risk level of the samples in the sample dataset is labeled, and the labeled sample dataset is divided into a training set and a validation set.

[0011] The normalization and physical enhancement processes are as follows:

[0012] The collected evaluation parameters constitute sample data. Normalization is performed on each sample. First, a physical minimum and maximum value are set for each evaluation parameter in the sample. Based on these values, each parameter is truncated, and then linear normalization is applied to the truncated parameters. Next, for each evaluation parameter, the sample mean and standard deviation are calculated. The positional relationship of the original parameter values ​​relative to the mean and standard deviation is then input into the logistic function to obtain a nonlinear mapping result. Finally, the linear normalization result and the nonlinear mapping result of each evaluation parameter are weighted and fused to obtain a normalized parameter vector.

[0013] Based on concrete fracture mechanics, creep theory, and structural design principles, and using each evaluation parameter and normalized parameter as a basis, stress ratio characteristics, temperature and humidity coupled creep factor, size effect factor, load stiffness ratio characteristics, and creep time factor are constructed. The above five physical enhancement features are spliced ​​together in a fixed order to obtain the physical enhancement feature vector of each sample. Then, the normalized parameter vector is spliced ​​together with the physical enhancement feature vector to obtain the initial feature vector.

[0014] The crack resistance risk classification model is constructed based on a deep neural network with feature distribution and risk prior. The model includes: a multi-scale feature extraction module, a deep neural network parameter adaptive initialization module, a global-local feature recalibration module, a deep feature extraction module, and a classification output module.

[0015] (1) The multi-scale feature extraction module adopts a multi-branch one-dimensional dilated convolution and an adaptive weighted fusion mechanism to perform multi-scale feature extraction on the initial feature vector to obtain a multi-scale fused feature vector;

[0016] The operation of the multi-scale feature extraction module is as follows:

[0017] The initial feature vector of each sample is used as the input sequence of a one-dimensional convolutional layer, which is fed into multiple one-dimensional convolutional branches set in parallel. Each branch has the same number of convolutional kernels and output channels, but different dilation rates. The convolution result of each branch is then connected to a non-linear activation function. After passing through multiple one-dimensional convolutional branches, the feature extraction results with the same multiple dimensions are output. Then, a set of learnable scoring parameters is set for each branch, and the adaptive fusion weights of each branch are calculated to obtain the weight coefficient of each branch. Finally, the feature extraction results of each branch are weighted and summed according to the weight coefficients to output the multi-scale fusion feature vector of each sample.

[0018] (2) Adaptive initialization of deep neural network parameters: First, calculate the higher-order statistics of the multi-scale fusion feature vectors of all samples, including the global mean vector, global variance vector, global skewness vector, and global kurtosis vector. Then, introduce the higher-order statistics and risk priors into the first layer of the deep neural network for adaptive weight initialization. Finally, use the risk level prior probability to directly initialize the output layer bias of the deep neural network.

[0019] The specific steps for adaptive parameter initialization in deep neural networks are as follows:

[0020] The mean, variance, skewness, and kurtosis of the multi-scale fusion feature vector corresponding to each sample are calculated on the feature dimension to obtain a higher-order statistic composed of the global mean vector, global variance vector, global skewness vector, and global kurtosis vector.

[0021] The first fully connected layer of the deep neural network is constructed. The initial variance corresponding to the dimensions of the input multi-scale fusion feature vector is corrected according to the global skewness vector and global kurtosis vector in the higher-order statistics to obtain the corrected variance. Then, the prior probability of the risk level is calculated based on the true label of the sample. The importance of the feature is determined according to expert knowledge. The risk importance factor on each dimension is calculated according to the prior probability of the risk level and the importance of the feature. Finally, the sampling variance is set for the weights corresponding to each input dimension in the first fully connected layer by combining the initial variance, the corrected variance and the risk importance factor. The bias vector of the first fully connected layer is initialized to a vector of all zeros.

[0022] Construct the output layer bias vector of the deep neural network, with the dimension corresponding to the number of wind direction level classification categories. Add a smoothing term to the prior probability of each risk level category. Then, initialize the bias of the output layer corresponding to the category to a value that matches the prior probability. Finally, keep the output layer weights in the conventional initialization method.

[0023] (3) The global-local feature recalibration module extracts global statistical information based on the multi-scale fusion feature vector, global skewness vector and global kurtosis vector, and then recalibrates to obtain the recalibrated feature vector;

[0024] The global-local feature recalibration module operates as follows:

[0025] The input multi-scale fusion feature vector is concatenated with the global skewness vector and the global kurtosis vector in a fixed order to obtain the global statistical vector. The multi-scale fusion feature vector and the global statistical vector are then recalibrated. The two feature vectors are projected onto the same feature space through a linear mapping, then summed and fused, and a bias term is added to obtain the fusion result. The fusion result is then subjected to Sigmoid activation to obtain the attention weight vector. Finally, the multi-scale fusion feature vector and the attention weight vector are multiplied element-wise to obtain the recalibrated feature vector.

[0026] (4) The deep feature extraction module introduces a learnable damage activation function and performs deep feature extraction through a series of fully connected layers in the deep neural network after adaptive initialization.

[0027] The deep feature extraction module is as follows: a multi-layer deep feature extraction layer is constructed, and a damage activation function is introduced after each layer. The recalibrated feature vector is input into the first deep feature extraction layer and passes through the damage activation function to obtain the first hidden layer features. Then, the output of the first layer is used as the input of the next layer. After layer-by-layer operation, the deep features are obtained.

[0028] (5) Input the deep features into the classification output module. After adaptive initialization, the output layer of the deep neural network performs risk level classification prediction. Based on the classification cross-entropy loss, introduce the loss function of physical consistency constraint and calculate the total loss of the model.

[0029] The specific operation of the classification output module is as follows:

[0030] An output layer is constructed, and deep features are input into the output layer to obtain the original scores corresponding to each risk level category. Then, a Softmax normalization operation is performed on each original score to obtain the predicted probability vector. The theoretical high-risk probability is constructed based on the stress ratio feature in the constructed physical enhancement features. The high-risk category probabilities in the prediction results are merged to obtain the total high-risk probability predicted by the model. Physical consistency loss is introduced as a penalty term for the total high-risk probability being lower than the theoretical high-risk probability. The total loss of the model is calculated in combination with the classification cross-entropy loss.

[0031] The pre-trained crack resistance risk classification model uses a mini-batch stochastic gradient descent optimization algorithm to iteratively update the parameters of the deep neural network. The specific training process is as follows:

[0032] The dataset consisting of the acquired parameters is divided into a training set and a validation set. A fixed number of samples are randomly selected from the training set each time to form a batch. After normalization and physical feature enhancement, these samples are input into the crack resistance risk classification model. The model sequentially passes through a multi-scale feature extraction module, a deep neural network parameter adaptive initialization module, a global-local feature recalibration module, a deep feature extraction module, and a classification output module, ultimately obtaining the total loss value for that batch of samples. Based on the total loss value, the gradients of all trainable parameters in the model are calculated using the backpropagation algorithm, and the parameters are updated using an adaptive moment estimation optimizer. After each training iteration, the current model is evaluated using the validation set until the iteration stopping condition is met, resulting in a trained crack resistance risk classification model.

[0033] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0034] This invention discloses a neural network-based method for classifying and grading the crack resistance risk of concrete main towers. It constructs a high-quality training dataset based on finite element simulation and orthogonal experimental design, and systematically generates multi-condition samples covering structural form, material grade, construction conditions, and environmental conditions using the controlled variable method. Particular attention is paid to the collection of extreme and boundary conditions. Combined with standards and specifications, it achieves refined labeling of five risk levels, providing a data foundation covering the complete risk spectrum for model training. A physical dimension-guided robust normalization method is proposed. Through a fusion strategy of "physical range truncation + linear normalization + nonlinear robust compression," it effectively suppresses gradient perturbations in anomalous samples while preserving the physical proportions of normal samples. Furthermore, it constructs five-dimensional physical enhancement features: stress ratio, temperature-humidity coupled creep factor, size effect factor, load stiffness ratio, and creep time factor. These features explicitly encode concrete fracture mechanics and creep theory into the input space, reducing the model's dependence on large-scale samples. A multi-scale dilated convolution feature fusion mechanism is adopted, using parallel one-dimensional convolutional branches with different dilation rates to extract neighboring features, medium-span and long-distance coupling relationships respectively, and dynamically integrating multi-scale information through sample adaptive weighted fusion. Simultaneously, a global-local feature recalibration module is introduced, using the current sample features and the global skewness and kurtosis statistics of the training set together for attention weight generation, enabling the model to identify abnormally important features. A learnable activation function inspired by damage mechanics is proposed to simulate the nonlinear damage evolution law of concrete from "elasticity-strengthening-softening" to replace the traditional monotonic activation function. A physical consistency constraint is introduced into the loss function, constructing a high-risk probability based on the stress ratio theory, and applying a squared penalty when the model predicts a high-risk probability lower than the theoretical value. Combining higher-order statistics (skewness, kurtosis) of the fused features with the prior probability of risk level, adaptive initialization of network weights and output layer bias is achieved, improving training stability and minority class recognition ability. Attached Figure Description

[0035] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0036] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0037] Figure 2 Figure 1 shows the distribution of physical enhancement features. Figure 2 shows the sample distribution of strain characteristics, Figure 3 shows the sample distribution of temperature and humidity coupled creep factor, Figure 4 shows the sample distribution of size effect factor, Figure 5 shows the sample distribution of load stiffness ratio feature, and Figure 6 shows the sample distribution of creep time factor.

[0038] Figure 3 This is a schematic diagram of the physical consistency loss constraint. Detailed Implementation

[0039] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0040] Example 1

[0041] like Figure 1 As shown, a neural network-based method for classifying and grading the crack resistance risk of concrete main towers includes:

[0042] Obtain parameters used to assess the crack resistance risk of the concrete main tower under actual working conditions;

[0043] The parameters are normalized based on physical dimensions and abnormal robustness to obtain a normalized parameter vector; based on concrete fracture mechanics and creep theory, physical enhancement features are constructed according to the parameters, and the normalized parameter vector and physical enhancement features are concatenated to form an initial feature vector.

[0044] The initial feature vector is input into the pre-trained crack resistance risk classification model, and the crack resistance risk level classification result is output after the model is processed.

[0045] In a specific implementation, a data acquisition and training dataset based on finite element simulation is constructed;

[0046] This invention constructs a foundational dataset for crack resistance risk analysis of concrete main towers using the finite element method (FEM). Specifically, based on the structural design parameters, material performance parameters, construction process parameters, and environmental condition parameters of concrete main towers in actual engineering projects, a refined finite element analysis model is established. Geometric parameters include the tower height, cross-sectional width, cross-sectional thickness, and tilt angle; material parameters include concrete compressive strength, elastic modulus, Poisson's ratio, and creep coefficient; and environmental and construction parameters include ambient temperature, ambient humidity, construction load, and curing time. By changing the values ​​of these parameters, and employing the controlled variable method and orthogonal experimental design, multiple analysis conditions covering different structural forms, material grades, construction conditions, and environmental conditions are generated. Finite element calculations are performed on each condition, and the maximum principal tensile stress and maximum crack width are extracted from the structural response results as the basis for risk assessment.

[0047] For each working condition, its corresponding 14 input parameters and 2 response parameters are used to form an original sample. During the sample collection process, special attention is paid to the generation of extreme and boundary conditions, such as combinations of high stress ratio, large size, high temperature difference, high humidity, and high construction load, to ensure that the dataset can cover the complete range from low risk to high risk.

[0048] After completing the finite element simulation, all samples were cleaned to remove invalid samples with obvious physical anomalies or those failing to converge, forming the original sample set. Subsequently, the samples in the original sample set were labeled with risk levels, categorized into five levels: extremely low risk, low risk, medium risk, high risk, and extremely high risk. The labeling process was based on the crack control standards and crack resistance safety factor requirements in the concrete structure design code, combined with the ratio of the maximum principal tensile stress to the concrete tensile strength and the range of the maximum crack width: extremely low risk when the maximum principal tensile stress is much less than the concrete tensile strength and the maximum crack width is close to zero; low risk when the maximum principal tensile stress is less than the concrete tensile strength and the maximum crack width is within the small range allowed by the code; medium risk when the maximum principal tensile stress is close to the concrete tensile strength and the maximum crack width is in the middle of the code limit; high risk when the maximum principal tensile stress exceeds the concrete tensile strength but to a limited extent and the maximum crack width is close to the code upper limit; and extremely high risk when the maximum principal tensile stress significantly exceeds the concrete tensile strength and the maximum crack width exceeds the code limit.

[0049] After labeling, all samples are divided into training, validation, and test sets according to a preset ratio. The training set is used for subsequent model training and parameter learning, the validation set is used for hyperparameter tuning and model selection during training, and the test set is used for independent evaluation of the final model performance. Through the above data collection and labeling process, a training dataset containing rich working condition information and clear risk level labels is constructed, providing a high-quality data foundation for subsequent risk classification models based on neural networks.

[0050] In a specific implementation, the normalization process and physical enhancement process are as follows:

[0051] (1) While conventional Z-score standardization can eliminate dimensional differences, it weakens the original engineering-significant proportional relationships between some parameters. For example, the relative relationship between the maximum principal tensile stress and material strength is no longer intuitive after standardization, which is not conducive to the subsequent model's identification of crack risk. This invention maintains a stable mapping for normal engineering samples through "physical range truncation + linear normalization + nonlinear robust compression," while preserving the anomalousness of abnormal samples that exceed the physical range but suppressing their extreme effects. The specific steps are as follows:

[0052] 1> Obtain the original sample data, denoted as , indicating the first The original parameter set corresponding to each training sample , Indicates the total number of training samples;

[0053] Each sample contains 14 input parameters: main tower height H (in meters, symbol "m"), cross-sectional width B (in meters, symbol "m"), cross-sectional thickness D (in meters, symbol "m"), tilt angle θ (in degrees, symbol "°"), and concrete compressive strength. (Unit is "megapascal", symbol "MPa") Elastic modulus (Unit is "megapascal", symbol "MPa"), Poisson's ratio ν (dimensionless), creep coefficient (Dimensionless) Ambient temperature (Unit: Degrees Celsius, symbol: °C), Ambient humidity (RH, unit: percentage, symbol: %), Construction load (Unit is "kilonewton", symbol "kN"), maintenance time (Unit is "day", symbol "d") Maximum principal tensile stress (Unit is "megapascal", symbol "MPa") Maximum crack width (The unit is "millimeters", and the unit symbol is "mm"). The above 14 parameters together constitute the original input vector of a single sample.

[0054] Furthermore, physical minimum and maximum values ​​are set for each of the 14 input parameters. These physical minimum and maximum values ​​are used to limit the normal range of values ​​for the parameters within the acceptable range for engineering applications. In specific implementations, these ranges can be given based on concrete structure design codes, finite element modeling boundary conditions, and engineering experience. For example, the compressive strength of concrete... The physical range can be from 20MPa to 60MPa, with a maximum crack width. The physical range can be from 0mm to 0.5mm. For other parameters, corresponding physical range tables are also established.

[0055] 2> Perform physical range truncation on each parameter in each sample, limiting extreme values ​​that significantly deviate from common engineering sense to a controllable range. Specifically,

[0056] When a parameter value is lower than its physical minimum value, replace the parameter value with the physical minimum value; when a parameter value is higher than its physical maximum value, replace the parameter value with the physical maximum value; when a parameter value is within the physical range, retain the original value.

[0057] 3> Perform linear normalization on the parameters after physical range truncation, mapping each parameter to a uniform numerical range around 0 to 1. The obtained linear normalization result is used to preserve the relative proportional relationship of each parameter within the normal physical range as much as possible.

[0058] 4. Construct nonlinear robust mapping components based on statistical information from the training set. Specifically,

[0059] For the The parameters are used to calculate the sample mean and sample standard deviation across the entire training set, where the _th parameter is the _th parameter. The sample mean of each parameter is denoted as . , No. The sample standard deviation of each parameter is denoted as... Then, the positional relationship of the original parameter values ​​relative to the mean and standard deviation is input into the logistic function to obtain a non-linear mapping result. This non-linear mapping result is used to smooth and compress outliers so that samples far from the normal range will not generate excessive gradient perturbations in the early stage of training.

[0060] 5> The linear normalization result and the nonlinear mapping result are weighted and fused to obtain the normalization parameter, denoted as . , indicating the first In the nth sample The normalized result of each parameter is expressed as: ,in, This represents the value of the j-th parameter in the i-th sample after physical range truncation and linear normalization. It represents the nonlinear robust mapping component of the j-th parameter in the i-th sample. This represents the robustness adjustment coefficient, used to balance the contributions of both factors, for example... 0.1 is acceptable.

[0061] It should be noted that the parameter values ​​of normal samples are located within their physical range, therefore their linear normalization results are... It can already accurately reflect its relative position within the normal range. At this point, because The values ​​are relatively small, and the fusion result is mainly composed of The decision is made, but for outlier samples, their original parameter values ​​are far from the mean, and the nonlinear mapping components... It will be smoothly compressed to a saturation value close to 0 or 1 through the logistic function, when When the value is not 0, this compressed information will be introduced into the fusion result, so that the normalized value of the abnormal sample is "pulled" towards the boundary, thereby suppressing its extreme influence while retaining its abnormal attributes.

[0062] 6> Repeat the above process for all 14 parameters to obtain the... The 14-dimensional normalized parameter vector of each sample is denoted as... , indicating the first The input vector after normalization of each sample has a dimension of 14, representing the basic input features that have undergone dimensional unification and anomaly suppression.

[0063] It should be noted that, in order to simultaneously address the issues of "large dimensional differences" and "strong disturbances from anomalous samples," this invention first sets a physical range, then performs linear normalization and nonlinear robust compression. This improves the stability of the subsequent training process without disrupting the main physical proportional relationships. Compared to simply using mean-variance standardization, this approach is more suitable for handling crack resistance risk data of concrete main towers with clearly defined engineering upper and lower limits.

[0064] (2) Since only 14 original input parameters are fed into the neural network, although basic information is included, many physical coupling relationships directly related to crack resistance risk are still implicit, such as the relationship between stress and strength, the creep development relationship under the combined effects of temperature and humidity, the influence of geometric dimensions on slenderness ratio and stress concentration, and the matching relationship between construction load and structural stiffness. In order to enable the model to focus on key crack resistance mechanisms even with limited samples, this invention constructs 5-dimensional physical enhancement features based on concrete fracture mechanics, creep theory, and structural design principles. The specific steps are as follows:

[0065] 1> with the first Based on the original and normalized parameters of each sample, a stress ratio characteristic is constructed, denoted as... , indicating the first The stress ratio of each sample characterizes the degree to which the maximum principal tensile stress is close to the tensile strength of concrete.

[0066] In practical implementation, stress ratio characteristics Adopting the first Maximum principal tensile stress of each sample The ratio is calculated by comparing the current principal tensile stress with the simplified tensile strength estimate. This ratio intuitively reflects the level of the current principal tensile stress relative to the material's tensile capacity. The simplified tensile strength estimate can be calculated according to the... Concrete compressive strength of individual samples The value is taken as 0.1 times the value. The larger the stress ratio, the closer the structure is to or exceeds the cracking threshold, and the higher the risk of cracking.

[0067] Furthermore, with the first Based on the original and normalized parameters of each sample, a temperature-humidity coupled creep factor is constructed, denoted as . , represents the creep development promotion intensity index under the combined effects of ambient temperature, ambient humidity and curing time in the i-th sample, used to express the long-term impact of environmental and time conditions on crack risk;

[0068] In the specific implementation, the first The ambient temperature of each sample , No. Environmental humidity of each sample With the Maintenance time for each sample Perform combined calculations and introduce a reference temperature. Scale correction is performed, denoted as ,in, This indicates the reference temperature (20℃ is acceptable). and These are coefficients used to adjust the degree of influence of temperature and humidity (e.g., 0.05 is acceptable. (0.02 is acceptable). Logarithmic transformation of curing time was used to simulate the pattern of rapid early development and stabilization in the later stage of creep, so as to reflect the development characteristics of "rapid growth in the early stage and slow growth in the later stage".

[0069] Furthermore, with the first Based on the original and normalized parameters of each sample, a size effect factor is constructed, denoted as . , indicating the first The sample is a geometric risk index determined by the main tower height, cross-sectional width and cross-sectional thickness, used to reflect the problem that slender structures and unfavorable cross-sectional shapes are more prone to local tensile stress concentration and crack initiation.

[0070] In practical implementation, first determine the cross-sectional width. and cross-sectional thickness The equivalent dimensions of the structural cross-section, combined with the height of the main tower. The slenderness ratio is constructed, and a width-to-thickness ratio correction term is introduced to reflect the influence of cross-sectional shape on overall stiffness and stress distribution, expressed as follows: ,in, The slenderness ratio characterizes the overall slenderness of the structure; Characterizing the equivalent size of the cross section, This represents a coefficient used to adjust for the effect of the width-to-thickness ratio of the cross section (e.g., (0.2 is acceptable). The width-to-thickness ratio correction term is used to penalize cases where the cross-sectional shape deviates significantly from a square.

[0071] Furthermore, with the first Based on the original and normalized parameters of each sample, a load stiffness ratio characteristic is constructed, denoted as . , indicating the first The degree of matching between construction load and structural stiffness in each sample;

[0072] In practical implementation, first determine the cross-sectional width. and cross-sectional thickness Calculate the cross-sectional area The cross-sectional area is denoted as ,in, Indicates the first The area of ​​the main tower section of each sample; then the construction load. With elastic modulus and cross-sectional area The combined calculation yields the load stiffness ratio. To facilitate numerical stability processing, this result can be multiplied by a fixed scaling factor (e.g., ...). ), represented as ;

[0073] The larger the load stiffness ratio, the more unfavorable the external load is under the given stiffness condition, and the higher the risk of structural deformation and cracking.

[0074] Furthermore, with the first Based on the original and normalized parameters of each sample, a creep time factor is constructed, denoted as . , indicating the first The cumulative creep degree after the combined effect of creep coefficient and curing time in each sample is used to enhance the model's ability to express the risk of long-term creep-induced cracking.

[0075] In practical implementation, the creep coefficient is used. Based on this, an exponentially decaying term that gradually saturates over time is introduced to simulate the process of creep effect from rapid initial development to stabilization in the later stage, expressed as: ,in, Represents the creep coefficient. Indicates maintenance time. Represents the time constant (e.g., 30 can be selected to control the saturation rate of creep over time. The exponential term, which characterizes the gradual saturation over time, simulates the process of creep effect from rapid development in the early stage to stabilization in the later stage.

[0076] 2> By splicing the above 5 physical enhancement features in a fixed order, we obtain the first... The physical enhancement feature vector of each sample, denoted as , indicating the first The 5-dimensional physical enhancement feature vector of each sample includes stress ratio, temperature and humidity coupled creep factor, size effect factor, load stiffness ratio feature and creep time factor, in sequence.

[0077] 3> Normalize the 14-dimensional parameter vector With 5D Physical Enhancement Feature Vector The concatenation yields a 19-dimensional initial feature vector, denoted as... , indicating the first The initial feature vector of each sample contains both the normalized basic parameters and the enhanced features that express the physical coupling relationship.

[0078] It should be noted that this invention does not simply increase the input dimension, but explicitly encodes the most critical crack resistance discrimination relationship in engineering into the input space. This can reduce the burden on the neural network to rely entirely on large-scale samples to "implicitly learn" the physical relationship. It is especially suitable for scenarios where the number of finite element simulation samples is limited and the risk categories are unbalanced, thereby enhancing the model's ability to perceive the causes of high risks and improving the interpretability and generalization stability of the classification results.

[0079] In one embodiment, the distribution of physical enhancement features is analyzed, such as Figure 2 As shown, the sample distribution histograms of five physical enhancement features (stress ratio, temperature-humidity coupled creep factor, size effect factor, load stiffness ratio feature, and creep time factor) are presented. The enhancement features constructed based on concrete fracture mechanics and creep theory in this invention have clear physical meanings, and their numerical distribution can reflect the differences in crack resistance risk under different working conditions, providing crucial prior information for neural networks. The horizontal axis of each subplot represents the "eigenvalue" (dimensionless), and the vertical axis represents the "frequency" (dimensionless).

[0080] In a specific implementation, the crack resistance risk classification model is constructed based on a deep neural network with feature distribution and risk prior. The model includes: a multi-scale feature extraction module, a deep neural network parameter adaptive initialization module, a global-local feature recalibration module, a deep feature extraction module, and a classification output module.

[0081] (1) Multi-scale feature extraction module: The crack resistance risk of concrete main towers has obvious multi-scale characteristics. Geometric dimensions and cross-sectional shape reflect more macro-scale information, material parameters and creep parameters reflect more meso-scale information, and stress and crack indices are closer to the risk result itself. If the initial feature vector is directly fed into the fully connected layer, information of different scales will be mixed and processed, and the model will find it difficult to actively distinguish the feature contributions under different perception ranges. This invention adopts a multi-branch one-dimensional dilated convolution and adaptive weighted fusion mechanism to perform multi-scale feature extraction on the initial feature vector. The specific steps are as follows:

[0082] 1> Initialize the feature vector The sequence is directly used as the input sequence for a one-dimensional convolutional layer, that is, the sequence is... It is considered as a one-dimensional feature sequence arranged in a fixed order, and input into multiple one-dimensional convolutional branches set in parallel.

[0083] In one implementation, three parallel one-dimensional dilated convolution branches are set up. All three branches have a kernel size of 3 and an output channel count of 32, but their dilation rates are different. Specifically...

[0084] The expansion rate of the first branch is denoted as The value can be 1; the expansion rate of the second branch is denoted as... The value can be 2; the expansion rate of the third branch is denoted as... 4 is a possible value;

[0085] Different inflation rates indicate that each branch has a different receptive range on the input feature sequence. A small inflation rate focuses more on the local combination relationship between neighboring features, while a large inflation rate focuses more on the long-distance correlation relationship across feature locations.

[0086] The initial feature vector is processed by three one-dimensional dilated convolution branches respectively. The process is performed, and a non-linear activation function is applied after the convolution result of each branch. The non-linear activation function can be the ReLU activation function.

[0087] After three branches, three 32-dimensional feature vectors are obtained, denoted as... , and ,in, Indicates the first Feature extraction results of each sample on the first scale branch Indicates the first Feature extraction results of each sample on the second scale branch Indicates the first Feature extraction results of a sample on the third scale branch.

[0088] 2> Calculate adaptive fusion weights for each of the three scale branches. Specifically, set a set of learnable scoring parameters for each branch, perform linear scoring (i.e. linear mapping) on ​​the 32-dimensional feature vector output by the corresponding branch, obtain the importance score of the branch to the current sample, and then obtain the three weight coefficients through Softmax normalization.

[0089] The three weighting coefficients are denoted as follows: , and ,in, Indicates the first The fusion weights of each sample on the first scale branch Indicates the first The fusion weights of each sample on the second scale branch Indicates the first The fusion weights of the samples on the third scale branch, and the sum of the three fusion weights is 1.

[0090] 4> The outputs of the three branches (i.e., the feature extraction results on each scale branch) are weighted and summed according to the fusion weights to obtain the multi-scale fused feature vector, denoted as . , indicating the first The multi-scale fusion feature vector of each sample has a dimension of 32, representing the highly discriminative input features after fusing perceptual information at different scales.

[0091] 5> Repeat the above process for all training samples to obtain a multi-scale fusion feature set.

[0092] It should be noted that this invention avoids simply mixing information from different physical scales. By using multiple parallel convolutional branches with different dilation rates, it can extract nearest neighbor combination relationships, medium span relationships, and long-distance coupling relationships respectively. Then, the sample adaptive fusion weights determine the strength of the role of different scales in the current sample, which not only improves the richness of feature expression, but also enables the model to automatically select more valuable scale information according to specific working conditions.

[0093] (2) Adaptive initialization of deep neural network parameters:

[0094] (2-1) Calculation of higher-order statistics of fused feature distributions

[0095] To make the network initialization more closely resemble the actual input distribution, this invention performs statistical analysis on the multi-scale fusion features of all training samples, extracting statistical measures such as the global mean vector, global variance vector, global skewness vector, and global kurtosis vector. Specifically,

[0096] Collect the multi-scale fusion feature vectors corresponding to all training samples in the multi-scale fusion feature set, and calculate the sample mean for each of the 32 feature dimensions of the multi-scale fusion feature vector to obtain the global mean vector, denoted as . , represents the mean of all training samples across 32 fused feature dimensions, and represents the center position of each feature dimension;

[0097] Furthermore, the sample variance is calculated for each of the 32 feature dimensions to obtain the global variance vector, denoted as... , represents the variance of all training samples across 32 fusion feature dimensions, characterizing the degree of dispersion of each feature dimension;

[0098] Furthermore, the skewness is calculated for each of the 32 feature dimensions to obtain the global skewness vector, denoted as... , represents the skewness result of all training samples on the 32 fusion feature dimensions, characterizing the degree of asymmetry in the distribution of each feature dimension. The larger the absolute value of skewness, the more likely that the feature dimension is to have a significant long tail or skewed distribution.

[0099] Furthermore, the kurtosis is calculated for each of the 32 feature dimensions to obtain the global kurtosis vector, denoted as... , represents the kurtosis of all training samples across 32 fused feature dimensions, characterizing the tail thickness and peak strength of each feature dimension. The larger the kurtosis, the more likely that the feature dimension contains heavy-tailed samples or a small number of strong anomalous distributions.

[0100] (2-2) Adaptive initialization of weights by combining higher-order statistics and risk priors

[0101] Conventional He initialization primarily determines the initial variance based on the input dimension, without considering the actual distribution of different feature dimensions or reflecting the varying importance of different features for risk identification. This invention introduces higher-order statistics and risk priors into the first-layer weight initialization process, making the network more sensitive to key features from the beginning of training. The specific steps are as follows:

[0102] 1> Construct the first fully connected layer of the deep neural network. The weight matrix of the first fully connected layer is denoted as... , represents the weight matrix that maps 32-dimensional input features to 128-dimensional hidden layer features, with size . The bias vector of the first fully connected layer is denoted as The dimension is 128;

[0103] Let the number of input dimensions of the 32-dimensional input feature be denoted as... , which is the input dimension of the first fully connected layer, with the basic variance initialized by He as the initial reference scale.

[0104] 2> Based on the global skewness vector and global kurtosis vector The initial variance corresponding to each input feature dimension is corrected. Specifically, for feature dimensions with large absolute skewness or large kurtosis, the initial variance of the weights corresponding to that dimension is appropriately increased, so that the network can more fully perceive these key features with asymmetric distribution, heavy tails or strong anomalies in the early stages of training.

[0105] In the specific implementation, the correction factor is first defined. By global skewness and global kurtosis Joint decision, indicating as ,in, and These are hyperparameters that control the influence of skewness and kurtosis on intensity (e.g.) ),use This is because the kurtosis of the standard normal distribution is 3; a kurtosis higher than 3 indicates the presence of heavy tails or outliers. Then, the basic variance initialized with He is... Multiply by the correction factor This yields the corrected variance corresponding to the input dimension. , represented as .

[0106] 3> Calculate the prior probability of the risk level. The prior probability of the risk level is obtained by statistically analyzing the labels of the training samples. Suppose there are N samples in the training set, and the number of samples belonging to risk level r is... Then the prior probability of this category is ;

[0107] There are a total of 5 risk levels, corresponding to extremely low risk, low risk, medium risk, high risk, and extremely high risk. The prior probability of a risk level is denoted as , indicating the first in the training set The proportion of samples at each risk level .

[0108] 4. Establish feature importance descriptions based on expert knowledge. Specifically, assign importance strength to the correlation between each input feature dimension and each risk level. The importance strength is denoted as... , indicating the first The input feature dimension for the th input feature dimension The importance of risk level identification can range from 0 to 1, for example, Indicates the first The input feature dimension for the th input feature dimension The importance of risk level identification;

[0109] For dimensions closely related to high risk, such as stress ratio, crack width, and load stiffness ratio, a higher importance strength can be assigned.

[0110] For example, the stress ratio characteristic, which is crucial for identifying high-risk and extremely high-risk situations, can be set to... For extremely low risk, it is set as For the maximum crack width, it is also sensitive to high / extremely high risk and can be set to a higher value; for the load stiffness ratio characteristic, it has a significant impact on high risk and can be set to a lower value. For Poisson's ratio ν, it has little differentiation across all risk levels, so it can be uniformly set to a low value, such as 0.2.

[0111] 5> Based on the prior probability of the risk level and importance intensity Calculate the risk importance factor for each input feature dimension, denoted as . , indicating the first The degree of enhancement that each input feature dimension should receive during the initialization phase; the larger the risk importance factor, the more the feature dimension needs to be given priority in the early stages of training.

[0112] In practical implementation, the risk importance factor is obtained by weighting and summing the importance intensity of each risk level according to the prior probability of that level, and is expressed as follows: That is, the importance of a feature dimension depends on the strength of its importance at each risk level and the frequency with which that risk level appears in the dataset.

[0113] 6> Based on the combined results of the basic variance, higher-order statistics correction, and risk importance factors, set the sampling variance for the weights corresponding to each input dimension in the first fully connected layer. Specifically,

[0114] Random sampling is performed using a zero-mean normal distribution, so that the weights of the same input feature dimension on different output neurons share the same initial variance rule, but the random values ​​are different. For input feature dimensions with stronger skewness, higher kurtosis and higher risk importance, the corresponding weight variance is larger.

[0115] In practical implementation, the sampling variance is set for the weights corresponding to each input dimension in the first fully connected layer by comprehensively considering the basic variance, the correction results of higher-order statistics, and the risk importance factor. That is, for input dimension q, its corresponding weight matrix is... The weights in the algorithm (i.e., the weights connecting the q-th input to all 128 neurons) will be calculated from a mean of 0 and a variance of 0. Sampling from a normal distribution, denoted as ,in, This represents the variance after skewness and kurtosis correction. Hyperparameters (such as) This is used to control the extent to which the risk importance factor enhances the final variance.

[0116] 7> Set the first layer bias vector Initialize it as a vector of all zeros to ensure that the center of the linear combination does not shift during the early stages of training.

[0117] It should be noted that the core idea of ​​this invention is to "make the weight initialization conform to the data itself, rather than making the data adapt to a uniform initialization assumption". By introducing distribution skewness, kurtosis and risk importance into the initialization process, the response to key anti-cracking features can be strengthened at the beginning of network training, reducing early training oscillations and improving convergence speed and recognition stability.

[0118] (2-3) Output layer bias initialization based on prior probability

[0119] In multi-class classification tasks, if the number of samples in each class differs significantly, and the output layer biases are all initialized to 0, the model tends to output an approximately uniformly distributed class probability in the early stages of training. This is inconsistent with the true class distribution, resulting in a lengthy correction process during the initial training phase. This invention directly initializes the output layer biases using prior probabilities of risk levels. The specific steps are as follows:

[0120] 1> Construct the output layer bias vector, denoted as , represents the bias vector of the output layer, with a dimension of 5, corresponding to 5 risk level categories;

[0121] 2> Statistical analysis of the prior probabilities of the five risk levels in the training set And add a smoothing term to each class of prior probabilities. This represents a smoothing parameter to prevent division by zero and to prevent numerical instability caused by extreme probabilities; for example, it can be taken as... .

[0122] 3> Based on the prior probability of each risk level, initialize the bias of the corresponding category in the output layer to a value that matches the prior probability. Specifically, set it in log odds form so that the category probability distribution output by the model at the beginning of training is closer to the true distribution of the training set.

[0123] In the specific implementation, a logarithmic probability method is used to set the output bias for the r-th risk level. Initialize to ,in, Represents very small smoothing terms (e.g.) ), to prevent In this case, the initialization method ensures that when all other inputs are 0, the r-th class probability output by the Softmax function is approximately equal to... .

[0124] 4> After completing the output layer bias initialization, keep the output layer weights in the normal initialization method, such as the He initialization method.

[0125] It should be noted that, in order to shorten the time for the model to transition from a "uniform prediction state" to a "state that conforms to the true class distribution", this design is particularly helpful in mitigating the problem of slow learning of a few classes in the early training when the proportion of high-risk and extremely high-risk samples is small.

[0126] (3) Global-Local Feature Recalibration Module:

[0127] Conventional channel attention methods typically generate weights based solely on the features of the current sample, failing to determine the degree of anomaly of that sample within the global training distribution. For crack risk problems, many high-risk samples exhibit significant deviations in certain key features relative to the overall distribution, such as significantly higher stress ratios, crack widths, and load stiffness ratios. This invention combines global statistical information with the features of the current sample to generate attention weights, enabling the model to highlight anomalous and important feature dimensions. The specific steps are as follows:

[0128] 1> Take the first Multi-scale fusion feature vector of each sample As the current input feature, and using the global skewness vector and global kurtosis vector By concatenating the data in a fixed order, a global statistical vector is obtained, which represents the skewness and heavy-tailedness of the overall distribution of the training data.

[0129] Furthermore, the multi-scale fusion feature vector of the current sample is... The current sample features and the global statistical vector are simultaneously input into the feature recalibration module. Specifically, the current sample features and the global statistical vector are projected onto the same feature space through linear mapping, then summed and fused, with a bias term added to obtain the fused result. Finally, a Sigmoid activation is applied to the fused result to obtain a 32-dimensional attention weight vector, denoted as... , indicating the first Each sample corresponds to a feature recalibration weight vector, with each element ranging from 0 to 1, representing the importance of the corresponding feature dimension in the current sample.

[0130] 2> Fuse feature vectors across multiple scales With attention weight vector Element-wise multiplication yields the recalibrated eigenvectors, denoted as . , indicating the first The feature vector of each sample after global-local joint recalibration has a dimension of 32.

[0131] It should be noted that this invention incorporates both "current sample features" and "overall distribution statistics" into the attention generation process, which allows the model to not only know "what features the sample itself has", but also "whether these features are abnormal in the overall data", thus making it more conducive to identifying high-risk samples.

[0132] (4) Deep feature extraction module:

[0133] The stress-induced damage process of concrete materials typically involves an elastic stage, a nonlinear development stage, and a damage softening stage. Conventional activation functions such as ReLU struggle to simulate this "increase-then-decrease" damage evolution, resulting in limited expressive power of networks for the crack risk formation process. This invention introduces a learnable damage activation function and performs deep feature extraction through a 3-layer fully connected network. The specific steps are as follows:

[0134] 1> Construct the first deep feature extraction layer, and denote the weight matrix of the first layer as follows. The first layer bias vector is denoted as ,in, This represents the weight matrix that maps from a 32-dimensional input to a 128-dimensional hidden layer. This represents the bias vector of the first layer. The input to the first layer is the recalibrated feature vector. ;

[0135] Construct a second deep feature extraction layer, and denote the weight matrix of the second layer as follows. The second layer bias vector is denoted as ,in, This represents the weight matrix that maps from a 128-dimensional input to a 256-dimensional hidden layer. This represents the bias vector of the second layer;

[0136] Construct a third deep feature extraction layer, and denote the weight matrix of the third layer as follows: The third layer bias vector is denoted as ,in, This represents the weight matrix that maps from a 256-dimensional input to a 128-dimensional hidden layer. This represents the bias vector of the third layer.

[0137] 2> A damage activation function is introduced after each of the above three fully connected layers. The damage activation function is denoted as... , represents the nonlinear activation function used to simulate the damage evolution characteristics of concrete. This activation function consists of a "smoothing enhancement part in the front section" and a "softening gated part in the back section". The overall characteristics are as follows: when the input is small, the output increases approximately linearly; when the input increases to a certain extent, the output growth slows down; when the input further exceeds the threshold, the output begins to soften and decrease. This trend is closer to the evolution law of concrete materials under stress damage.

[0138] In practical implementation, the damage activation function The design motivation is to simulate the nonlinear behavior of concrete materials during stress damage, characterized by "elasticity-strengthening-softening." The calculation method is designed as a piecewise function or a continuous function with a learnable parameter. Specifically, it employs parametric GELU or Swish function variants. To more clearly represent the "softening" characteristic, the calculation method is expressed as follows: ,in, Represents the independent variable of the function. This indicates the shape control parameters of the front-end smoothing enhancement section. The parameter indicating the steepness of the softening gating section in the latter part. The threshold position parameter indicates when significant softening occurs. Let represent the Sigmoid activation function, and Control the growth rate of the smoothing enhancement section at the front (elastic section). Control the descent slope of the softening gating section in the later stage (damaged section). It is the threshold position where significant softening occurs; when When the function exhibits an approximately linear growth pattern followed by saturation (reinforcement), when... At this point, the function begins to decrease linearly (softening), and the learnable parameters... To enable the network to adaptively adjust this damage evolution curve to better fit the actual physical process, the initial value can be set to... , and .

[0139] 3> Recalibrate the feature vectors Enter the first The first layer is a fully connected layer that is processed by a damage activation function to obtain the hidden layer features, denoted as . , indicating the first The 128-dimensional feature vector of each sample after the first layer of deep feature extraction;

[0140] Furthermore, the features of the first hidden layer... The input to the second fully connected layer is processed by a damage activation function to obtain the hidden layer features of the second layer, denoted as . , indicating the first The 256-dimensional feature vector of each sample after the second layer of deep feature extraction;

[0141] Furthermore, the features of the second hidden layer... The input to the third fully connected layer is processed by a damage activation function to obtain the hidden layer features of the third layer, denoted as . , indicating the first The 128-dimensional feature vector of a sample after the third layer of deep feature extraction.

[0142] It should be noted that this invention replaces the traditional monotonic activation function with an activation function that is more consistent with the damage mechanism of concrete. This allows the network to not only represent general linear and nonlinear mapping relationships, but also to express the characteristic change law of "development-saturation-softening" which is closer to the crack failure process, thereby improving the model's ability to fit high-risk critical states and damage evolution processes.

[0143] (5) Classification output module:

[0144] While training solely based on cross-entropy loss may achieve high accuracy at the label level, it lacks constraints on predictions that clearly conflict with engineering physics principles. For example, when the stress ratio significantly exceeds a reasonable cracking threshold, the model might still output a high probability of low risk. Even if such a prediction aligns with some simulation labels, it may still violate fundamental principles of concrete fracture mechanics, reducing the model's credibility in real-world engineering applications. This invention incorporates physical consistency constraints into the classification loss, with the specific steps as follows:

[0145] 1> Construct the output layer, and denote the output layer weight matrix as follows: This represents the weight matrix that maps 128-dimensional deep features to 5 risk level outputs. The output layer bias vector is the one initialized in step S303. ;

[0146] The third hidden layer features The input-output layer obtains the raw scores corresponding to the five categories. Then, Softmax normalization is performed on the raw scores of the five categories to obtain the predicted probability vector, denoted as... , indicating the first The predicted probability distribution of 5 risk levels corresponding to samples, where the in the vector is the _th ... The elements are denoted as , indicating the first The sample belongs to the first The predicted probability of a risk level.

[0147] 2> Based on the stress ratio construction theory, the probability of high risk is high. Specifically, the stress ratio obtained from step S202 is taken. As input, it is mapped to a theoretical high-risk probability using the Sigmoid function, denoted as . , indicating the first Each sample has a theoretical high-risk probability obtained from prior knowledge of fracture mechanics.

[0148] Furthermore, to reflect the pattern that "risk increases rapidly when the stress ratio approaches 0.8", the center of the Sigmoid function can be set near 0.8, and a large slope parameter can be set (for example, the slope parameter can be set to 20), so that the probability of high risk increases rapidly when the stress ratio exceeds 0.8.

[0149] 3> Combine the probabilities of high-risk categories in the model prediction results to obtain the total high-risk probability predicted by the model, denoted as . , indicating the first The total probability that a sample is predicted to be at high risk or very high risk is specifically obtained by adding the probability of the fourth risk level and the probability of the fifth risk level.

[0150] 4> Calculate the physical consistency loss, denoted as This indicates that the model's prediction result must not be lower than the loss term used to constrain the trend of theoretically high-risk judgment. Specifically,

[0151] Physical consistency loss Used in the high-risk total probability predicted by the model Lower than the theoretical high-risk probability Imposing punishment at the time is represented as ,in, This represents the function that takes the maximum value. The function ensures that loss only occurs when the predicted probability is lower than the theoretical probability, which is in line with the design intention of "penalizing underestimation". The squared term makes the penalty increase exponentially as the deviation increases.

[0152] In practice, when the total probability of high risk predicted by the model is lower than the theoretical probability of high risk, a squared penalty is applied. When the total probability of high risk predicted by the model is no lower than the theoretical probability of high risk, the penalty is no longer applied, thus ensuring that high stress ratio samples are not wrongly assigned an excessively low probability of high risk by the model.

[0153] 5> Calculate the classification cross-entropy loss to constrain the predicted probability distribution to be consistent with the true label, where the true label vector is denoted as... , indicating the first The true class label vectors of each sample are encoded using one-hot encoding.

[0154] 6> The weighted sum of the classification cross-entropy loss and the physical consistency loss yields the total loss function, expressed as: ,in, Represents the classification cross-entropy loss. This represents the weight hyperparameter used to balance the two types of losses. The preferred value can be set to 0.2.

[0155] It should be noted that the constraint direction of the physical consistency loss is "when theoretical judgment shows that the probability of high risk should be high, the probability of high risk predicted by the model should not be significantly lower." Therefore, in the loss design, penalties should be imposed on the case where "the predicted probability of high risk is lower than the theoretical probability of high risk", rather than on the opposite case. This will enable the model to retain its data-driven learning ability while avoiding classification results that obviously violate the fracture mechanism of concrete, thereby improving the reliability and interpretability of the model in practical engineering applications.

[0156] like Figure 3 As shown, the scatter plot represents the model's predicted total probability of high risk for samples with different stress ratios. The black dashed line is the theoretical high-risk probability curve constructed based on the stress ratio, and the red filled area represents the portion where the predicted value is lower than the theoretical value (i.e., the area where a penalty is imposed). This invention constrains the model's predicted total probability of high risk to be no lower than the theoretical high-risk probability (based on the sigmoid mapping of the stress ratio) through physical consistency loss, ensuring that samples with high stress ratios are not incorrectly classified as low-risk, thereby improving the engineering reliability of the model's predictions. The horizontal axis represents the "stress ratio" (dimensionless), and the vertical axis represents the "high-risk probability" (dimensionless, range 0-1).

[0157] In a specific implementation, the model training process is as follows:

[0158] After completing the construction and parameter initialization of the deep neural network, the model training phase begins. This invention employs a mini-batch stochastic gradient descent optimization algorithm to iteratively update the network parameters. Each time, a fixed number of samples are randomly selected from the training set to form a batch. These samples undergo normalization, physical enhancement feature construction, and multi-scale dilated convolution feature fusion in step S2 to obtain a multi-scale fused feature vector for each sample. Then, through global-local feature recalibration, deep feature extraction based on damage mechanics, and calculation of the physical consistency loss function in step S4, the total loss value for that batch of samples is obtained.

[0159] Based on the total loss value, the gradients of all trainable parameters in the network are calculated using the backpropagation algorithm, and the parameters are updated using the adaptive moment estimation optimizer. The learning rate adopts an exponential decay strategy, maintaining a high learning rate in the early stage of training to accelerate convergence, and gradually reducing the learning rate as the number of iterations increases for fine-tuning.

[0160] After each training round, the current model is evaluated using the validation set. The classification cross-entropy loss and physical consistency loss on the validation set are calculated, and evaluation metrics such as classification accuracy, macro-average precision, and macro-average recall for the five risk levels on the validation set are recorded.

[0161] The model training stops iteratively using an early stopping strategy. Specifically, when the total loss on the validation set no longer decreases for several consecutive training epochs, or the macro-average accuracy on the validation set no longer improves for several consecutive training epochs, the early stopping mechanism is triggered, the model training is terminated, and the model parameters at the optimal performance on the validation set are saved as the final training result.

[0162] Optionally, to prevent model overfitting, weight decay regularization is introduced during training to impose norm penalty on the weight matrix in the network. At the same time, the loss difference between the training set and the validation set is monitored. If the difference continues to widen, training is terminated early to avoid overfitting.

[0163] Through the above training and stopping iteration judgment process, it is ensured that the final model fully learns the mapping relationship between crack resistance risk and input features on the training set, while maintaining good generalization ability on the validation set.

[0164] In a specific implementation, the trained model is applied to actual working conditions:

[0165] Once the model training is complete and the optimal parameters are saved, it can be applied to the classification and grading task of crack resistance risk assessment for concrete main towers. When a crack resistance risk assessment is required for a new concrete main tower, 14 input parameters for the main tower are first collected, including the tower height, cross-sectional width, cross-sectional thickness, tilt angle, concrete compressive strength, elastic modulus, Poisson's ratio, creep coefficient, ambient temperature, ambient humidity, construction load, curing time, and the maximum principal tensile stress and maximum crack width obtained through finite element analysis or on-site monitoring.

[0166] These parameters are organized in the same order and units as the training data to form the original input sample. This sample is then input into the trained deep neural network. First, it undergoes the fusion processing of physical range truncation, linear normalization, and nonlinear robust mapping in step S2 to obtain a normalized parameter vector. Simultaneously, based on concrete fracture mechanics and creep theory, five-dimensional physical enhancement features are constructed, including stress ratio, temperature-humidity coupled creep factor, size effect factor, load stiffness ratio feature, and creep time factor. These features are concatenated with the normalized parameter vector to obtain the initial feature vector. Finally, multi-branch one-dimensional dilated convolution and adaptive weighted fusion mechanism are used to extract multi-scale fusion features.

[0167] Then, the multi-scale fused feature enters the global-local feature recalibration module in step S4. It combines pre-stored global skewness and global kurtosis vectors to generate attention weights, recalibrating the fused feature. This is then passed through three fully connected layers using damage activation functions to extract deep nonlinear features. Finally, the output layer yields the raw scores corresponding to the five risk levels, which are then normalized using Softmax to output a predicted probability vector. The five elements in this probability vector correspond to the probability values ​​of extremely low risk, low risk, medium risk, high risk, and extremely high risk, respectively. The category with the highest probability is taken as the final crack resistance risk classification result for the concrete main tower.

[0168] Meanwhile, to enhance the interpretability of the prediction results, in addition to outputting the classification results, the specific values ​​of key physical enhancement features such as stress ratio characteristics, temperature and humidity coupled creep factor, and load stiffness ratio characteristics, as well as the model's prediction probability for high-risk and extremely high-risk categories, can be output to provide decision-making reference for engineering technicians.

[0169] Through the above process, the present invention can achieve rapid and accurate classification of crack resistance risk in concrete main towers, providing a scientific basis for crack control during construction and operation.

[0170] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A neural network-based method for classifying and grading the crack resistance risk of concrete main towers, characterized in that, include: Obtain parameters used to assess the crack resistance risk of the concrete main tower under actual working conditions; The parameters are normalized based on physical dimensions and abnormal robustness to obtain a normalized parameter vector; Based on concrete fracture mechanics and creep theory, physical reinforcement features are constructed according to parameters, and the normalized parameter vector is spliced ​​with the physical reinforcement features to form an initial feature vector. The normalization and physical enhancement processes are as follows: The collected evaluation parameters constitute sample data. Each sample is normalized. First, a physical minimum and a physical maximum value are set for each evaluation parameter in the sample. Based on the set physical minimum and physical maximum values, each parameter is truncated. Then, linear normalization is performed on the truncated parameters. Next, for each evaluation parameter, the sample mean and sample standard deviation are calculated. The positional relationship of the original parameter values ​​relative to the mean and standard deviation is then input into the logistic function to obtain the nonlinear mapping result. Finally, the linear normalization result of each evaluation parameter is weighted and fused with the nonlinear mapping result to obtain the normalized parameter vector; Based on concrete fracture mechanics, creep theory and structural design principles, and using each evaluation parameter and normalized parameter as a basis, stress ratio characteristics, temperature and humidity coupled creep factor, size effect factor, load stiffness ratio characteristics and creep time factor are constructed. The above five physical enhancement features are spliced ​​in a fixed order to obtain the physical enhancement feature vector of each sample. Then, the normalized parameter vector is spliced ​​with the physical enhancement feature vector to obtain the initial feature vector. The initial feature vector is input into the pre-trained crack resistance risk classification model, and the crack resistance risk level classification result is output after the model is processed. The crack resistance risk classification model is constructed based on a deep neural network with feature distribution and risk prior. The model includes: a multi-scale feature extraction module, a deep neural network parameter adaptive initialization module, a global-local feature recalibration module, a deep feature extraction module, and a classification output module. (1) The multi-scale feature extraction module adopts a multi-branch one-dimensional dilated convolution and an adaptive weighted fusion mechanism to perform multi-scale feature extraction on the initial feature vector to obtain a multi-scale fused feature vector; (2) Adaptive initialization of deep neural network parameters: First, calculate the higher-order statistics of the multi-scale fusion feature vectors of all samples, including the global mean vector, global variance vector, global skewness vector, and global kurtosis vector. Then, introduce the higher-order statistics and risk priors into the first layer of the deep neural network for adaptive weight initialization. Finally, use the risk level prior probability to directly initialize the output layer bias of the deep neural network. (3) The global-local feature recalibration module extracts global statistical information based on the multi-scale fusion feature vector, global skewness vector and global kurtosis vector, and then recalibrates to obtain the recalibrated feature vector; (4) The deep feature extraction module introduces a learnable damage activation function and performs deep feature extraction through a series of fully connected layers in the deep neural network after adaptive initialization. (5) Input the deep features into the classification output module. After adaptive initialization, the output layer of the deep neural network performs risk level classification prediction. Based on the classification cross-entropy loss, introduce the loss function of physical consistency constraint. Introduce physical consistency loss as a penalty term for the total probability of high risk being lower than the theoretical probability of high risk. Combine the classification cross-entropy loss to calculate the total loss of the model. The global-local feature recalibration module operates as follows: The input multi-scale fusion feature vector is concatenated with the global skewness vector and the global kurtosis vector in a fixed order to obtain the global statistical vector. The multi-scale fusion feature vector and the global statistical vector are then recalibrated. The two feature vectors are projected onto the same feature space through a linear mapping, then summed and fused, and a bias term is added to obtain the fusion result. The fusion result is then subjected to Sigmoid activation to obtain the attention weight vector. Finally, the multi-scale fusion feature vector and the attention weight vector are multiplied element-wise to obtain the recalibrated feature vector.

2. The method for classifying and grading the crack resistance risk of concrete main towers based on neural networks according to claim 1, characterized in that, The operation of the multi-scale feature extraction module is as follows: The initial feature vector of each sample is used as the input sequence of a one-dimensional convolutional layer, which is fed into multiple one-dimensional convolutional branches set in parallel. Each branch has the same number of convolutional kernels and output channels, but different dilation rates. The convolution result of each branch is then connected to a non-linear activation function. After passing through multiple one-dimensional convolutional branches, the feature extraction results with the same multiple dimensions are output. Then, a set of learnable scoring parameters is set for each branch, and the adaptive fusion weights of each branch are calculated to obtain the weight coefficient of each branch. Finally, the feature extraction results of each branch are weighted and summed according to the weight coefficients to output the multi-scale fusion feature vector of each sample.

3. The method for classifying and grading the crack resistance risk of concrete main towers based on neural networks according to claim 1, characterized in that, The specific steps for adaptive parameter initialization in deep neural networks are as follows: (1) Calculate the mean, variance, skewness and kurtosis of the multi-scale fusion feature vector corresponding to each sample in the feature dimension, and obtain the higher-order statistics composed of global mean vector, global variance vector, global skewness vector and global kurtosis vector; (2) Construct the first fully connected layer of the deep neural network. Correct the initial variance of the multi-scale fusion feature vector dimension of the input according to the global skewness vector and global kurtosis vector in the higher-order statistics to obtain the corrected variance. Then calculate the prior probability of the risk level based on the true label of the sample. Then determine the importance of the feature according to expert knowledge. Calculate the risk importance factor on each dimension according to the prior probability of the risk level and the importance of the feature. Finally, combine the initial variance, the corrected variance and the risk importance factor to set the sampling variance for the weights corresponding to each input dimension in the first fully connected layer. And initialize the bias vector of the first fully connected layer to a vector of all zeros; (3) Construct the output layer bias vector of the deep neural network, with the dimension corresponding to the number of risk level classification categories. Add a smoothing term to the prior probability of each risk level, then initialize the bias of the corresponding category of the output layer to a value that matches the prior probability, and finally keep the output layer weights in the normal initialization method.

4. The method for classifying and grading the crack resistance risk of concrete main towers based on neural networks according to claim 1, characterized in that, The deep feature extraction module is as follows: Construct a multi-layer deep feature extraction layer, and introduce a damage activation function after each layer. Input the recalibrated feature vector into the first deep feature extraction layer and pass it through the damage activation function to obtain the first hidden layer features. Then, use the output of the first deep feature extraction layer as the input of the next deep feature extraction layer. After layer-by-layer operation, the deep features are obtained.

5. The method for classifying and grading the crack resistance risk of concrete main towers based on neural networks according to claim 1, characterized in that, The specific operation of the classification output module is as follows: An output layer is constructed, and deep features are input into the output layer to obtain the original scores corresponding to each risk level category. Then, a Softmax normalization operation is performed on each original score to obtain the predicted probability vector. The theoretical high-risk probability is constructed based on the stress ratio feature in the constructed physical enhancement feature. The high-risk category probabilities in the prediction results are merged to obtain the total high-risk probability predicted by the model.

6. The method for classifying and grading the crack resistance risk of concrete main towers based on neural networks according to claim 1, characterized in that: actual The parameters used to assess the crack resistance risk of the concrete main tower in the working conditions are obtained by the finite element model. The finite element model is constructed based on the crack resistance risk assessment parameters of the concrete main tower in the actual project, and a sample dataset covering multiple working conditions is generated. The risk level of the samples in the sample dataset is labeled, and the labeled sample dataset is divided into training set and validation set.

7. The method for classifying and grading the crack resistance risk of concrete main towers based on neural networks according to claim 1, characterized in that, The pre-trained crack resistance risk classification model uses a mini-batch stochastic gradient descent optimization algorithm to iteratively update the parameters of the deep neural network. The specific training process is as follows: The dataset consisting of the acquired parameters is divided into a training set and a validation set. A fixed number of samples are randomly selected from the training set each time to form a batch. After normalization and physical feature enhancement, these samples are input into the crack resistance risk classification model. The model sequentially passes through a multi-scale feature extraction module, a deep neural network parameter adaptive initialization module, a global-local feature recalibration module, a deep feature extraction module, and a classification output module, ultimately obtaining the total loss value for that batch of samples. Based on the total loss value, the gradients of all trainable parameters in the model are calculated using the backpropagation algorithm, and the parameters are updated using an adaptive moment estimation optimizer. After each training iteration, the current model is evaluated using the validation set until the iteration stopping condition is met, resulting in a trained crack resistance risk classification model.

Images