CN-122020346-A - Concrete main tower crack resistance risk classification method based on neural network

CN122020346ACN 122020346 ACN122020346 ACN 122020346ACN-122020346-A

Abstract

The invention relates to the technical field of bridge structure health monitoring, in particular to a classification method for concrete main tower crack resistance risk based on a neural network, which comprises the following steps of obtaining parameters for evaluating the concrete main tower crack resistance risk in actual working conditions; the method comprises the steps of carrying out normalization processing on parameters based on physical dimension and abnormal robustness to obtain normalized parameter vectors, constructing physical enhancement features according to the parameters based on concrete fracture mechanics and creep theory, splicing the normalized parameter vectors and the physical enhancement features to form initial feature vectors, inputting the initial feature vectors into a pre-trained crack risk classification model, and outputting a crack risk classification result after model processing. The invention can accurately and efficiently finish the grading classification of the cracking risk of the concrete main tower through the neural network guided by physical information.

Inventors

LI XIAO
HAN QINGJUN
WANG ZHE
CUI DEQIANG
FANG LEI
WANG ZHIPENG
Yu Qianhao

Assignees

山东鹏程路桥集团有限公司
济宁济邹高速公路有限责任公司
山东鲁东交通建设集团有限公司

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. A concrete main tower crack resistance risk classification method based on a neural network is characterized by comprising the following steps: acquiring parameters for evaluating crack resistance risk of the concrete main tower in actual working conditions; Based on concrete fracture mechanics and creep theory, constructing physical enhancement features according to the parameters, and splicing the normalized parameter vectors and the physical enhancement features to form initial feature vectors; The initial feature vector is input into a pre-trained cracking risk classification model, and a cracking risk classification result is output after model processing.
2. The classification method for classifying crack resistance risks of concrete main towers based on the neural network according to claim 1, wherein the normalization processing and physical enhancement processes are as follows: The collected evaluation parameters form sample data, each sample is normalized, a physical minimum value and a physical maximum value are set for each evaluation parameter in the sample, each parameter is truncated based on the set physical minimum value and physical maximum value, and then linear normalization is carried out on the parameters after the truncation; calculating a sample mean value and a sample standard deviation of each parameter for each evaluation parameter, inputting a position relation of an original parameter value relative to the mean value and the standard deviation into a logic Style function to obtain a nonlinear mapping result, and finally carrying out weighted fusion on a linear normalization result and the nonlinear mapping result of each evaluation parameter to obtain a normalization parameter vector; Based on concrete fracture mechanics, creep theory and structural design principle, constructing stress ratio characteristics, temperature-humidity coupling creep factors, size effect factors, load stiffness ratio characteristics and creep time factors based on each evaluation parameter and normalization parameters, splicing the 5 physical reinforcement characteristics according to a fixed sequence to obtain physical reinforcement characteristic vectors of each sample, and splicing the normalization parameter vectors and the physical reinforcement characteristic vectors to obtain initial characteristic vectors.
3. The concrete main tower crack resistance risk classification method based on the neural network is characterized in that a crack resistance risk classification model is constructed based on a deep neural network with characteristic distribution and risk priori, and the model comprises a multi-scale characteristic extraction module, a deep neural network parameter self-adaptive initialization, a global-local characteristic recalibration module, a deep characteristic extraction module and a classification output module; (1) The multi-scale feature extraction module performs multi-scale feature extraction on the initial feature vector by adopting a multi-branch one-dimensional cavity convolution and self-adaptive weighted fusion mechanism to obtain a multi-scale fusion feature vector; (2) The method comprises the steps of carrying out parameter self-adaptive initialization on a deep neural network, firstly calculating high-order statistics of multi-scale fusion feature vectors of all samples, including a global mean value vector, a global variance vector, a global skewness vector and a global kurtosis vector, then introducing the high-order statistics and risk prior into a first layer of the deep neural network together for carrying out weight self-adaptive initialization, and then directly initializing output layer bias of the deep neural network by using risk level prior probability; (3) A global-local feature recalibration module is used for extracting global statistical information based on the multi-scale fusion feature vector, the global skewness vector and the global kurtosis vector, and then recalibration is carried out to obtain a feature vector after recalibration; (4) The depth feature extraction module introduces a leachable damage activation function and performs depth feature extraction through multiple serially connected full-connection layers in the depth neural network after self-adaptive initialization; (5) And (3) inputting the depth characteristics into a classification output module, carrying out risk level classification prediction on the output layer of the depth neural network after self-adaptive initialization, introducing a loss function of physical consistency constraint on the basis of classification cross entropy loss, and calculating the total loss of the model.
4. The method for classifying crack resistance risks of the concrete main tower based on the neural network according to claim 3, wherein the operation of the multi-scale feature extraction module is specifically as follows: The method comprises the steps of taking an initial feature vector of each sample as an input sequence of a one-dimensional convolution layer, inputting the initial feature vector into a plurality of one-dimensional convolution branches which are arranged in parallel, enabling convolution kernels and output channels of each branch to be the same in number and different in expansion rate, accessing a nonlinear activation function after convolution results of each branch, outputting feature extraction results with the same dimensions after the plurality of one-dimensional convolution branches, setting a group of leachable scoring parameters for each branch, respectively calculating self-adaptive fusion weights of the branches to obtain weight coefficients of each branch, and then carrying out weighted summation on the feature extraction results of each branch according to the weight coefficients to output multi-scale fusion feature vectors of each sample.
5. The classification method for classifying crack resistance risks of concrete main towers based on the neural network according to claim 3, wherein the parameter self-adaptive initialization operation of the deep neural network is as follows: (1) Respectively calculating the mean value, variance, skewness and kurtosis of the multi-scale fusion feature vector corresponding to each sample in the feature dimension to obtain a high-order statistic consisting of a global mean value vector, a global variance vector, a global skewness vector and a global kurtosis vector; (2) Constructing a first full-connection layer of the deep neural network, correcting an initial variance corresponding to the dimension of an input multi-scale fusion feature vector according to a global bias vector and a global kurtosis vector in a high-order statistic to obtain a corrected variance, calculating prior probability of a risk level based on a real label of a sample, determining the degree of feature importance according to expert knowledge, calculating a risk importance factor on each dimension according to the prior probability of the risk level and the degree of feature importance, and finally setting a sampling variance for the weight corresponding to each input dimension in the first full-connection layer by combining the initial variance, the corrected variance and the risk importance factor; (3) And constructing an output layer bias vector of the deep neural network, classifying the number of categories by the dimension corresponding to the wind direction level, adding a smoothing term for the prior probability of each type of risk level, initializing the bias of the category corresponding to the output layer to a numerical value matched with the prior probability, and finally keeping the weight of the output layer in a conventional initialization mode.
6. The method for classifying crack resistance risk classification of a concrete main tower based on a neural network according to claim 3, wherein the global-local feature recalibration module is operated as follows: Inputting a multiscale fusion feature vector, splicing the multiscale fusion feature vector and the global kurtosis vector according to a fixed sequence to obtain a global statistical vector, carrying out recalibration on the multiscale fusion feature vector and the global statistical vector, projecting the two feature vectors into the same feature space through linear mapping, adding and fusing, superposing offset items to obtain a fusion result, then carrying out Sigmoid activation on the fusion result to obtain an attention weight vector, and finally multiplying the multiscale fusion feature vector and the attention weight vector element by element to obtain a feature vector after recalibration.
7. The method for classifying crack resistance risk of a concrete main tower based on a neural network according to claim 3 is characterized in that the depth feature extraction module is specifically characterized by constructing a plurality of depth feature extraction layers, introducing a damage activation function after each layer, inputting a recalibrated feature vector into a first depth feature extraction layer, obtaining a first hidden layer feature through the damage activation function, then taking the first layer output as the input of the next layer, and obtaining the depth feature after layer-by-layer operation.
8. The classification method for classifying the crack resistance risk of the concrete main tower based on the neural network according to claim 3, wherein the classification output module is operated as follows: The method comprises the steps of constructing an output layer, inputting depth features into the output layer to obtain original scores corresponding to risk class categories, performing Softmax normalization operation on each original score to obtain a prediction probability vector, constructing theoretical high risk probability according to stress ratio features in constructed physical enhancement features, combining high risk category probabilities in prediction results to obtain model prediction high risk total probability, introducing physical consistency loss as a penalty item of which the high risk total probability is lower than the theoretical high risk probability, and calculating the total loss of a model by combining classification cross entropy loss.
9. The neural network-based concrete main tower crack resistance grading classification method is characterized in that parameters for evaluating the crack resistance of the concrete main tower in actual working conditions are obtained by a finite element model, the finite element model is built according to crack resistance evaluation parameters of the concrete main tower in actual engineering, a sample data set covering various working conditions is generated, risk grade labeling is conducted on samples in the sample data set, and the labeled sample data set is divided into a training set and a verification set.
10. The classification method for classifying the crack resistance risk of the concrete main tower based on the neural network according to claim 1, wherein a pre-trained crack resistance risk classification model adopts a small batch random gradient descent optimization algorithm to iteratively update the parameters of the deep neural network, and the training process is specifically as follows: The method comprises the steps of dividing an acquired data set formed by parameters into a training set and a verification set, randomly extracting a fixed number of samples from the training set each time to form a batch, inputting the batch into an anti-cracking risk classification model after normalization processing and physical feature enhancement operation, sequentially carrying out multi-scale feature extraction module, deep neural network parameter self-adaptive initialization, global-local feature recalibration module, deep feature extraction module and classification output module to finally obtain a total loss value of the batch of samples, calculating gradients of all trainable parameters in the model according to the total loss value through a back propagation algorithm, updating the parameters by utilizing an adaptive moment estimation optimizer, and after each training round is finished, evaluating the current model by utilizing the verification set until the iteration stop condition is met, and finally training to obtain a trained anti-cracking risk classification model.

Description

Concrete main tower crack resistance risk classification method based on neural network Technical Field The invention relates to the technical field of bridge structure health monitoring, in particular to a concrete main tower crack resistance risk classification method based on a neural network. Background Along with the development of modern bridge engineering to large-span, high pier stud and complex structure, the concrete main tower is used as the core bearing member of the large-span bridge such as cable-stayed bridge, suspension bridge and the like, and the crack resistance is directly related to the safety, durability and service life of the bridge structure. In practical engineering, a concrete main tower faces complex stress and environment coupling action in the construction period and the operation period, and the cracking risk is jointly influenced by various factors including geometric dimension parameters, material performance parameters, construction process parameters, environment condition parameters and the like, wherein strong nonlinear coupling relation exists between the parameters, the dimension difference of different parameters is huge, and the physical meaning is different. Although the traditional anti-cracking risk analysis method based on finite element simulation can calculate the stress field and the crack width under the specific working condition more accurately, the problems of high calculation cost, long analysis period, high requirements on professional ability of engineering personnel and the like exist, and the actual engineering requirements of quick evaluation, real-time early warning and batch working condition analysis are difficult to meet. In recent years, along with the rapid development of artificial intelligence technology, engineering risk prediction methods based on neural networks are paid attention to gradually, however, most of the prior art directly takes finite element simulation results as training data, and the original input parameters are simply normalized and then sent into a fully-connected network for training, so that the processing mode has obvious defects. The prior concrete structure crack resistance risk assessment multi-dependency finite element simulation or empirical formula is adopted, data sources are single, working conditions are limited, a complete training sample set covering extreme working conditions and boundary conditions is difficult to effectively generate, the model is insufficient in generalization capability when facing high-risk scenes such as high stress ratio, large size, complex temperature and humidity environment and the like, the risk critical state cannot be accurately identified, the traditional data preprocessing method adopts unified standardization or normalization processing, the inherent physical meaning and dimension differences of all parameters are ignored, key physical proportion relations of stress and strength, temperature and humidity and creep and the like are easy to break, meanwhile, for abnormal samples appearing in finite element simulation, the conventional method lacks an effective robust processing mechanism, abnormal values are easy to cause gradient oscillation in the initial training stage, the model convergence stability is influenced, the prior neural network model directly stacks original parameters on the input feature structure, prior physical knowledge such as concrete fracture mechanics, creep theory, size effect and the like is not explicitly coded, the model needs to rely on a large number of sample physical coupling relations, the prior data preprocessing method adopts unified standardization or normalization processing, the inherent physical meaning and dimension differences of all parameters are easy to break stress and strength, temperature and humidity and creep and the like are easy to break, the prior art is difficult to capture the inherent key physical proportion relations of the stress and strength, the prior model is difficult to break down-level stress coefficient, the prior neural network model is obviously limited to break down, the prior model has a crack resistance threshold value is greatly limited by the prior model (the prior model has a crack resistance threshold value is greatly limited by the prior model, the prior model has a crack resistance threshold value is greatly limited by the crack resistance threshold value when the prior to the model is greatly exceeds the threshold value, has a threshold-restricted by the prior stress-level model has a crack resistance threshold-oriented by the prior stress model has a threshold-oriented stress limitation and has a high-level structure and a high-level structure-down-a high-level structure and a high-quality structure, response delay or gradient instability in initial training stage easily occurs under the condition of non-Gaussian characteristic distribution Therefore, the invention provides a classification method f