A bridge pier damage position recognition method and system based on deep learning and a medium

Abstract

The present disclosure relates to a kind of bridge pier damage position identification method, system and medium based on deep learning, the method includes: obtaining bridge pier initial parameter from bridge standard drawing set, divide multiple bridge pier damage positions, combine bridge pier type with bridge pier damage position;According to bridge pier finite element model, construct including the inherent frequency of bridge pier under each bridge pier damage working condition, bridge pier first-order frequency, bridge pier first-order frequency change rate, bridge pier structure parameter;Original data set is standardized, and training set and test set are divided, and feature is increased;The multilayer perception machine neural network model including Dropout regularization module, weight attenuation regularization module, batch standardization module, cosine annealing learning rate scheduler is constructed, and the output is probability prediction distribution;Multiple network models are trained, and bridge pier damage position prediction model is obtained by average integration;Prediction is carried out using prediction model.The present disclosure is more stable and accurate to bridge pier damage position identification.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of bridge data processing, in particular to a bridge pier damage position identification method and system based on deep learning and a medium. BACKGROUND

[0002] The bridge pier is the core load-bearing component of the bridge, which bears the huge load of the upper part of the bridge and transmits it to the foundation. During use, the bridge pier is often damaged due to water erosion, freeze-thaw cycle, ship and vehicle impact, and natural disasters such as earthquakes. Bridge pier damage can seriously affect the safety and performance of the bridge, reduce the bearing capacity and stiffness of the bridge pier, affect the stability and durability, and even cause catastrophic collapse accidents. Therefore, it is of great significance to quickly identify the damage position of the bridge pier to ensure the long-term safe operation of the bridge.

[0003] Traditional bridge pier damage detection relies on manual visual inspection and local contact detection. This method relies on the subjectivity of the detection personnel and is difficult to inspect closely in dangerous areas such as high altitude and deep water. There are blind spots in the overall state evaluation of large bridge piers. To solve the above problems, more and more bridge pier damage evaluation methods have been proposed.

[0004] For example, the paper "A Railway Bridge Pier Damage Identification Method Based on Frequency Response Function Similarity" published in the journal "China Railway Science" proposes a quantitative identification method for railway bridge pier damage based on model correction theory and frequency response function similarity. The similarity coefficient of the self-vibration frequency and the frequency response function within a specific frequency band is used to construct the objective function, and the pier body damage degree and foundation constraint spring damage degree are used as identification indicators. An optimization algorithm is used to achieve damage evaluation of the railway simply supported beam bridge pier.

[0005] For example, the paper "Model Test Research on Railway Bridge Pier Damage Identification" published in the journal "Vibration and Shock" establishes a model bridge pier suitable for low piers of railway bridges. The measured frequency and mode shape are used as input, and an optimization algorithm is used to identify the damage of the model bridge pier.

[0006] For example, the Chinese patent with patent publication number CN119309633A and the name "Bridge Pier Impact Damage Detection Device and Method Based on Multi-source Data Fusion" proposes a method for comprehensively determining the damage state of the bridge pier after being impacted by arranging piezoelectric sensors in an array inside the bridge pier and setting a crack detection device on the surface.

[0007] For example, the Chinese patent with publication number CN114036974A and the name "Bridge Pier Impact Damage Detection Device and Method Based on Multi-source Data Fusion" proposes a method for dynamically identifying the foundation scour depth by analyzing the dynamic characteristics of the structure system.

[0008] The above studies demonstrate the feasibility of damage location identification methods based on dynamic characteristics. With the development of artificial intelligence technology, more and more artificial intelligence technologies are being applied in the field of bridge damage identification engineering. Neural networks can extract complex mapping features of input data and output target data to predict complex data relationships.

[0009] As the paper "Bridge Structure Damage Identification Method Based on Transfer Convolutional Neural Network" published in the Journal of Railway Science and Engineering, a structure damage identification method combining transfer learning and one-dimensional convolutional neural network is proposed. It provides an effective solution to the problem of structure damage identification under limited data conditions.

[0010] As the paper "Structure Damage Identification Based on Convolutional Neural Network" published in the journal "Vibration and Shock", convolutional neural networks are used to extract structural features to identify damage. To better improve the noise resistance, a hybrid noise training mode is proposed, which can identify more order frequency information.

[0011] As the paper "Structure Damage Identification Based on Self-training Semi-supervised Neural Network" published in the journal "Earthquake Engineering and Engineering Vibration", a structure damage identification framework based on self-training semi-supervised neural network is proposed. The self-training semi-supervised method is used to train the multi-layer perceptron neural network, and the samples with high confidence are selected to label pseudo-labels from unlabeled samples, expand the sample training set, and use normalized frequency change rate and damage characteristic index as input features to identify structural damage.

[0012] As the paper "Bridge Multi-risk Damage Identification Research Based on RF-MLP" published in the journal "Transportation Technology", a multi-layer perceptron (MLP) model based on random forest (RF) algorithm optimization is used for intelligent unsupervised feature learning, and the damage identification effect of the trained neural network is verified to realize the identification of damage location, damage type and damage degree under multi-damage working conditions.

[0013] As the paper "Bridge Early Warning Method and System Based on Physical Information Neural Network and Machine Vision" with publication number CN119964334A and name "Bridge Early Warning Method and System Based on Physical Information Neural Network and Machine Vision" proposed by the Chinese patent, the scheme is: real-time acquisition of bridge dynamic response data, input into the trained physical information neural network model, online evaluation of bridge operation state, checking of abnormal response and generation of warning signal.

[0014] As the paper "Bridge Structure Detection Method and System Based on 3D Gaussian Scattering" with publication number CN119559167A and name "Bridge Structure Detection Method and System Based on 3D Gaussian Scattering" proposed by the Chinese patent, based on a pre-trained deep learning model, the collected 3D Gaussian scattering data and environmental sensor data are dynamically fused to intelligently detect the health status of the bridge structure, identify potential damage and provide maintenance recommendations.

[0015] At present, studies have proved the feasibility of using neural networks for damage location identification methods. However, the existing bridge pier damage location identification method and system based on neural networks use existing bridge pier damage data for model training, which is extracted through field experiments, and the extraction is difficult and costly, thereby resulting in poor stability and accuracy of bridge pier damage location identification. SUMMARY

[0016] Therefore, it is necessary to provide a bridge pier damage location identification method, system and medium based on deep learning to improve the stability and accuracy of bridge pier damage location identification.

[0017] To solve the above problems, the present disclosure adopts the following technical solutions:

[0018] In a first aspect, the present disclosure provides a bridge pier damage location identification method based on deep learning, comprising the following steps:

[0019] Step 1: Obtain the initial parameters of the bridge pier from the bridge standard drawing set, divide the bridge pier damage location, and combine the bridge pier type with the bridge pier damage location to obtain the combination of the bridge pier damage working condition;

[0020] Step 2: According to the combination of the bridge pier damage working condition, a bridge pier finite element model is constructed by using finite element software, and an original data set for bridge pier damage location identification is constructed, wherein the original data set includes the bridge pier natural frequency, the bridge pier first-order frequency, the bridge pier first-order frequency change rate, and the bridge pier structure parameters under each bridge pier damage working condition;

[0021] Step 3: Standardize the original data set, divide the training set and the test set, and increase the features of the training set and the test set to obtain the training set after feature increase and the test set after feature increase;

[0022] Step 4: Construct a multi-layer perception neural network model, wherein the multi-layer perception neural network model includes a Dropout regularization module, a weight decay regularization module, a batch normalization module, and a cosine annealing learning rate scheduler for training the model;

[0023] Step 5: Train a plurality of multi-layer perception neural network models, wherein the output of the multi-layer perception neural network model is a probability prediction distribution, and the trained multi-layer perception neural network model is averaged to obtain a bridge pier damage location prediction model;

[0024] Step 6: Identify the damage location of the bridge pier using the bridge pier damage location prediction model.

[0025] In a preferred embodiment, the pier initial parameters include pier diameter, column height, pile-column ratio, damage location and soil parameters; and the obtaining of the pier initial parameters from the bridge standard drawing set comprises: obtaining pier initial parameter sample sets from the bridge standard drawing set, and expanding the pier initial parameter sample sets by interpolation method.

[0026] In a preferred embodiment, the step 2 comprises:

[0027] The horizontal resistance coefficient of the soil under each working condition is obtained, the horizontal action effect of the elastic pile is calculated by using the m method, and the soil constraint at different depths is simulated by using spring stiffness;

[0028] The calculation formula of the segmented spring stiffness of the soil is:

[0029] ;

[0030] In the formula, denotes the spring stiffness of the first segment; denotes the calculation width of the pile; denotes the horizontal resistance coefficient of the soil; denotes the segment length; denotes the distance from the ground to the spring of the first segment;

[0031] The calculation formula of the stiffness of the top concentrated spring is:

[0032] ;

[0033] In the formula, denotes the stiffness of the top concentrated spring; denotes the area of the foundation coefficient distribution diagram of the top spring of the pile within the length;

[0034] The constraint of the pier column top is calculated by using the spring unit to simulate the constraint action of the rubber bearing at the top of the pile and the main beam on the column, and the calculation formula of the spring stiffness of the pier column top is as follows:

[0035] ;

[0036] ;

[0037] ;

[0038] In the formula, denotes the spring stiffness of the pier column top; denotes the resistance stiffness of the plate rubber bearing; denotes the shear modulus of the rubber material; denotes the bearing area of the bearing; represents the thickness of the rubber sheet; represents the push stiffness of the main beam to the pier; represents the elastic modulus of the main beam concrete; represents the cross-sectional area of the main beam; 、 represents the length of the main beam on both sides of the pier top;

[0039] The pier first-order natural frequency change rate calculation formula is:

[0040] ;

[0041] In the formula, represents the pier first-order natural frequency change rate; represents the pier first-order frequency; represents the pier natural frequency.

[0042] In a preferred embodiment, the pier structure parameters include pile ratio, pile diameter, area, length, and moment of inertia.

[0043] In a preferred embodiment, the dividing the training set and the test set specifically includes: analyzing whether the data quantity of different damage positions in the original data set is balanced, if balanced, randomly dividing the standardized original data set to obtain the training set and the test set; if not balanced, based on the standardized original data set, using a stratified sampling method to determine the training set and the test set.

[0044] In a preferred embodiment, the calculating the polynomial features includes calculating square terms; and the calculating the statistical features includes calculating mean, standard deviation, maximum value, and minimum value.

[0045] In a preferred embodiment, the calculating the polynomial features, the performing feature cross, the calculating the statistical features

[0046] In a preferred embodiment, the training multiple multilayer perception neural network models includes: defining a hyperparameter space, the hyperparameters including hidden layer structure, Dropout rate, learning rate, weight decay, and batch normalization; using a random search method to randomly sample in the hyperparameter space, and performing multiple rounds of experiments to obtain an optimal hyperparameter combination; using the best hyperparameter combination, combined with different initializations, to train multiple neural network models.

[0047] In a preferred embodiment, the method further includes evaluating and verifying the pier damage position prediction model using evaluation indexes, the evaluation indexes including accuracy, recall rate, and F1 score.

[0048] In a second aspect, the present disclosure provides a pier damage position recognition system based on deep learning, comprising:

[0049] The dividing and combining module is used for obtaining initial parameters of the bridge pier from a bridge standard drawing set, dividing a plurality of bridge pier damage positions, and combining a bridge pier type and the bridge pier damage positions to obtain a combination of bridge pier damage working conditions.

[0050] The data set construction module is used for constructing a bridge pier finite element model through finite element software according to the combination of the bridge pier damage working conditions, and constructing an original data set of bridge pier damage position identification, which includes a bridge pier natural frequency, a bridge pier first-order frequency, a bridge pier first-order frequency change rate and a bridge pier structure parameter under each bridge pier damage working condition.

[0051] The data set processing module is used for data standardization of the original data set, division of a training set and a test set, feature addition of the training set and the test set, and obtaining a feature-added training set and a feature-added test set.

[0052] The model construction module is used for constructing a plurality of multilayer perceptron neural network models, which include a Dropout regularization module, a weight decay regularization module, a batch normalization module and a cosine annealing learning rate scheduler for training the models.

[0053] The training module is used for training the plurality of multilayer perceptron neural network models, the output of the multilayer perceptron neural network models being a probability prediction distribution, and obtaining a bridge pier damage position prediction model through average integration of the trained multilayer perceptron neural network models.

[0054] The prediction module is used for identifying a damage position of a bridge pier by using the bridge pier damage position prediction model.

[0055] In a third aspect, the disclosure provides a non-transitory computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the bridge pier damage position identification method based on deep learning according to the first aspect.

[0056] The bridge pier damage position identification method, system and medium based on deep learning described above are used for obtaining initial parameters of the bridge pier from a bridge standard drawing set, dividing a plurality of bridge pier damage positions, combining a bridge pier type and the bridge pier damage positions to obtain a combination of bridge pier damage working conditions, obtaining an original data set of bridge pier damage position identification based on finite element software, and performing feature addition on the original data set. Compared with the prior art, the original data set is more easily obtained, the original data set covers a wider range, and the features of the data set used for training are more comprehensive. Based on the feature-added training set and the feature-added test set, a bridge pier damage position prediction model with high accuracy and stability is obtained through model construction and average integration, and then the bridge pier damage position prediction is more accurate and stable. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 Flowchart of the bridge pier damage position identification method in one embodiment of the present disclosure;

[0058] Figure 2 Flowchart of step 3 of the bridge pier damage position identification method in one embodiment of the present disclosure;

[0059] Figure 3 Flowchart of step 5 of the bridge pier damage position identification method in one embodiment of the present disclosure;

[0060] Figure 4 Training loss diagram in the training process of the bridge pier damage position identification method in one embodiment of the present disclosure;

[0061] Figure 5 Learning rate change diagram in the training process of the bridge pier damage position identification method in one embodiment of the present disclosure;

[0062] Figure 6 Comparison diagram of test set accuracy of each damage position of the bridge pier damage position identification method in one embodiment of the present disclosure;

[0063] Figure 7 Confusion matrix of the prediction result of the bridge pier damage position identification method in one embodiment of the present disclosure;

[0064] Figure 8 Comparison diagram of ROC curve and AUC of each damage position of the bridge pier damage position identification method in one embodiment of the present disclosure;

[0065] Figure 9 Structural diagram of the bridge pier damage position identification system in one embodiment of the present disclosure. DETAILED DESCRIPTION

[0066] The technical solutions of the present disclosure will be described in detail below with reference to the accompanying drawings and preferred embodiments.

[0067] Referring to Figure 1 The present disclosure provides a bridge pier damage position identification method based on deep learning, comprising:

[0068] Step 1: Obtain the initial parameters of the bridge pier from the standard bridge drawing set, divide the various damage positions of the bridge pier, and combine the types of the bridge pier and the damage positions of the bridge pier to obtain the combination of the damage working conditions of the bridge pier;

[0069] Step 2, according to the combination of the pier damage working condition, the finite element software is used to construct the pier finite element model, and the original data set for identifying the pier damage position is constructed, the original data set includes the pier natural frequency, the pier first-order frequency, the pier first-order frequency change rate, and the pier structure parameters under each pier damage working condition.

[0070] Step 3, the original data set is standardized, the training set and the test set are divided, the feature of the training set and the test set is increased, and the training set and the test set after the feature is increased are obtained.

[0071] Step 4, a multi-layer perception neural network model is constructed, the multi-layer perception neural network model includes a Dropout regularization module, a weight attenuation regularization module, a batch normalization module, and a cosine annealing learning rate scheduler for training the model.

[0072] Step 5, a plurality of the multi-layer perception neural network models are trained, the output of the multi-layer perception neural network model is a probability prediction distribution, the trained multi-layer perception neural network model is averaged and integrated to obtain a pier damage position prediction model.

[0073] Step 6, the pier damage position prediction model is used to identify the damage position of the pier.

[0074] The steps of the pier damage position identification method based on deep learning are described below.

[0075] Step 1, the pier initial parameters of several sizes of piers are obtained from the bridge standard drawing set, that is, the data of typical piers are obtained, the pier initial parameter sample set is expanded by interpolation method to cover most of the pier sizes, the piers are divided into 15 different damage positions, the pier type and the damage position are combined, and the working condition of the pier damage combination is constructed.

[0076] Specifically, the bridge standard drawing set can be obtained from the industry standard drawing set or from the engineering resource library.

[0077] In this embodiment, the pier type is classified according to the structure form and divided into gravity piers, hollow piers, column piers, thin-walled piers, and flexible piers.

[0078] The pier initial parameters include the pier diameter, the column height, the pile column ratio, the damage position, and the soil parameters.

[0079] Step 2, according to the combination of the pier damage working condition, the finite element software is used to construct the pier finite element model, and the original data set for identifying the pier damage position is constructed, the original data set for identifying the pier damage position includes the pier natural frequency, the pier first-order frequency, the pier first-order frequency change rate, and the pier structure parameters under each pier damage working condition.

[0080] The pier structure parameters include pile ratio, pile diameter, area, length, and moment of inertia.

[0081] Different types and sizes of pier finite element models are cyclically constructed using ANSYS APDL finite element analysis software, the spring element constraint stiffness under each pier combination working condition is calculated, the pier first-order frequency and the pier natural frequency under different damage conditions are solved, the first-order natural frequency change rate is calculated, and the pier structure parameters, i.e., the pier pile ratio, pile diameter, area, length, and moment of inertia, are extracted, and the pier damage location identification original data set is constructed for neural network learning data features.

[0082] To simulate the column foundation constraint in the real situation, the pile is divided into multiple calculation sections, the constraint of the soil on the pile is discretized as linear spring elements on the contact surface of different calculation sections and the soil, and the stiffness coefficients of these springs are adjusted to reflect the change of the lateral pressure coefficient of the soil under different burial depths. The calculation method of the spring stiffness is as follows:

[0083] The horizontal resistance coefficient of the soil under each working condition is obtained , the horizontal action effect of the elastic pile is calculated by using the m method, and the soil constraint under different burial depths is simulated by using the spring stiffness.

[0084] The calculation formula of the soil segmented spring stiffness is:

[0085]

[0086] In the formula, represents the stiffness of the spring in the th section; represents the calculation width of the pile; represents the horizontal resistance coefficient of the soil; represents the segmented length; represents the distance from the ground to the spring in the th section.

[0087] The calculation formula of the stiffness of the top concentrated spring is:

[0088]

[0089] In the formula, represents the stiffness of the top concentrated spring; represents the area of the foundation coefficient distribution diagram within the length of the pile top spring.

[0090] The calculation of the column top constraint: the spring element is used to simulate the rubber bearing at the top of the column and the spring element is used to simulate the constraint effect of the main beam on the column. The calculation formula of the column top spring stiffness of the bridge pier is as follows: the calculation formula of the anti-push stiffness of the plate type rubber bearing, the calculation formula of the anti-push stiffness of the main beam on the bridge pier and the calculation formula of the column top spring stiffness of the bridge pier.

[0091] The calculation formula of the anti-push stiffness of the plate type rubber bearing is as follows:

[0092]

[0093] In the formula, represents the anti-push stiffness of the plate type rubber bearing; represents the shear modulus of the rubber material; represents the bearing pressure area of the bearing; represents the thickness of the rubber sheet;

[0094] The calculation formula of the anti-push stiffness of the main beam on the bridge pier is as follows:

[0095]

[0096] In the formula, represents the anti-push stiffness of the main beam on the bridge pier; represents the elastic modulus of the concrete of the main beam; represents the cross-sectional area of the main beam; 、 represents the length of the main beam on both sides of the pier top;

[0097] The calculation formula of the column top spring stiffness of the bridge pier is as follows:

[0098]

[0099] In the formula, represents the column top spring stiffness of the bridge pier.

[0100] The first-order frequency of the bridge pier (the first-order natural frequency of the bridge pier) is convenient to obtain, the damage of the bridge pier can cause the local stiffness to decrease, and the natural frequency of the bridge pier to decrease. Different damage positions have different influences on the overall stiffness of the structure, the first-order vibration mode determines the overall influence of the stiffness at different positions on the vibration characteristics, which is specifically manifested as: the greater the displacement of the damage in the first-order vibration mode, the greater the influence of the overall stiffness, and the greater the change of the frequency, and vice versa, the smaller the displacement of the damage in the first-order vibration mode, the smaller the change of the frequency. Therefore, the damage position can be deduced by the size of the frequency change rate.

[0101] The calculation formula of the first-order natural frequency change rate of the bridge pier is as follows:

[0102]

[0103] In the formula, represents the rate of change of the first-order natural frequency of the pier; represents the first-order frequency of the pier; represents the natural frequency of the pier.

[0104] Step 3, data standardization is performed on the original data set, a training set and a test set are divided, feature addition is performed on the training set and the test set, and a feature-added training set and a feature-added test set are obtained; the process is as shown in Figure 2 , including the following steps:

[0105] (1) Data loading

[0106] The pier damage location recognition data set is read, the damage location is labeled and coded, and is converted into a digital form of 0~15.

[0107] (2) Feature data standardization

[0108] Since the pier damage location recognition data set includes pier parameters and damage location recognition parameters, the unit dimensions and data scales of these data are not in the same quantity set, and different scales will cause large differences in gradient update amplitude, causing model convergence difficulties. Therefore, in order to eliminate the influence caused by the dimensional difference between data, the original data features need to be standardized.

[0109] In this embodiment, Z-score standardization is adopted to eliminate the dimensional difference between the original data through mathematical transformation, so that the data conforms to the standard normal distribution. The standardization formula is as follows:

[0110]

[0111] In the formula, represents the standardized feature data; represents the original feature data; represents the feature mean, represents the standard deviation of the feature;

[0112] This method converts the original data into standardized data with a mean of 0 and a standard deviation of 1 without changing the basic shape of the data distribution, so that the data feature dimensions are unified. This effectively avoids the problems of calculation instability and feature masking caused by extreme values, and improves the stability and accuracy of the model.

[0113] (3) Data division

[0114] The adaptive stratified sampling strategy is adopted for the unbalanced samples of different damage locations and the cases of sufficient samples and insufficient samples. Whether the data quantity of different damage locations in the (standardized) original data set is balanced is analyzed and judged. If balanced, the standardized original data set is randomly divided to obtain the training set and the test set. If unbalanced, the training set and the test set are determined by using the stratified sampling method based on the standardized original data set.

[0115] In the case of sufficient samples, the original data is divided into a training set and a test set by using the random division method, wherein the training set accounts for 80% and is used for model training, the test set accounts for 20% and is used for evaluating the performance of the model, and the minimum sample quantity is controlled to guarantee the quantity of each damage location.

[0116] In the case of insufficient samples, the standard stratified sampling method is adopted to divide the original data into a training set and a test set, so as to guarantee the original distribution ratio of the data and guarantee the balanced sample extraction of each damage location.

[0117] The adaptive stratified sampling method solves the problem of uneven sample distribution and incorrect evaluation of the recognition ability of a certain damage location.

[0118] (4) Feature Augmentation

[0119] Feature augmentation is performed on each feature (bridge pier damage location recognition feature, referred to as original feature, including pile column ratio, pile column diameter, area, length, moment of inertia, bridge pier natural frequency, bridge pier first-order frequency and bridge pier first-order frequency change rate) in the original data set.

[0120] The feature augmentation method includes calculating polynomial features, performing feature cross and calculating statistical features.

[0121] The feature engineering method is adopted to expand the original 8 bridge pier damage location recognition features into a total of 48 recognition features. The calculation of polynomial features includes calculating square terms, taking 8 groups of original feature square terms as polynomial features to capture the nonlinear relationship of the features. Feature cross is performed: 28 groups of feature combination terms are created as cross features to capture the synergistic effect between the features. The calculation of statistical features includes calculating the mean, standard deviation, maximum value and minimum value. The 4 groups of features are taken as statistical features to represent the overall structural state of the features. Through the feature enhancement operation, the recognition ability of the model for the damage location is significantly improved.

[0122] Step 4, a multilayer perception neural network model is constructed, which includes a Dropout regularization module for training the model, a weight decay regularization module for training the model, a batch normalization module for training the model and a cosine annealing learning rate scheduler for training the model.

[0123] In this step, a multi-layer perceptron neural network (MLP) is constructed to prevent overfitting and improve generalization ability through Dropout + Weight Decay double regularization. A cosine annealing learning rate scheduler is added to dynamically adjust the learning rate and improve convergence. An adaptive batch normalization module is used to balance the training speed and model performance.

[0124] First, load the data and convert the NumPy array to a PyTorch tensor, enable GPU acceleration, apply GPU parallel operation capability, and improve neural network calculation speed.

[0125] Initialize the model, initialize the weights, optimizer, and learning rate scheduler. Use Dropout regularization in the neural network to randomly drop neurons during training to prevent neuron co-adaptation and improve generalization ability. The formula is as follows:

[0126]

[0127] In the formula, represents the original output value of the neuron; represents the output value after Dropout; represents the Dropout probability; represents the scaling factor;

[0128] Use weight decay regularization to penalize large weight values to prevent overfitting. The formula is as follows:

[0129]

[0130] In the formula, represents the total loss; represents the original cross-entropy loss; represents the number of trainable weight parameters; represents the total number of trainable weight parameters, represents the regularization intensity coefficient; represents the th trainable weight parameter of the model;

[0131] Use Batch Norm batch normalization to solve the internal covariate shift problem, allowing the model to use a higher learning rate. The formula is as follows:

[0132]

[0133]

[0134] In the formula, represents the normalized , The raw input to the presentation layer; This represents the average value of the current batch. Indicates the variance of the current batch; Represents the stable terms of a numerical value; This represents the learnable scaling parameter. This represents the learnable offset parameter; This represents the standardized output.

[0135] Forward propagation transforms data features into probabilities of lesion locations. This includes performing a linear transformation, using the ReLU activation function to preserve positively correlated features, and using the Softmax function to output the probability distribution of lesion locations. The formulas for the linear transformation and activation function are shown below. The formulas for the Softmax function and the Softmax formula are as follows:

[0136] Linear transformation:

[0137]

[0138] In the formula, Indicates the first The output of the layer linear transformation; This represents the output of the upper-layer neurons after activation; Indicates the first The weight matrix of the layer linear transformation; Indicates the first Bias vector of the layer linear transformation;

[0139] Activation function :

[0140] ;

[0141] Softmax function:

[0142]

[0143]

[0144] In the formula, A number indicating the location of the damage; It also indicates the number of the damaged location; Indicates the total number of damage locations; Indicates the sample number; The model predicts the first... The sample belongs to the first The probability of each damage location; Indicates the first The sample belongs to the first The raw scores for each damage location are the output of the last linear layer of the neural network. indicates the original score of the th sample belonging to the th damage location, which is the output of the last linear layer of the neural network.

[0145] The loss is calculated by forward propagation of data, and the loss function is cross-entropy loss, the formula is as follows:

[0146]

[0147] In the formula, indicates the number of samples, indicates the number of samples in the batch; indicates an indicator function, if the true damage location of the th sample is the th damage location, it is 1, otherwise it is 0;

[0148] The loss is back-propagated, the gradient is calculated by chain rule, the parameters are updated, and the damage identification ability is optimized. The gradient descent optimization formula is as follows:

[0149]

[0150] In the formula, indicates the current iteration, indicates the next iteration; indicates the weight at the th iteration; indicates the weight at the th iteration after updating; indicates the learning rate; indicates the gradient of the loss function;

[0151] In the formula, the learning rate The dynamic learning rate scheduling is adopted, the Adam optimizer adapts to the update requirements of different features, the cosine annealing avoids falling into local optimal solution, the learning rate is reduced when the validation loss stagnates, the oscillation is prevented, and the risk of overfitting is reduced.

[0152] Step 5, using the training set with increased features and the test set with increased features to train a plurality of the multilayer perceptron neural network models, the output of the multilayer perceptron neural network model is a probability prediction distribution, and the trained multilayer perceptron neural network model is averaged to obtain a pier damage location prediction model.

[0153] It can be understood that the input of the pier damage location prediction model is the pier inherent frequency, the pier first-order frequency, the pier first-order frequency change rate and the pier structure parameters of the pier to be predicted, and the output is a damage location probability prediction distribution.

[0154] Based on the training set after feature increase, random search method is used to randomly sample in the defined hyperparameter space, and multiple rounds of tests are performed to obtain the optimal hyperparameter combination. Using the best hyperparameters, combined with different initializations, multiple neural network models are trained and integrated to obtain the best prediction model, which is used for accurate identification of the bridge pier damage location. The process of step 5 is shown in Figure 3 , including the following steps:

[0155] (1) Hyperparameter space definition

[0156] The hyperparameter space is defined according to the experience parameter range. The hyperparameters in this embodiment include hidden layer structure, Dropout rate, learning rate, weight decay, and batch normalization specific parameter range and optimization effect as shown in Table 1:

[0157] Table 1

[0158]

[0159] (2) Finding the optimal hyperparameter combination

[0160] Random sampling (randomly generating hyperparameter combinations) is performed in the self-defined hyperparameter space, the model is trained, the accuracy is evaluated using the test set, and the accuracy improvement index is used as the judgment standard (whether the accuracy is greater than the preset best accuracy). After 20 rounds of tests, the optimal hyperparameter combination is found, and the model parameter is updated using the optimal hyperparameter combination.

[0161] (3) Training multiple models

[0162] This embodiment uses the optimal hyperparameter combination to train multiple neural network models. Some of the best hyperparameters are: the first hidden layer neuron number is 256, the second hidden layer neuron number is 128, the Dropout rate is 0.228, the weight decay is 0.000602, the batch normalization is enabled, and the batch size is 128.

[0163] (4) Model integration

[0164] Multiple neural network models are averaged and integrated to obtain the best prediction model, which is used as the bridge pier damage location prediction model. This model is used for accurate identification of the bridge pier damage location. The model integration formula is as follows: the prediction probability distribution formula of a single neural network model and the prediction probability distribution formula of the integrated model.

[0165] Prediction probability distribution of a single neural network model:

[0166]

[0167] In the formula, represents the input of the neural network model, denotes the number of neural network models, denotes the th neural network model predicts the probability prediction distribution of the bridge pier damage location; denotes the th neural network model predicts the probability of the bridge pier damage location belonging to the th damage location.

[0168] The prediction probability distribution of the integrated model:

[0169]

[0170] In the formula, denotes the total number of neural network models; denotes the probability prediction distribution of the integrated model (bridge pier damage location prediction model) for the input .

[0171] Step 6, identifying the damage location of the bridge pier using the bridge pier damage location prediction model.

[0172] The model outputs a probability prediction distribution, and usually takes the maximum value of the damage location probability prediction distribution as the final prediction result.

[0173] This embodiment stores the loss and learning rate of each training step, as shown in Figure 4 and Figure 5 .

[0174] Further, after step 5 and before step 6, the following steps are further included:

[0175] The bridge pier damage location prediction model is evaluated and verified using evaluation indicators, including accuracy, recall rate, and F1 score. The confusion matrix and multi-class ROC curve are used to visualize the comparison between the predicted value and the true value. The evaluation indicator calculation formula is as follows: the accuracy calculation formula, the recall rate calculation formula, and the F1 score calculation formula.

[0176] Accuracy:

[0177]

[0178] In the formula, denotes the proportion of actual positive samples to predicted positive samples, which measures the accuracy of the model in predicting positive samples; denotes the number of correctly predicted samples; denotes the number of incorrectly predicted samples;

[0179] Recall rate:

[0180]

[0181] wherein, represents the proportion of correctly predicted positive samples to actual positive samples, measuring the ability of the model to find all positive samples;

[0182] F1 score:

[0183]

[0184] wherein, represents the F1 score, which is the harmonic mean of precision and recall;

[0185] The evaluation indexes of the prediction results of each damage position in this embodiment are shown in Table 2, and the accuracy of each damage position is shown in Figure 6 Figure 6 The overall accuracy is also shown to be 0.8962.

[0186] Table 2

[0187]

[0188] In addition, the confusion matrix of the prediction results, and the ROC curve and AUC comparison chart of each damage position are shown in Figure 7 and Figure 8 Position 1 in the figure represents the first damage position, and the meanings of other position-related expressions in the figure can be understood in the same way.

[0189] Referring to Figure 9 , the disclosure provides a bridge pier damage position recognition system based on deep learning, comprising:

[0190] The division and combination module is used to obtain the initial parameters of the bridge pier from the standard bridge drawing set, to divide a plurality of bridge pier damage positions, and to combine the bridge pier type and the bridge pier damage position to obtain the combination of the bridge pier damage working condition.

[0191] The data set construction module is used to construct a bridge pier finite element model through finite element software according to the combination of the bridge pier damage working condition, to construct an original data set for bridge pier damage position recognition, and the original data set includes the bridge pier natural frequency, the bridge pier first-order frequency, the bridge pier first-order frequency change rate, and the bridge pier structure parameters under each bridge pier damage working condition.

[0192] The data set processing module is used to perform data standardization on the original data set, to divide a training set and a test set, and to increase the features of the training set and the test set to obtain the training set after feature increase and the test set after feature increase.

[0193] ​a model construction module, configured to construct a plurality of multilayer perceptron neural network models, wherein the multilayer perceptron neural network model comprises a Dropout regularization module, a weight decay regularization module, a batch normalization module, and a cosine annealing learning rate scheduler for training the model;

[0194] a training module, configured to train the plurality of multilayer perceptron neural network models, wherein the output of the multilayer perceptron neural network model is a probability prediction distribution, and the trained multilayer perceptron neural network models are averaged to obtain a pier damage location prediction model;

[0195] a prediction module, configured to identify the damage location of the pier by using the pier damage location prediction model.

[0196] In the embodiment, the data set construction module is configured to obtain the horizontal resistance coefficient of the soil body under each working condition, and the m method is used to calculate the horizontal action effect of the elastic pile, and a spring stiffness is used to simulate the constraint of the soil body at different depths.

[0197] In the embodiment, the training of the plurality of multilayer perceptron neural network models specifically comprises defining a hyperparameter space, wherein the hyperparameters include a hidden layer structure, a Dropout rate, a learning rate, a weight decay, and batch normalization; a random search method is used to randomly sample in the hyperparameter space, and a plurality of rounds of tests are performed to obtain an optimal hyperparameter combination; and the optimal hyperparameter combination is used in combination with different initializations to train a plurality of neural network models.

[0198] In the embodiment, the system further comprises an evaluation module configured to evaluate and verify the pier damage location prediction model by using evaluation indexes, wherein the evaluation indexes include an accuracy, a recall rate, and an F1 score.

[0199] The pier damage location identification system based on deep learning can be implemented by referring to any one of the pier damage location identification methods based on deep learning in the above embodiments, and the specific implementation steps will not be described again.

[0200] The disclosure also provides a computer-readable storage medium comprising instructions that, when executed on a computer, cause the computer to perform the steps of the pier damage location identification method based on deep learning described in any one of the embodiments.

[0201] The bridge pier damage position recognition method, system and medium based on deep learning of the present disclosure, by obtaining the initial parameters of the bridge pier from the bridge standard drawing set, dividing a plurality of bridge pier damage positions, combining the bridge pier type and the bridge pier damage position to obtain the combination of the bridge pier damage working condition, and obtaining the original data set of the bridge pier damage position recognition based on the finite element software, and adding features to the original data set, compared with the prior art, the original data set is more easily obtained, and the original data set covers a wider range, and the features of the data set for training are more comprehensive; based on the training set after feature addition and the test set after feature addition, a bridge pier damage position prediction model with accurate and stable bridge pier damage position recognition is obtained through model construction and average integration, and then the bridge pier damage position prediction is more accurate and stable.

[0202] The present disclosure adopts deep learning, finite element analysis and other technologies, and combines various data processing and model optimization methods, which can effectively locate the bridge pier damage position and improve the positioning accuracy, and provides reliable technical support for intelligent operation and maintenance of bridge infrastructure. Compared with the traditional bridge pier damage positioning method, the present disclosure has the advantages of safety, practicality, economy and the like, and has a broad application prospect.

[0203] The technical features of the above-described embodiments can be combined arbitrarily, and to make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present disclosure.

[0204] The above-described embodiments only express several implementation manners of the present disclosure, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled persons in the art, without departing from the concept of the present disclosure, a number of modifications and improvements can be made, which are all within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure patent should be subject to the appended claims.

Claims

1. A method for identifying a damage location of a pier based on deep learning, characterized in that, The method comprises the following steps: Step 1, obtaining initial parameters of a bridge pier from a bridge standard drawing set, dividing a plurality of damage positions of the bridge pier, and combining a type of the bridge pier and the damage positions of the bridge pier to obtain a combination of damage working conditions of the bridge pier; Step 2, constructing a bridge pier finite element model according to the combination of the damage working conditions of the bridge pier by using a finite element software, and constructing an original data set for identifying the damage positions of the bridge pier, wherein the original data set comprises a natural frequency of the bridge pier, a first-order frequency of the bridge pier, a first-order frequency variation rate of the bridge pier, and a structure parameter of the bridge pier under each damage working condition of the bridge pier; Step 3, performing data standardization on the original data set, dividing a training set and a test set, and increasing features of the training set and the test set to obtain a training set after feature increase and a test set after feature increase; Step 4, constructing a multi-layer perception neural network model, wherein the multi-layer perception neural network model comprises a Dropout regularization module, a weight decay regularization module, a batch normalization module, and a cosine annealing learning rate scheduler for training the model, the Dropout regularization is used to randomly discard neurons during training in the neural network, and the weight decay regularization is used to punish large weight values; Step 5, training a plurality of the multi-layer perception neural network models, wherein an output of the multi-layer perception neural network model is a probability prediction distribution, the plurality of trained multi-layer perception neural network models are averaged to obtain a bridge pier damage position prediction model, and a prediction probability distribution of the integrated model is as follows: In the formula, denotes the probability prediction distribution of the bridge pier damage location prediction model for the input denotes the total number of neural network models, denotes the number of neural network models, denotes the prediction probability distribution of a single neural network model;​ Step 6, identifying the damage position of the bridge pier by using the bridge pier damage position prediction model. 2.The bridge pier damage location recognition method based on deep learning according to claim 1, characterized in that, The initial parameters of the bridge pier comprise a diameter of the bridge pier, a column height, a pile-column ratio, a damage position, and a soil parameter, the initial parameters of the bridge pier are obtained from the bridge standard drawing set to obtain an initial parameter sample set of the bridge pier, and the initial parameter sample set of the bridge pier is expanded by using an interpolation method. 3.The bridge pier damage location recognition method based on deep learning according to claim 1, characterized in that, The step 2 comprises: obtaining a horizontal resistance coefficient of the ground soil under each working condition, calculating a horizontal effect of the elastic pile by using an m method, and simulating constraints of the soil at different depths by using a spring stiffness; a segmented spring stiffness calculation formula of the soil is as follows: ； wherein denotes the first segment spring stiffness; denotes the calculated width of the pile; denotes the horizontal resistance coefficient of the ground soil; denotes the segment length; denotes the distance from the ground to the first segment spring; a stiffness calculation formula of a top concentrated spring is as follows: ； wherein represents the top concentrated spring stiffness; represents the pile top spring area of the foundation coefficient profile over the length. a bridge pier column top constraint is calculated by using a spring unit to simulate a constraint action of a rubber bearing at a top of the pile and a main beam on the column, and a spring stiffness calculation formula of the bridge pier column top is as follows: ； ； ； wherein, Kp represents the spring stiffness at the top of the pier column; Kp represents the lateral stiffness of the plate rubber bearing; G represents the shear modulus of the rubber material; A represents the bearing pressure area; t represents the thickness of the rubber sheet; Kp represents the lateral stiffness of the main girder to the pier; Ec represents the elastic modulus of the main girder concrete; A represents the main girder cross-sectional area; , L represents the length of the main girder on both sides of the pier top; a first-order natural frequency variation rate calculation formula of the bridge pier is as follows: ； In the formula, represents the rate of change of the first-order natural vibration frequency of the pier; represents the first-order frequency of the pier; represents the natural frequency of the pier. 4.The method of claim 1, wherein, The structure parameter of the bridge pier comprises a pile-column ratio, a pile-column diameter, an area, a length, and a moment of inertia. 5.The method of claim 1, wherein, The dividing of the training set and the test set specifically comprises: analyzing whether data quantities of different damage positions in the original data set are balanced, if balanced, randomly dividing the standardized original data set to obtain the training set and the test set, and if unbalanced, determining the training set and the test set by using a stratified sampling method based on the standardized original data set. 6.The method of claim 1, wherein, The feature increase method comprises: calculating polynomial features, performing feature cross, or calculating statistical features. 7.The method of claim 1, wherein, The training of the plurality of multi-layer perception neural network models comprises: defining a hyperparameter space, the hyperparameters comprising a hidden layer structure, a Dropout rate, a learning rate, a weight decay, and batch normalization; randomly sampling in the hyperparameter space by using a random search method, performing multiple rounds of tests to obtain an optimal hyperparameter combination; and training a plurality of neural network models using the optimal hyperparameter combination and different initializations. 8.The method of claim 1, wherein, The method further comprises evaluating and verifying the pier damage location prediction model by using evaluation indexes, the evaluation indexes comprising an accuracy, a recall rate, and an F1 score. 9.A bridge pier damage location recognition system based on deep learning, characterized in that, The method comprises: a division and combination module configured to obtain initial pier parameters from a bridge standard drawing set, divide a plurality of pier damage locations, and combine pier types and the pier damage locations to obtain a combination of pier damage working conditions; a data set construction module configured to construct a pier finite element model by using finite element software according to the combination of the pier damage working conditions, and construct an original data set for pier damage location identification, the original data set comprising pier natural frequencies, pier first-order frequencies, pier first-order frequency change rates, and pier structure parameters under each pier damage working condition; a data set processing module configured to perform data standardization on the original data set, divide a training set and a test set, and increase features of the training set and the test set to obtain a feature-increased training set and a feature-increased test set; a model construction module configured to construct a multi-layer perception neural network model, the multi-layer perception neural network model comprising a Dropout regularization module, a weight decay regularization module, a batch normalization module, and a cosine annealing learning rate scheduler, the Dropout regularization module being configured to randomly discard neurons during training in the neural network, and the weight decay regularization module being configured to penalize large weight values; a training module configured to train a plurality of the multi-layer perception neural network models, the outputs of the multi-layer perception neural network models being probability prediction distributions, the training module being configured to average the plurality of trained multi-layer perception neural network models to obtain a pier damage location prediction model, and the prediction probability distribution of the integrated model being: In the formula, denotes the probability prediction distribution of the pier damage location prediction model for the input denotes the total number of neural network models, denotes the number of neural network models, denotes the prediction probability distribution of a single neural network model;​ a prediction module configured to identify a damage location of a pier by using the pier damage location prediction model.

10. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the deep learning-based pier damage location identification method according to any one of claims 1 to 8.

Images