Bridge deck continuous structure concrete damage location prediction method, system, and medium

Abstract

The present disclosure relates to a bridge deck continuous structure concrete damage position prediction method, system and medium, belonging to the field of intelligent prediction of bridge damage, the bridge deck continuous structure adopts self-sensing concrete material, the method comprises: establishing a force-electricity coupling finite element model to simulate the bridge deck continuous structure concrete damage to obtain a data set comprising bridge geometric size parameters, and current rate of change data, damage initial position data, damage termination position data of the bridge deck continuous structure; a stratified sampling method is used to divide the training set and the validation set; a multi-task ordered classification model with Coral loss is trained, and an early stop mechanism is used to select the best multi-task ordered classification model to obtain a prediction model of the initial position and the termination position of the bridge deck continuous structure concrete damage; and the prediction model is used to predict the initial position and the termination position of the bridge deck continuous structure concrete damage. The present disclosure can achieve timely and accurate prediction without laying sensors.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent prediction of bridge damage, and in particular to methods, systems and media for predicting the location of damage in continuous structural concrete of bridge decks. Background Technology

[0002] To eliminate or reduce bridge deck expansion joints and improve driving comfort and safety, continuous bridge deck structures have become a widely used structural form in small and medium-span bridges. However, in actual bridge operation, although the bridge deck is connected as a whole, the substructure remains disconnected at the pier tops. Under vehicle loads, the continuous bridge deck structure, as the main tension member, bears enormous negative bending moments, making it prone to transverse cracking of the pavement layer. With water infiltration and continuous vehicle impact, the concrete of the continuous structure cracks and peels off, even exposing the internal reinforcing steel, seriously affecting driving safety.

[0003] Current damage detection technologies primarily rely on manual visual inspection and sensor detection. Manual visual inspection can only detect damage when it reaches a stage visible to the human eye, exhibiting a significant time lag. Furthermore, it requires highly experienced personnel and is subject to greater subjectivity. Sensor detection can overcome this issue; however, sensor detection requires point-based sensor deployment (typically strain, accelerometer, or displacement sensors), reflecting only localized responses at the deployed locations and creating blind spots in areas without sensors. To ensure accurate damage prediction, a large number of sensors need to be densely deployed, often pre-embedded. These sensors are prone to aging and drift in the environment, requiring regular maintenance and calibration; otherwise, the accuracy of damage location prediction is compromised. Therefore, sensor detection methods suffer from high workload, high cost, and high maintenance costs, resulting in poor engineering applicability.

[0004] Therefore, in order to avoid the problems of the aforementioned sensor detection methods, it is urgent to develop a new method for predicting the location of damage to the continuous concrete structure of bridge decks that does not require the deployment of sensors. Summary of the Invention

[0005] Therefore, it is necessary to provide methods, systems, and media for predicting the location of damage in continuous concrete structures on bridge decks to address the aforementioned issues.

[0006] To solve the above problems, the present disclosure adopts the following technical solution: In a first aspect, this disclosure provides a method for predicting the location of concrete damage in a continuous bridge deck structure, wherein the continuous bridge deck structure uses self-sensing concrete material, and the prediction method includes: Obtain structural parameters of the superstructure and continuous bridge deck construction of bridges with different geometric dimensions from bridge design drawings; Based on the structural parameters, a force-electric coupling finite element model of the entire bridge, including the continuous structure of the bridge deck, is established, and a dataset is obtained by simulating the concrete damage of the continuous structure of the bridge deck. The dataset includes the bridge's geometric parameters, as well as the current change rate data, damage initiation location data, and damage termination location data of the continuous structure of the bridge deck. Based on the dataset, a stratified sampling method is used to divide the training set and the validation set; A multi-task ordered classification model using Coral loss is trained on the training set, and an early stopping mechanism is used on the validation set to select the optimal multi-task ordered classification model, resulting in a prediction model for the initial and final damage locations of the continuous concrete structure on the bridge deck. The multi-task ordered classification model includes a shared feature extraction layer, a shared fully connected layer, a first output head, and a second output head. The shared feature extraction layer extracts features from the input data and outputs a multi-dimensional vector. The shared fully connected layer performs dimensionality reduction and compression on the multi-dimensional vector and implements nonlinear feature transformation using the ReLU activation function. A random deactivation regularization method is used to suppress overfitting and output a compressed nonlinear damage feature vector. The first output head predicts the initial damage location based on the compressed nonlinear damage feature vector, and the output head predicts the final damage location based on the compressed nonlinear damage feature vector. The current change rate data of the continuous structural concrete of the bridge deck to be damaged and the corresponding bridge geometric parameters are obtained. The prediction model is then used to predict the initial and final damage locations of the continuous structural concrete of the bridge deck.

[0007] In a preferred embodiment, the step of dividing the training set and validation set according to the dataset using stratified sampling specifically involves dividing the dataset into a training set, a validation set, and a test set using stratified sampling. The prediction method also includes the step of evaluating the prediction model using the test set.

[0008] In a preferred embodiment, the step of dividing the dataset into a training set, a validation set, and a test set using stratified sampling specifically includes: Each initial damage location data and its corresponding initial damage location data in the dataset are fused into a composite label. Based on the dataset with the composite label, a stratified sampling method is used to divide it into a training set, a validation set, and a test set.

[0009] In a preferred embodiment, dividing the dataset into training, validation, and test sets using stratified sampling specifically includes: For infinity values ​​in the training, validation, and test sets, convert the infinity values ​​into missing values ​​and fill in all missing values; Outliers in the training, validation, and test sets are handled using the quantile Windsor truncation method. Z-score standardization is performed on the data in the training, validation, and test sets.

[0010] In a preferred embodiment, dividing the dataset into training, validation, and test sets using stratified sampling specifically includes: generating synthetic samples from the training set using synthetic minority class oversampling, wherein the synthetic samples serve as data in the training set.

[0011] In a preferred embodiment, the loss function of the multi-task ordered classification model is: in, The loss value for the Coral loss function for ordered classification; The number of samples in a batch; This is a sample index used to identify the first sample in a batch. One sample; The total number of categories in an ordered classification; The number of binary classification boundaries that need to be predicted for each sample; The threshold index represents the threshold value in an ordered classification task. A binary classification text; For the first The sample at the th The true label on the binary sub-problem; It is the sigmoid activation function; For the model to the first The sample at the th Logit output on a binary subproblem.

[0012] In a preferred embodiment, the shared feature extraction layer includes a first convolutional block, a second convolutional block, a third convolutional block, and an average pooling layer arranged sequentially. The first and second convolutional blocks are used to perform batch normalization, ReLU activation, max pooling, and random deactivation on the data in sequence. The third convolutional block is a residual block, which is used to first increase the number of channels by using two convolutional layers, and then achieve identity mapping through skip connections. The average pooling layer is used to compress features of arbitrary length into a fixed dimension using adaptive average pooling and then perform random deactivation.

[0013] In a preferred embodiment, the parameter updates for training the multi-task ordered classification model using Coral loss are implemented using the AdamW optimizer, as shown in the following formula: in, These are the updated parameter values; These are the parameter values ​​before the update; Represented as the learning rate; This is expressed as the bias correction estimate of the first moment of the gradient; This is expressed as the bias correction estimate of the second moment of the gradient; This is expressed as the weight decay coefficient; To prevent division by zero of extremely small constants; The learning rate scheduler for the multi-task ordered classification model trained using Coral loss employs cosine annealing, causing the learning rate to periodically decay from its initial value to near zero according to a cosine function, as shown in the following formula: in, Indicates the first Learning rate at each round; This represents the minimum learning rate; This is represented as the initial maximum learning rate; Indicates the number of rounds currently performed; The total number of cycles in a complete cosine cycle.

[0014] Secondly, this disclosure provides a system for predicting the location of concrete damage in a continuous bridge deck structure, wherein the continuous bridge deck structure uses self-sensing concrete material, and the prediction system includes: The structural parameter acquisition module is used to obtain the structural parameters of the superstructure and continuous bridge deck structure of bridges with different geometric dimensions from bridge design drawings. The finite element model building module is used to build a force-electric coupling finite element model of the entire bridge, including the continuous structure of the bridge deck, based on the structural parameters. The dataset acquisition module uses the force-electric coupling finite element model to simulate the damage of the continuous concrete structure of the bridge deck to obtain a dataset. The dataset includes bridge geometric parameters, as well as current change rate data, damage initiation location data, and damage termination location data of the continuous structure of the bridge deck. The partitioning module is used to divide the dataset into a training set and a validation set using stratified sampling. The training module is used to train a multi-task ordered classification model using Coral loss on the training set, and to select the best multi-task ordered classification model on the validation set using an early stopping mechanism, so as to obtain the prediction model of the initial and termination positions of damage in the continuous concrete structure of the bridge deck. The multi-task ordered classification model includes a shared feature extraction layer, a shared fully connected layer, a first output head, and a second output head. The shared feature extraction layer is used to extract features from the input data and output a multi-dimensional vector. The shared fully connected layer is used to reduce the dimensionality of the multi-dimensional vector and to perform nonlinear feature transformation through the ReLU activation function. The random deactivation regularization method is used to suppress overfitting and output a compressed nonlinear damage feature vector. The first output head is used to predict the initial position of damage based on the compressed nonlinear damage feature vector, and the output head for the termination position of damage is used to predict the termination position of damage based on the compressed nonlinear damage feature vector. The prediction module is used to acquire the current change rate data of the continuous structural concrete of the bridge deck and the corresponding bridge geometric parameters for damage location prediction, and to use the prediction model to predict the initial and final damage locations of the continuous structural concrete of the bridge deck.

[0015] Thirdly, this disclosure provides a storage medium storing a computer program, which, when executed by a processor, implements the method for predicting the location of damage to continuous structural concrete on bridge decks as described in the first aspect.

[0016] The aforementioned method, system, and medium for predicting the location of damage in continuous concrete structures on bridge decks do not require the deployment of sensors. The design employs a multi-task ordered classification model with Coral loss and obtains a prediction model for the initial and final locations of damage in continuous concrete structures on bridge decks based on a dataset acquired through a force-electric coupling finite element model. Based on the prediction model, current change rate data, and corresponding bridge geometric parameters, accurate identification and prediction of damage locations in continuous concrete structures on bridge decks can be achieved. The identification can be completed when the damage is in the microcrack stage, and the damage location prediction has good timeliness and high sensitivity. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for predicting the location of damage to continuous concrete structures on bridge decks in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the process of obtaining a dataset using finite element simulation technology according to an embodiment of the present invention; Figure 3 This is a training process loss curve diagram according to an embodiment of the present invention; Figure 4 This is a graph showing the accuracy variation of an embodiment of the present invention; Figure 5 This is a heatmap of the confusion matrix for predicting the initial damage location in the test set according to an embodiment of the present invention. Figure 6 This is a heatmap of the confusion matrix for predicting the damage termination location in the test set according to an embodiment of the present invention. Figure 7 This is a schematic diagram of the bridge deck continuous concrete damage location prediction system in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0019] See Figure 1 This disclosure provides a method for predicting the location of damage in a continuous bridge deck concrete structure, wherein the continuous bridge deck structure uses self-sensing concrete material, and the method for predicting the location of damage includes: Obtain structural parameters of the superstructure and continuous bridge deck construction of bridges with different geometric dimensions from bridge design drawings; A force-electric coupling finite element model of the entire bridge was established, and a dataset was obtained by simulating the damage of the continuous concrete structure of the bridge deck. The dataset includes the bridge's geometric parameters, as well as the current change rate data, damage initiation location data, and damage termination location data of the continuous structure of the bridge deck. Based on the dataset, a stratified sampling method is used to divide the training set and the validation set; A multi-task ordered classification model using Coral loss is trained on the training set, and an early stopping mechanism is used on the validation set to select the optimal multi-task ordered classification model, resulting in a prediction model for the initial and final damage locations of the continuous concrete structure on the bridge deck. The multi-task ordered classification model includes a shared feature extraction layer, a shared fully connected layer, a first output head, and a second output head. The shared feature extraction layer is used to extract features from the input data and output a multi-dimensional vector. The shared fully connected layer is used to reduce the dimensionality of the multi-dimensional vector and perform nonlinear feature transformation using the ReLU activation function. A random deactivation regularization method is used to suppress overfitting and output a compressed nonlinear damage feature vector. The first output head is used to predict the initial damage location based on the compressed nonlinear damage feature vector, and the second output head is used to predict the final damage location based on the compressed nonlinear damage feature vector. The current change rate data of the continuous structural concrete of the bridge deck to be damaged and the corresponding bridge geometric parameters are obtained. The prediction model is then used to predict the initial and final damage locations of the continuous structural concrete of the bridge deck.

[0020] The method will be described in detail below.

[0021] The continuous bridge deck concrete is self-sensing concrete, meaning the continuous bridge deck structure uses self-sensing concrete material, which forms the basis for the method described above. The method includes the following steps: Step 1: Extract the structural parameters of the bridge based on the bridge design drawings.

[0022] In this embodiment, structural parameters of the superstructure and continuous bridge deck with different geometric dimensions (i.e., multiple geometric dimensions) are obtained from several bridge design drawings, providing a sufficient data foundation for subsequent finite element simulation and deep learning models.

[0023] For bridges with different geometric dimensions, the structural parameters of both the superstructure and the continuous structure are obtained from the bridge design drawings. Specifically, these structural parameters refer to the structural parameters of the superstructure and the continuous structure of the bridge deck. The structural parameters of the superstructure include beam height, beam length, beam width, top plate thickness, bottom plate thickness, web thickness, chamfer dimensions, flange dimensions, and hinge joint width. The structural parameters of the continuous structure include the length of the continuous structure, the length of the debonding zone, the length of the anchorage zone, and the thickness of the pavement.

[0024] Step 2: Establish a force-electric coupling finite element model of the entire bridge based on the structural parameters. The entire bridge includes the bridge structure and the continuous bridge deck structure. Use the force-electric coupling finite element model to simulate the concrete damage of the continuous bridge deck structure to obtain a dataset. The dataset includes the bridge geometric parameters, the current change rate data of the continuous bridge deck structure, the damage initiation location data of the continuous bridge deck structure, and the damage termination location data of the continuous bridge deck structure. The bridge's geometric dimensions include beam length, beam height, and beam width.

[0025] Based on the structural parameters of the bridge, finite element simulation was carried out. By introducing the piezoresistive matrix through the electro-mechanical coupling field mechanism of self-sensing materials, a full bridge finite element model including the superstructure and the continuous structure of the bridge deck was established. The finite element model is a force-electric coupling finite element model. The current response characteristics of the continuous structure concrete of the bridge deck were simulated when damage occurred. Then, the current change rate of the continuous structure concrete of the bridge deck was obtained, and a dataset was obtained.

[0026] This embodiment employs finite element modeling, using Solid65 elements to simulate ordinary concrete and Solid226 elements to simulate the self-sensing concrete of the continuous bridge deck structure. A force-electric coupling finite element model is established, taking into account both structural mechanical response and electrical signal response. Establishing a force-electric coupling finite element model requires introducing key electrical parameters such as resistivity and piezoresistive matrix on top of the mechanical parameters. At least these two electrical parameters must be introduced. Resistivity determines the current density vector field distribution and potential gradient field distribution within the model after applying an electric potential. The piezoresistive matrix, acting as the bridge, organically combines mechanics and electricity, specifically manifested as the resistivity change caused by strain in the conductive network within the concrete. Based on the structural parameters obtained in step 1, a finite element model is established that is compatible with bridges of different geometric dimensions for both the bridge superstructure and the continuous bridge deck structure (i.e., the finite element model includes both the bridge superstructure and the continuous bridge deck structure). To ensure the reliability of continuity between the two, overlapping nodes at the interface are merged, integrating the bridge and the continuous bridge deck structure into a unified finite element model, achieving continuity of the displacement field and electric potential field at the interface. In other words, the bridge superstructure and the continuous bridge deck structure form a multi-scale mesh nesting in the numerical model. The coarse mesh of the bridge superstructure provides the macroscopic geometric boundary for the local continuous structure, while the fine mesh of the continuous bridge deck structure is a local magnification and refined characterization of the key nodes of the bridge superstructure.

[0027] In this embodiment, during the application of potential load, an alternating potential excitation method is used on the continuous bridge deck structure. This involves alternately applying positive and negative potentials between adjacent finite element elements, creating a uniform electric field within the continuous bridge deck structure. Essentially, this method simulates the continuous bridge deck structure as several continuously arranged sensor units (or sensing units), constructing a distributed sensing array. First, the current density corresponding to each sensor unit in the undamaged state is extracted as a reference parameter. When the structure is damaged, the rate of change of the current density corresponding to each sensor unit relative to the reference parameter is calculated. Utilizing the sensitivity of the electric field to concrete damage, the rate of change of current is used as a damage characteristic, thereby achieving effective identification of damage at different locations and in local areas. This method embodies the core concept that the self-sensing concrete structure itself is a sensor, achieving accurate simulation of the force-electric response characteristics of the continuous bridge deck structure. The detailed process for obtaining basic data for the force-electric coupling finite element model of the entire bridge is as follows: Figure 2 As shown, the process includes: obtaining the resistivity and piezoresistive effect (i.e., piezoresistive matrix) of self-sensing concrete through experiments; establishing a force-electric coupling finite element model; pre-setting electrodes in the continuous concrete structure of the bridge deck to output the actual potential gradient; and obtaining the rate of change of current in the continuous structure of the bridge deck when it is damaged.

[0028] Understandably, as one embodiment, a single force-electric coupling finite element model can be used to accommodate bridges with different geometries. Understandably, as one embodiment, more than one force-electric coupling finite element model can also be established.

[0029] Step 3 involves partitioning the dataset obtained from the force-electric coupling finite element model of the full bridge in Step 2 to obtain a training set and a validation set. Further, a test set is also obtained through this partitioning operation. Step 3 may also include preprocessing the training set, validation set (and test set), with the preprocessed results used in subsequent steps. The preprocessing includes at least one of missing value handling, outlier handling, and minority class oversampling.

[0030] In a preferred embodiment, step 3 specifically includes: Step 3.1, Dataset partitioning: The initial location data of the continuous structural concrete damage of the bridge deck and its corresponding termination location data are merged into a composite label to obtain a dataset with the composite label. Based on the dataset with the composite label, a stratified sampling method is used to partition the dataset into a training set, a validation set and a test set.

[0031] This embodiment employs stratified sampling to partition the dataset, ensuring the reliability of the multi-task ordered classification task during the evaluation phase. A composite label is constructed, fusing the combined information of the initial damage location and the termination damage location of the continuous concrete structure of the bridge deck into a single category. In other words, two ordered classification labels are integrated into a single integer label using mathematical operations, called a composite label, used for joint stratified sampling to ensure the joint distribution of the two labels during dataset partitioning. Each initial damage location data point and its corresponding termination damage location data result in a composite label. Specifically, the initial damage location value is multiplied by a sufficiently large coefficient, and then added to the termination damage location value. This ensures that different combinations do not overlap. Assuming the initial damage location is... The location where the damage terminates is ,coefficient If we take 1000, then the composite tag for this combination is: 。 If stratification is performed solely based on the initial or final location of the damage, it may disrupt the distribution of the other label. By merging the two into a composite type, stratified sampling ensures that the proportion of each combination in the training, validation, and test sets remains consistent with the original data (the original data is the dataset obtained in step 2 / data in the dataset), thereby enabling the model to capture the internal relationship between the two and improving the reliability of the evaluation results.

[0032] Based on the integrated composite labels, the dataset was split into a first dataset and a test set at a ratio of 70% to 30%. Stratified sampling ensured that the distribution of composite categories in the test set was consistent with the original data, avoiding the possibility of missing minority class combinations in the test set due to random partitioning, which could affect the reliability of the evaluation results. Subsequently, the first dataset was split into a training set and a validation set at a ratio of 80% to 20% using the same method. Ultimately, the training set, validation set, and test set accounted for 56%, 14%, and 30% of the original data, respectively. The training set was used for model learning, while the test set was only used for model selection, early stopping, and hyperparameter tuning, without parametric gradient updates, effectively preventing overfitting. The test set was only used for evaluation after the model was fully trained, truly reflecting the model's generalization ability.

[0033] Step 3.2, missing value handling: Convert infinity values ​​to missing values ​​(NaN), and use the Simplemputer tool (a class in the Scikit-Learn library for filling missing values) based on the median strategy to fill all missing values ​​(both original missing values ​​and missing values ​​caused by outliers).

[0034] The median strategy is more resistant to outlier interference than the mean, especially for data where current values ​​are prone to extreme values ​​during damage. The imputation scheme is fitted only on the training set, while the validation and test sets still use the same scheme. This ensures the model's evaluation results on location data are realistic and reliable, preventing information leakage from affecting model accuracy. While ensuring data integrity, it also preserves the original class distribution to the greatest extent possible. This approach is applicable to common missing value problems and improves the stability and generalization ability of model training, making it well-suited for the needs of multi-task ordered classification tasks.

[0035] Step 3.3, Outlier Handling: Outlier handling is performed using the Windsor quantile truncation method. This involves using the Windsor truncation method to determine the normal fluctuation range of the data through quantiles, such as the 5th percentile and the 95th percentile. Outliers are then modified to their corresponding boundary values, which are the boundary values ​​of the normal fluctuation range.

[0036] This embodiment uses the Winsorizing quantile method for outlier handling. By reasonably determining the interval through quantiles, extreme outliers are truncated to the interval boundary rather than deleted, thus preserving the original data to the greatest extent possible. The calculation formula is as follows: Assume the original dataset is The original dataset after processing with the Windsor truncation method is then... any data Defined as: in, Data from the original dataset , Values Integers in For the original dataset The total number of data in the data, for The data obtained after processing using the Windsor truncation method For quantiles, usually In this embodiment, the value is 0.01; For the first quantiles; For the first Quantiles.

[0037] The truncation boundaries are calculated only on the training set. Using a quantile of 0.01, the 1st percentile of each feature is taken as the lower truncation boundary, and the 99th percentile as the upper truncation boundary. Sample values ​​exceeding the truncation boundary are truncated to the corresponding boundary, while values ​​within the boundary retain their original values. For the validation and test sets, the boundaries determined in the training set are used directly to avoid data leakage. This method effectively suppresses the interference of extreme noise on model training, preserves the original data distribution and the relative relationships of ordered features, and can be applied to multi-task ordered classification tasks. It also possesses robustness in automatically correcting inverted upper and lower bounds.

[0038] Step 3.4, Minority Class Oversampling: For the training set, synthetic samples are generated using the Synthetic Minority Over-sampling Technique (SMOTE), which serves as the data in the training set. In this embodiment, the minority class oversampling uses the Synthetic Minority Over-sampling Technique (SMOTE) to balance the number of composite label samples in the training set. Its core idea is to perform linear interpolation between minority class samples to generate new synthetic samples, rather than simply copying the original samples. This effectively alleviates the class imbalance problem while ensuring that the geometric structure of the feature space remains unchanged. The SMOTE oversampling formula is shown below: in, This indicates the newly generated sample; Represents the original minority class sample; Indicates from of Minority class samples randomly selected from the nearest neighbors; It is represented as a random number between [0,1], which controls the degree of interpolation; This represents the number of nearest neighbors, used to determine which samples can be used for synthesis.

[0039] By generating diverse synthetic samples through SMOTE oversampling, the number of samples across different classes can be balanced, effectively avoiding overfitting the model to the majority class and improving the accuracy of minority class identification. Synthetic samples are generated within the neighborhood of the original feature space, fully preserving the correlation between features, making it particularly suitable for physical quantities such as bridge geometric parameters and current change rates. Complementing the stratified sampling in the preprocessing stage, SMOTE oversamples only within the training set, ensuring that the validation and test sets maintain their original distribution, thus realistically evaluating the model's generalization ability. This method significantly enhances the model's robustness to imbalanced data, improves the accuracy of minority class prediction, and reduces the risk of overfitting.

[0040] Step 3.5, Data Standardization: This example uses Z-score standardization, which follows missing value handling, outlier truncation, and oversampling. Its purpose is to eliminate dimensions. Since the data features include bridge geometric parameters and current change rates, which differ by several orders of magnitude, without standardization, larger features would dominate gradient updates, ignoring smaller values ​​and increasing the difficulty of model training. Standardized data is closer to the mean and unit variance, helping the gradient descent optimizer converge faster. It also reduces numerical instability caused by feature differences, making the model more robust to small changes in input features. The Z-score standardization formula is as follows: in, This represents the standardized feature values; Represents the features in the training set; It represents the mean of a certain feature in the training set; It represents the standard deviation of a feature in the training set.

[0041] Step 4: Based on the processed training set, a multi-task ordered classification model that combines CNN and Coral loss is trained using a preset combination of hyperparameters. On the validation set, an early stopping mechanism is used to select the best model, which is then used to predict the initial and final locations of damage to the continuous concrete structure of the bridge deck.

[0042] The multi-task ordered classification model includes: a shared feature extraction layer, a shared fully connected layer, a first output head (output head for the initial damage location), and a second output head (output head for the final damage location). The shared feature extraction layer extracts features from the input data and outputs a multi-dimensional vector. The shared fully connected layer performs dimensionality reduction and compression on this multi-dimensional vector, using the ReLU activation function to achieve nonlinear feature transformation and Dropout regularization to suppress overfitting. It outputs a compressed nonlinear damage feature vector shared by both the initial and final damage location tasks. The shared fully connected layer supports knowledge transfer between the initial and final damage location learning processes. The output head for the initial damage location is used to predict the initial damage location based on the output of the shared fully connected layer. The output head for the final damage location is used to predict the final damage location based on the output of the shared fully connected layer.

[0043] The multi-task ordered classification model, a deep learning model, learns the sequential information of damage development from its initial location to its termination location. The task objective is to predict where concrete damage begins and ends. Essentially, it segments the damage area according to cross-sectional location. These segments (e.g., "no damage area → initial damage area → termination damage area → no damage area") exhibit a clear order from nothing to something and from something back to nothing, conforming to the core characteristic of ordered classification: "categories have natural levels / order." This deep learning model not only identifies the location of damage but also utilizes the principle of "orderly damage development" to output results that are more consistent with engineering realities.

[0044] This embodiment uses the data preprocessed in step 3 to model and validate the prediction model. The prediction model adopts a multi-task ordered classification model with Coral loss. The model takes ordered multi-classification tasks as the core objective, adopts a multi-task learning mode that combines shared feature extraction with task-specific outputs, uses convolutional neural networks to extract local features, and introduces residual connections to enhance gradient flow. Specifically, the high-dimensional features extracted by convolution are mapped to a 256-dimensional shared representation to promote knowledge transfer between the two tasks. Independent Coral (correlation-aligned) ordered regression heads are designed for the initial and final damage locations.

[0045] The shared feature extraction layer consists of three sequentially arranged convolutional blocks, referred to as the first convolutional block, the second convolutional block, and the third convolutional block, as well as an average pooling layer located after the third convolutional block.

[0046] The first two convolutional blocks sequentially perform batch normalization, ReLU activation, max pooling, and Dropout (random deactivation) to alleviate internal covariate bias, reduce computational cost, prevent overfitting, and enhance regularization. The third convolutional block is a residual block, employing two convolutional layers to increase the number of channels to 128, and then using skip connections to achieve identity mapping. This residual structure provides a direct path for gradients, effectively mitigating the gradient vanishing problem in deep networks, making training more stable, and shifting the optimization objective from fitting complex mappings to learning residual functions, accelerating convergence and reducing sensitivity to the learning rate. The final shared feature extraction layer (i.e., the average pooling layer) uses adaptive average pooling to compress features of arbitrary length to a fixed dimension and performs Dropout, extracting local dependencies while ensuring the model's robustness to changes in the number of input features. That is, the output of the average pooling layer is the compressed nonlinear damage feature vector.

[0047] The multi-task ordered classification model employs a fully connected design, consisting of a shared fully connected layer and task-specific output heads. The task-specific output heads include a first output head and a second output head. The shared fully connected layer compresses the multi-dimensional vector output from the feature extraction layer, using ReLU and Dropout for non-linear transformation and regularization, facilitating knowledge transfer between the two tasks. Both the first and second output heads consist of two linear layers, supplemented by ReLU and Dropout. The entire design balances parameter sharing with task independence, reducing overfitting risk while ensuring the model learns unique classification boundaries for each task. The first linear layer maps the compressed non-linear damage feature vector to a hidden feature space suitable for the current task (prediction of damage initiation / termination location). After the first linear layer outputs, ReLU introduces a non-linear transformation, enabling the model to learn complex damage feature patterns; Dropout then suppresses overfitting; finally, the second linear layer outputs the final prediction result (prediction of damage initiation / termination location), ensuring the model's feature extraction capability for the current task while controlling overfitting risk.

[0048] In this embodiment, the model employs the ordered classification Coral loss function. This loss function can utilize the natural sequence between classifications, such as the spatial order of lesion locations, and is superior to multi-class cross-entropy. It transforms the complex ordered classification into multiple binary classification problems, making training more stable. The formula is as follows: in, The loss value for the Coral loss function for ordered classification; The number of samples in a batch; This is a sample index used to identify the first sample in a batch. One sample; The total number of categories in an ordered classification; The number of binary classification boundaries that need to be predicted for each sample; The threshold index represents the threshold value in an ordered classification task. A binary classification text; For the first The sample at the th The true label on the binary sub-problem; It is the sigmoid activation function; For the model to the first The sample at the th The logit output of a binary subproblem is the result of a linear transformation without activation.

[0049] Authentic Labels The formula is as follows: in For the first The true category of each sample; This is an indicator function; it takes the value 1 when the condition within the parentheses is true, and 0 otherwise.

[0050] The AdamW optimizer is used for parameter updates, improving regularization performance and avoiding the issue of weight decay failing with learning rate scaling in traditional Adam. Furthermore, each parameter has an independent learning rate, which accelerates convergence. The formula is as follows: in, These are the updated parameter values; These are the parameter values ​​before the update; Indicates the current iteration step; Represented as the learning rate; This is expressed as the bias correction estimate of the first moment of the gradient; This is expressed as the bias correction estimate of the second moment of the gradient; This is expressed as the weight decay coefficient; To prevent division by zero of extremely small constants.

[0051] The learning rate scheduler uses cosine annealing to cause the learning rate to decay periodically from its initial value to near zero according to a cosine function, as shown in the following formula: in, Represented as the first Learning rate per epoch (round); This represents the minimum learning rate, which defaults to 0 but can also be set to a non-zero value. This is represented as the initial maximum learning rate; This indicates the number of epochs currently performed. The total number of epochs in a complete cosine period, i.e. from decay to The number of iterations required.

[0052] The AdamW optimizer, combined with the learning rate scheduler, provides efficient and stable parameter updates in the early stages of training, and enables fine-tuning in the later stages, ultimately achieving higher accuracy and lower latency on the test set. This reduces the risk of overfitting. The AdamW optimizer ensures fast and stable gradient updates, while cosine annealing ensures that the learning rate decays reasonably as training progresses. Together, they promote the model to converge to the optimal solution.

[0053] Furthermore, after obtaining the prediction model of the initial and final damage locations of the continuous concrete structure of the bridge deck in step 4, step 5 is also included. It is understood that in some embodiments, step 5 can be omitted and step 6 can be performed directly.

[0054] Step 5, evaluating the prediction model using the test set: The deep learning model is evaluated using evaluation metrics, including accuracy, precision, recall, F1 score, and mean absolute error. The evaluation metrics are calculated as follows: Accuracy: in, Accuracy, or accuracy, represents the proportion of samples that are correctly predicted out of the total number of samples. It represents the number of true positives, that is, the number of instances that the model correctly predicts as positive. It is the number of true negative instances, that is, the number of instances that the model correctly predicts as negative. It is the number of false positives, that is, the number of negative class instances that the model incorrectly predicts as positive. It represents the number of false negatives, i.e., the number of positive class instances that the model incorrectly predicts as negative.

[0055] Accuracy: in, Precision rate, which represents the proportion of samples predicted as positive that are actually positive. and Same as above.

[0056] Recall rate: in, Recall rate, also known as the positive rate, is the proportion of samples that are actually positive that are predicted to be positive. and Same as above.

[0057] F1 score: in, The F1 score, or F1 score, is the harmonic mean of precision and recall, which balances precision and recall and takes into account the impact of false positives and false negatives on model performance.

[0058] Since both the initial and final damage locations are ordered categories, the absolute difference between the predicted and true categories is of practical significance. The smaller the difference, the closer the predicted result is to the true result. Therefore, this embodiment uses the mean absolute error as a supplementary evaluation index to make full use of the order information of the order categories and to measure the average offset between the predicted and true locations in more detail. This can intuitively reflect the accuracy of the model in ordered classification tasks.

[0059] Mean absolute error: in, The units are consistent with the units in the category labels; This represents the total number of samples, i.e., the number of test set samples used in the evaluation. Indicates the first The model prediction value for each sample; in an ordered classification task, this value is the label of the model's output class. Indicates the first The true value of a sample, that is, the actual class label of the sample.

[0060] Step 6: Obtain the current change rate data and corresponding bridge geometric parameters of the continuous concrete structure of the bridge deck to be damaged (the bridge geometric parameters of the bridge corresponding to the continuous concrete structure of the bridge deck to be damaged). Use the prediction model to predict the initial and final damage locations of the continuous concrete structure of the bridge deck.

[0061] The method for obtaining the current change rate data of the continuous bridge deck concrete for damage location prediction is as follows: The continuous bridge deck structure uses self-sensing concrete material, which has its own sensing function, eliminating the need for pre-embedded sensors. Sensors are arranged at equal intervals along the transverse direction within the concrete. Each pair of electrodes, with positive and negative poles between any two adjacent electrodes, forms a sensor. The concrete between each pair of positive and negative electrodes constitutes a sensing unit. Under the excitation of an electric field, the concrete in each sensing unit can reflect changes in its internal resistance in real time. The entire bridge deck is formed by a continuous structure. Several sensors cover various sections of the continuous bridge deck structure. The rate of change of current is calculated by collecting the current response of the sensing units before and after structural damage.

[0062] Figure 3 The diagram shows the training loss curve of a multi-task ordered classification model in one embodiment, illustrating the training loss and validation loss. It demonstrates that the training process is healthy, converges well, shows no obvious signs of overfitting, and that the model's fit to the data gradually improves, indicating satisfactory generalization ability.

[0063] Figure 4 The graph shows the accuracy variation of a multi-task ordered classification model in one embodiment. The initial position is the initial position of loss, and the termination position is the termination position of loss. It can be seen that the accuracy gradually increases with the training rounds and eventually tends to stabilize and converge, indicating that the model's predictive ability continuously improves during training and eventually reaches a stable performance level, with high model prediction accuracy.

[0064] Figure 5 This is a heatmap of the confusion matrix based on the prediction of the initial location of damage from the test set in one embodiment; Figure 6 This is a heatmap of the confusion matrix based on the damage termination location prediction of a test set in one embodiment; Figure 5 and Figure 6 Each example includes 20 categories, named Category 1 through Category 20. From... Figure 5 and Figure 6 It can be seen that both types of prediction tasks have high overall accuracy and good model prediction performance.

[0065] The evaluation metrics for damage location prediction in this embodiment are shown in Table 1: Table 1 Evaluation Indicators for Damage Location Prediction

[0066] According to the evaluation metrics, the model performs significantly on both the initial damage location and the final damage location tasks. The model can predict the damage location with extremely high accuracy, and the prediction bias in ordered classification is small, less than 0.03 categories. It has strong classification performance, meets the needs of high-precision prediction in practical applications, and provides technical support for practical engineering applications.

[0067] See Figure 7 This disclosure provides a system for predicting the location of concrete damage in a continuous bridge deck structure, wherein the continuous bridge deck structure uses self-sensing concrete material, and the prediction system includes: The structural parameter acquisition module is used to obtain the structural parameters of the superstructure and continuous bridge deck structure of bridges with different geometric dimensions from bridge design drawings. The finite element model building module is used to build a force-electric coupling finite element model of the entire bridge, including the continuous structure of the bridge deck, based on the structural parameters. The dataset acquisition module uses the force-electric coupling finite element model to simulate the damage of the continuous concrete structure of the bridge deck to obtain a dataset. The dataset includes bridge geometric parameters, as well as current change rate data, damage initiation location data, and damage termination location data of the continuous structure of the bridge deck. The partitioning module is used to divide the dataset into a training set and a validation set using stratified sampling. The training module is used to train a multi-task ordered classification model using Coral loss on the training set, and to select the best multi-task ordered classification model on the validation set using an early stopping mechanism, so as to obtain the prediction model of the initial and termination positions of damage in the continuous concrete structure of the bridge deck. The multi-task ordered classification model includes a shared feature extraction layer, a shared fully connected layer, a first output head, and a second output head. The shared feature extraction layer is used to extract features from the input data and output a multi-dimensional vector. The shared fully connected layer is used to reduce the dimensionality of the multi-dimensional vector and to perform nonlinear feature transformation through the ReLU activation function. The random deactivation regularization method is used to suppress overfitting and output a compressed nonlinear damage feature vector. The first output head is used to predict the initial position of damage based on the compressed nonlinear damage feature vector, and the output head for the termination position of damage is used to predict the termination position of damage based on the compressed nonlinear damage feature vector. The prediction module is used to acquire the current change rate data of the continuous structural concrete of the bridge deck and the corresponding bridge geometric parameters for damage location prediction, and to use the prediction model to predict the initial and final damage locations of the continuous structural concrete of the bridge deck.

[0068] In this embodiment, the step of dividing the training set and validation set according to the dataset using stratified sampling specifically involves dividing the dataset into a training set, a validation set, and a test set using stratified sampling. The prediction system further includes an evaluation module, which is used to evaluate the prediction model using the test set.

[0069] In this embodiment, the partitioning module is also used to convert infinity values ​​in the training set, validation set, and test set into missing values ​​and fill in all missing values; to process outliers in the training set, validation set, and test set using the quantile Windsor truncation method; and to perform Z-score standardization on the data in the training set, validation set, and test set.

[0070] In this embodiment, the partitioning module is further configured to generate synthetic samples using synthetic minority class oversampling technology, and the synthetic samples serve as data in the training set.

[0071] It is understood that the training set and validation set involved in the training module are the final output training set and validation set of the partitioning module.

[0072] In specific implementation, the bridge deck continuous structural concrete damage location prediction system can refer to the bridge deck continuous structural concrete damage location prediction method in any of the above embodiments to predict the bridge deck continuous structural concrete damage location. The specific implementation steps will not be repeated.

[0073] An electronic device can be implemented according to the method of this disclosure, the electronic device comprising: a memory; one or more processors; one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing the method for predicting the location of damage to continuous concrete in bridge decks according to any of the above embodiments.

[0074] This disclosure also provides a storage medium storing a computer program, which, when executed by a processor, implements the various steps of the method for predicting the location of damage to continuous concrete structures on bridge decks as described in any of the above embodiments.

[0075] This disclosure presents a method, system, and medium for predicting the location of damage in continuous concrete structures on bridge decks. It achieves timely and accurate predictions without the need for sensor deployment. The disclosure employs a force-electric coupling finite element model to acquire datasets, a multi-task ordered classification model using Coral loss, and a verification method. This yields prediction models for the initial and final locations of damage in continuous concrete structures on bridge decks. Based on these prediction models, current change rate data, and corresponding bridge geometric parameters, accurate identification and prediction of damage locations in continuous concrete structures on bridge decks can be achieved. Identification can be completed even when damage is at the micro-crack stage, demonstrating good timeliness and high sensitivity in damage location prediction.

[0076] Specifically: This disclosure does not rely on the deployment of sensors, that is, it does not use sensor detection methods, but adopts a completely new prediction method. Therefore, it does not have the problems of large workload, high cost and high maintenance cost of point-based sensors. Compared with point-based sensors, this disclosure has less workload (the number of electrodes that need to be deployed is less than the number of point-based sensors), lower cost (the deployment cost of electrodes is much lower than the cost of sensors), and lower maintenance cost (compared with sensors, electrodes do not have problems such as easy aging and drift). This disclosure improves the timeliness and sensitivity of damage detection: by collecting the current change rate data of the self-sensing concrete of the continuous bridge deck structure, and using its current change rate under damage conditions, the early electrical response capture of internal damage in the continuous bridge structure can be achieved, and the identification can be completed when the damage is in the microcrack stage. This disclosure requires only one force-electric coupling finite element model to obtain training data under different working conditions, resulting in low economic and time costs for data acquisition. Field tests struggle to cover all possible geometric dimensions, damage locations, and other parameters, and can cause irreversible damage to the structure. This disclosure, based on parametric modeling, can systematically control these parameters, accurately simulate damage location and extent, and provide high-precision training data for deep learning. It effectively solves the problems of high cost, limited sample size, and low working condition coverage in the original data acquisition process, providing a solid data foundation for the construction of deep learning models. This disclosure selects the Coral loss function, which transforms the ordered classification problem into multiple binary classification problems through binarization, effectively preserving the spatial sequence of damage locations. Simultaneously, the prediction model employs an architecture combining convolutional neural networks and residual connections, along with normalization, adaptive average pooling, and multi-level Dropout regularization strategies. Cross-layer connections alleviate the gradient vanishing problem in deep networks, enhance feature reuse capabilities, and thus improve the model's ability to extract local features from current change rate data and its generalization performance. This disclosure constructs a multi-task ordered classification model. By adopting a sharing strategy of shared feature extraction layer and shared fully connected layer, the two tasks of damage initiation position and damage termination position share the same set of convolutional neural network parameters during the feature learning stage, realizing information complementarity and feature enhancement. At the same time, the model sets two independent classification heads and synchronously inputs the prediction results of the two tasks. By jointly optimizing the loss functions of the two tasks, the model can autonomously learn the relationship between the two tasks during training and explore the spatial distribution patterns of the two, thereby improving the accuracy of prediction and recognition. To address the issue of uneven distribution of damage samples in practical engineering, this disclosure integrates a synthetic minority class oversampling technique. This technique uses linear interpolation of the minority class samples in the feature space to generate new samples, effectively balancing the number of samples in each class and mitigating the bias in model accuracy towards the majority class. Simultaneously, a data quality control process is constructed, including missing value imputation, outlier handling, and standardization. Quantile-based outlier truncation effectively suppresses interference from sensor noise, and Z-score standardization eliminates dimensional differences. These data augmentation and quality control measures significantly improve robustness and generalization performance in actual operation, providing solid data support for deep learning models in bridge damage identification.

[0077] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0078] The embodiments described above are merely illustrative of several implementations of this disclosure, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this disclosure, and these all fall within the protection scope of this disclosure. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A method for predicting the location of damage in continuous concrete structures of bridge decks, characterized in that, The continuous bridge deck structure uses self-sensing concrete material, and the prediction method includes: Obtain structural parameters of the superstructure and continuous bridge deck construction of bridges with different geometric dimensions from bridge design drawings; A force-electric coupling finite element model of the entire bridge was established, and a dataset was obtained by simulating the damage of the continuous concrete structure of the bridge deck. The dataset includes the bridge's geometric parameters, as well as the current change rate data, damage initiation location data, and damage termination location data of the continuous structure of the bridge deck. Based on the dataset, a stratified sampling method is used to divide the training set and the validation set; A multi-task ordered classification model using Coral loss is trained on the training set, and an early stopping mechanism is used on the validation set to select the optimal multi-task ordered classification model, resulting in a prediction model for the initial and final damage locations of the continuous concrete structure on the bridge deck. The multi-task ordered classification model includes a shared feature extraction layer, a shared fully connected layer, a first output head, and a second output head. The shared feature extraction layer extracts features from the input data and outputs a multi-dimensional vector. The shared fully connected layer performs dimensionality reduction and compression on the multi-dimensional vector and implements nonlinear feature transformation using the ReLU activation function. A random deactivation regularization method is used to suppress overfitting and output a compressed nonlinear damage feature vector. The first output head predicts the initial damage location based on the compressed nonlinear damage feature vector, and the output head predicts the final damage location based on the compressed nonlinear damage feature vector. The current change rate data of the continuous structural concrete of the bridge deck to be damaged and the corresponding bridge geometric parameters are obtained. The prediction model is then used to predict the initial and final damage locations of the continuous structural concrete of the bridge deck.

2. The method for predicting the location of damage to continuous concrete structures on bridge decks according to claim 1, characterized in that, The step of dividing the dataset into training and validation sets using stratified sampling specifically involves dividing the dataset into training, validation, and test sets using stratified sampling. The prediction method also includes the step of evaluating the prediction model using the test set.

3. The method for predicting the location of damage to continuous concrete structures on bridge decks according to claim 1, characterized in that, The step of dividing the dataset into training, validation, and test sets using stratified sampling specifically includes: Each initial damage location data and its corresponding initial damage location data in the dataset are merged into a composite label. Based on the dataset with the composite label, a stratified sampling method is used to divide the dataset into a training set, a validation set, and a test set.

4. The method for predicting the location of damage to continuous concrete structures on bridge decks according to claim 3, characterized in that, Before training the multi-task ordered classification model using Coral loss based on the training set, the following steps are also included: For infinity values ​​in the training, validation, and test sets, convert the infinity values ​​into missing values ​​and fill in all missing values; Outliers in the training, validation, and test sets are handled using the quantile Windsor truncation method. Z-score standardization is performed on the data in the training, validation, and test sets.

5. The method for predicting the location of damage to continuous concrete structures on bridge decks according to any one of claims 1 to 4, characterized in that, Before training the multi-task ordered classification model using Coral loss based on the training set, the following step is also included: generating synthetic samples using a synthetic minority class oversampling technique, wherein the synthetic samples are used as data in the training set.

6. The method for predicting the location of damage to continuous concrete structures on bridge decks according to claim 1, characterized in that, The loss function of the multi-task ordered classification model is: in, The loss value for the Coral loss function for ordered classification; The number of samples in a batch; This is a sample index used to identify the first sample in a batch. One sample; The total number of categories in an ordered classification; The number of binary classification boundaries that need to be predicted for each sample; The threshold index represents the threshold value in an ordered classification task. A binary classification text; For the first The sample at the th The true label on the binary sub-problem; It is the sigmoid activation function; For the model to the first The sample at the th Logit output on a binary subproblem.

7. The method for predicting the location of damage to continuous concrete structures on bridge decks according to claim 1, characterized in that, The shared feature extraction layer includes a first convolutional block, a second convolutional block, a third convolutional block, and an average pooling layer arranged sequentially. The first and second convolutional blocks are used to perform batch normalization, ReLU activation, max pooling, and random deactivation on the data in sequence. The third convolutional block is a residual block, which is used to first increase the number of channels by using two convolutional layers, and then achieve identity mapping through skip connections. The average pooling layer is used to compress features of arbitrary length into a fixed dimension using adaptive average pooling and then perform random deactivation.

8. The method for predicting the location of damage in continuous bridge deck concrete according to claim 1, characterized in that, The parameters of the multi-task ordered classification model trained using Coral loss are updated using the AdamW optimizer, as shown in the following formula: in, These are the updated parameter values; These are the parameter values ​​before the update; Represented as the learning rate; It is expressed as the bias correction estimate of the first moment of the gradient; This is expressed as the bias correction estimate of the second moment of the gradient; This is expressed as the weight decay coefficient; To prevent division by zero of extremely small constants; The learning rate scheduler for the multi-task ordered classification model trained using Coral loss employs cosine annealing, causing the learning rate to periodically decay from its initial value to near zero according to a cosine function, as shown in the following formula: in, Indicates the first Learning rate at each round; This represents the minimum learning rate; This is represented as the initial maximum learning rate; Indicates the number of rounds currently performed; The total number of cycles in a complete cosine cycle.

9. A system for predicting the location of damage in continuous concrete structures on bridge decks, characterized in that, The continuous bridge deck structure uses self-sensing concrete material, and the prediction system includes: The structural parameter acquisition module is used to obtain the structural parameters of the superstructure and continuous bridge deck structure of bridges with different geometric dimensions from bridge design drawings. The finite element model building module is used to build a force-electric coupling finite element model of the entire bridge, including the continuous structure of the bridge deck, based on the structural parameters. The dataset acquisition module uses the force-electric coupling finite element model to simulate the damage of the continuous concrete structure of the bridge deck to obtain a dataset. The dataset includes bridge geometric parameters, as well as current change rate data, damage initiation location data, and damage termination location data of the continuous structure of the bridge deck. The partitioning module is used to divide the dataset into a training set and a validation set using stratified sampling. The training module is used to train a multi-task ordered classification model using Coral loss on the training set, and to select the best multi-task ordered classification model on the validation set using an early stopping mechanism, so as to obtain the prediction model of the initial and termination positions of damage in the continuous concrete structure of the bridge deck. The multi-task ordered classification model includes a shared feature extraction layer, a shared fully connected layer, a first output head, and a second output head. The shared feature extraction layer is used to extract features from the input data and output a multi-dimensional vector. The shared fully connected layer is used to reduce the dimensionality of the multi-dimensional vector and to perform nonlinear feature transformation through the ReLU activation function. The random deactivation regularization method is used to suppress overfitting and output a compressed nonlinear damage feature vector. The first output head is used to predict the initial position of damage based on the compressed nonlinear damage feature vector, and the output head for the termination position of damage is used to predict the termination position of damage based on the compressed nonlinear damage feature vector. The prediction module is used to acquire the current change rate data of the continuous structural concrete of the bridge deck and the corresponding bridge geometric parameters for which damage location prediction is to be performed, and to use the prediction model to predict the initial and final damage locations of the continuous structural concrete of the bridge deck.

10. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for predicting the location of damage to continuous concrete structures on bridge decks as described in any one of claims 1-8.

Images