Method and system for predicting concrete damage of expansion joint based on multi-task deep learning

Abstract

The present disclosure relates to a multi-task deep learning-based expansion joint concrete damage prediction method and system, designed for the field of intelligent prediction. The method comprises: simulating the piezoresistive effect of self-sensing concrete to obtain a force-electricity coupling finite element model; setting a multi-level damage degree coefficient, calculating the current change rate before and after damage for each damage position of each damage degree coefficient, obtaining an original data set and performing data preprocessing; training a multi-task MLP model with a shared bottom layer, wherein the feature learning is realized by combining damage degree regression loss, cross-entropy-based loss level classification loss, and focus loss-based damage position classification loss; performing hyperparameter optimization using three-fold cross-validation; and predicting the damage of the expansion joint concrete using the prediction model. The present disclosure can cooperatively predict the damage position, damage degree and damage level, is accurate in prediction, and can identify early damage.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent prediction of concrete damage, and in particular to a method and system for predicting concrete damage in expansion joints based on multi-task deep learning. Background Technology

[0002] To accommodate deformation caused by temperature changes, concrete shrinkage, and vehicle loads, expansion joints are necessary at locations where structural displacements are concentrated in bridges. However, in actual engineering, repeated impacts from vehicle loads can lead to microcracks in the anchorage zone concrete. Furthermore, the stress conditions in the anchorage zone concrete are complex, bearing not only the vertical pressure from vehicle wheels but also the horizontal shear force during vehicle braking. Environmental factors cannot be ignored either; extreme temperature differences cause repeated expansion and contraction of the expansion joints, exacerbating fatigue damage; rainwater and corrosive media such as chloride ions seep into the concrete through microcracks, causing corrosion of the reinforcing steel in the anchorage zone. The expansion of the steel steel then leads to cracking of the concrete.

[0003] Damage to the concrete in the anchorage zone not only affects traffic safety but also accelerates fatigue damage at beam ends and supports, potentially leading to substructure failure and impacting structural safety. There is an urgent need to design a method and system for predicting concrete damage in expansion joints. This system should be able to accurately identify the specific extent of concrete damage during bridge operation, enabling the coordinated prediction of multiple damage indicators, particularly the identification of early-stage micro-cracks, to provide early warning of damage and support preventative maintenance. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and system for predicting concrete damage in expansion joints based on multi-task deep learning to address the above problems.

[0005] To solve the above problems, the present disclosure adopts the following technical solution: Firstly, a method for predicting concrete damage in expansion joints based on multi-task deep learning is provided, including: Collect bridge structural parameters based on bridge design drawings; A finite element model of the bridge was established based on the bridge structural parameters, in which self-sensing concrete was used in the expansion joint anchorage zone. The piezoresistive effect of the self-sensing concrete was simulated to obtain a force-electric coupling finite element model. Multi-level damage degree coefficients were set, and the rate of change of current before and after damage was calculated for each damage location of each damage degree coefficient, resulting in an original data set including bridge structural parameters, rate of change of current, damage location corresponding to the rate of change of current, and damage degree coefficient corresponding to the rate of change of current. Data preprocessing was performed on the original dataset, including: constructing a feature matrix composed of current change rate; using damage location as classification label, extracting continuous floating-point values ​​of damage degree as regression label, and mapping damage degree to loss level to generate auxiliary classification label; using stratified sampling to divide the training set, validation set and test set, with damage location as stratified variable; performing feature standardization on bridge structural parameters and current change rate, and performing regression target standardization on regression label; Training employs a shared-bottom-level multi-task MLP model to obtain a predictive model capable of simultaneously predicting the location, extent, and grade of concrete damage in the expansion joint anchorage zone. This model includes: joint damage extent regression loss, cross-entropy-based loss level classification loss, and focus-based loss damage location classification loss to achieve feature learning. The damage extent regression loss function is selected from MSE, Huber, and LogCosh using an automatic hyperparameter search method. Triple-fold cross-validation is used for hyperparameter tuning. An early stopping mechanism is used to determine the hyperparameters. Obtain the bridge structural parameters and the current change rate of the expansion joint concrete of the bridge to be predicted, and use the prediction model to predict the damage of the expansion joint concrete.

[0006] In a preferred embodiment, the bridge structural parameters include bridge geometry, support type, and expansion joint size.

[0007] In a preferred embodiment, the multi-task MLP model includes a shared feature extraction layer and three output heads. The shared feature extraction layer is used to extract shared features, and the output heads are used to classify damage location, damage level, or damage severity based on the shared features.

[0008] In a preferred embodiment, the shared feature extraction layer consists of multiple fully connected blocks, each of which is used to perform linear transformation, ReLU activation, and random activation operations in sequence.

[0009] In a preferred embodiment, the number of hidden layers and neurons in the shared feature extraction layer are both adaptively determined by Optuna; the hyperparameter tuning using tri-fold cross-validation specifically involves: achieving hyperparameter tuning through random search of Optuna hyperparameters combined with tri-fold cross-validation.

[0010] In a preferred embodiment, the formula for the weighted combination index of hyperparameter search in the hyperparameter tuning using triple cross-validation is: in, This represents a weighted combination index of the comprehensive damage degree regression, damage level classification, and damage location classification tasks under a certain fold cross-validation in the hyperparameter search. The mean absolute error of the damage severity regression task; Represents the weighted F1 score for the damage severity classification task; This represents the weighted F1 score for the damage location classification task.

[0011] In a preferred embodiment, LogCosh is selected from MSE, Huber, and LogCosh as the function of the damage degree regression loss using an automatic hyperparameter search method.

[0012] In a preferred embodiment, the method for predicting concrete damage in expansion joints further includes the step of evaluating the prediction model using multi-dimensional evaluation indicators.

[0013] In a preferred embodiment, accuracy, precision, recall, and F1 score are used as evaluation indicators for both the prediction of damage location and the prediction of damage level; mean absolute error, root mean square error, and coefficient of determination are used as evaluation indicators for the prediction of damage severity.

[0014] Secondly, a multi-task deep learning-based system for predicting concrete damage in expansion joints is provided, including: The data acquisition module is used to collect bridge structural parameters based on bridge design drawings. The finite element model construction module is used to establish a finite element model of the bridge based on the bridge structural parameters, wherein the concrete in the expansion joint anchorage zone is self-sensing concrete; it is used to simulate the piezoresistive effect of the self-sensing concrete to obtain a force-electric coupling finite element model; it is used to set multi-level damage degree coefficients, and for each damage location of each damage degree coefficient, it calculates the current change rate before and after the damage, and obtains the original data set including bridge structural parameters, current change rate, damage location corresponding to the current change rate, and damage degree coefficient corresponding to the current change rate. The data preprocessing module is used to preprocess the data based on the original dataset. Specifically, it is used to: construct a feature matrix composed of the current change rate; use the damage location as a classification label, extract continuous floating-point values ​​of the damage degree as regression labels, and map the damage degree to the loss level to generate auxiliary classification labels; use stratified sampling to divide the training set, validation set, and test set, with the damage location as a stratification variable; perform feature standardization on the bridge structural parameters and current change rate, and perform regression target standardization on the regression labels. The training module is used to train a multi-task MLP model with a shared underlying layer to obtain a prediction model that can simultaneously predict the location, degree, and level of concrete damage in the expansion joint anchorage zone. Specifically, it is used for: feature learning through joint damage degree regression loss, cross-entropy-based loss level classification loss, and focus-based loss damage location classification loss. The function for damage degree regression loss is selected from MSE, Huber, and LogCosh using an automatic hyperparameter search method; hyperparameter tuning is performed using three-fold cross-validation; and hyperparameters are determined using an early stopping mechanism. The prediction module is used to obtain the bridge structural parameters of the bridge to be predicted and the current change rate of the expansion joint concrete, and to predict the damage of the expansion joint concrete using the prediction model.

[0015] The beneficial effects of the aforementioned method and system for predicting damage in expansion joint concrete based on multi-task deep learning are as follows: A force-electric coupling finite element model is constructed, and based on this, a raw dataset including bridge structural parameters, current change rate, damage location, and damage degree coefficient is built. Specifically, by introducing self-sensing concrete into the finite element model, its piezoresistive effect causes a change in the local current change rate when micro-cracks appear inside the concrete. Furthermore, a data preprocessing and multi-task MLP model training method is specifically designed, including the design of three loss terms acting synergistically on a shared underlying layer. This design enables the method and system to collaboratively predict damage location, damage degree, and damage level with accuracy, and to identify early micro-cracks. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method in one embodiment of the present disclosure; Figure 2 This is a loss curve diagram of the training and verification process in an embodiment of this disclosure; Figure 3 This is a confusion matrix diagram of damage location prediction for the test set in an embodiment of this disclosure; Figure 4 This is a scatter regression plot of the damage level of the test set in the embodiments of this disclosure; Figure 5 This is a violin plot showing the residual damage levels of the test set in an embodiment of this disclosure. Figure 6 This is a confusion matrix diagram for damage level prediction of the test set in the embodiments of this disclosure; Figure 7 This is a schematic diagram of the system structure in one embodiment of the present disclosure. Detailed Implementation

[0017] The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0018] See Figure 1This paper presents a method for predicting concrete damage in expansion joints based on multi-task deep learning, including: Collect bridge structural parameters based on bridge design drawings; A finite element model of the bridge was established based on the bridge structural parameters, in which self-sensing concrete was used in the expansion joint anchorage zone. The piezoresistive effect of the self-sensing concrete was simulated to obtain a force-electric coupling finite element model. Multi-level damage degree coefficients were set, and the rate of change of current before and after damage was calculated for each damage location of each damage degree coefficient, resulting in an original data set including bridge structural parameters, rate of change of current, damage location corresponding to the rate of change of current, and damage degree coefficient corresponding to the rate of change of current. Data preprocessing was performed on the original dataset, including: constructing a feature matrix composed of current change rate; using damage location as classification label, extracting continuous floating-point values ​​of damage degree as regression label, and mapping damage degree to loss level to generate auxiliary classification label; using stratified sampling to divide the training set, validation set and test set, with damage location as stratified variable; performing feature standardization on bridge structural parameters and current change rate, and performing regression target standardization on regression label; Training employs a shared-bottom-level multi-task MLP model to obtain a predictive model capable of simultaneously predicting the location, extent, and grade of concrete damage in the expansion joint anchorage zone. This model includes: joint damage extent regression loss, cross-entropy-based loss level classification loss, and focus-based loss damage location classification loss to achieve feature learning. The damage extent regression loss function is selected from MSE, Huber, and LogCosh using an automatic hyperparameter search method. Triple-fold cross-validation is used for hyperparameter tuning. An early stopping mechanism is used to determine the hyperparameters. Obtain the bridge structural parameters and the current change rate of the expansion joint concrete of the bridge to be predicted, and use the prediction model to predict the damage of the expansion joint concrete.

[0019] The following details the method for predicting concrete damage in expansion joints based on multi-task deep learning, including: Step 1: Based on the bridge design drawings, collect the bridge structural parameters (also known as the basic structural parameters of the bridge), including the bridge's geometric dimensions, support type, expansion joint dimensions, etc., to provide input parameters for subsequent finite element simulation.

[0020] In this embodiment, based on the bridge design drawings, the bridge's geometric parameters, support types, and expansion joint dimensions are collected. Preferably, bridge structural details are also collected, and the properties of materials such as concrete and steel are recorded to provide data support for subsequent precise finite element simulations. Bridge structural details include superstructure details and ancillary structure details.

[0021] Step 2: Establish a finite element model of the bridge based on the basic structural parameters of the bridge. The concrete in the expansion joint anchorage zone of the bridge finite element model is self-sensing concrete. Based on the bridge finite element model, simulate the piezoresistive effect of the self-sensing concrete to obtain a force-electric coupling finite element model. Based on the force-electric coupling finite element model, set multiple levels of damage degree coefficients and preset damage locations for each level of damage degree coefficient. Based on the force-electric coupling finite element model, calculate the rate of change of current before and after damage for each damage location of each level of damage degree coefficient.

[0022] Preferably, a finite element model of the bridge is established, including the bridge superstructure, bridge deck pavement, expansion joints, and bearing types, but excluding the bridge substructure and bearings. It is understood that, correspondingly, the force-electric coupling finite element model is implemented based on the aforementioned bridge finite element model, and also excludes the bridge substructure and bearings.

[0023] Understandably, the concrete in the expansion joint anchorage zone is poured between the steel anchor bars of the expansion joint and the pre-reserved groove in the bridge beam, directly and firmly anchoring the expansion joint device and the main bridge structure together, playing a crucial role in load transfer. The expansion joint device includes steel components and elastic sealant, among other parts.

[0024] A self-sensing concrete material is defined in the expansion joint anchorage zone, and its piezoresistive coefficient matrix is ​​imported to achieve force-electric multiphysics coupling. This simulates the piezoresistive effect of the self-sensing concrete, constructing a numerical model of the bridge and expansion joint—that is, establishing a force-electric coupling finite element model based on the piezoresistive effect. Simultaneously, different damage degree coefficients are set to characterize the concrete damage state, and the current change rate of the concrete in the expansion joint anchorage zone is extracted.

[0025] The bridge finite element model established based on the basic structural parameters of the bridge fully considers the actual structural details of the expansion joint device, including the geometric dimensions and material properties of the steel components and elastic sealant. The concrete material in the expansion joint anchorage zone is self-sensing concrete, and a piezoresistive coefficient matrix and resistivity are introduced, together forming a force-electric coupling constitutive model. Based on this, a force-electric coupling finite element model is established, and alternating potentials are set on the self-sensing concrete in the expansion joint anchorage zone to simulate the electrode arrangement in actual engineering. By assigning the piezoresistive coefficient matrix and resistivity to the self-sensing concrete, under external loads, the deformation inside the concrete causes a change in resistivity, thereby altering the current response between the electrodes. As an example rather than a limitation, multiple levels of damage degree coefficients are set within the range of 0 to 0.6, for example, 60 levels of damage degree coefficients are uniformly set within the range. For each level of damage coefficient, all damage locations are traversed, the current response before and after damage is extracted, and the rate of change of current is calculated. For each level of damage coefficient, the current response after damage needs to be extracted separately for each damage location. Specifically, when traversing all preset damage locations, the steady-state current response at each location should be collected immediately after damage loading is completed, and the damage degree coefficient at the corresponding location should be recorded synchronously. Then, the operation is repeated at the next location to ensure that the current change rate calculation at each location has a unique correspondence.

[0026] In this embodiment, the basic data required for training the MLP model is generated through step 2, which is called the original data set. The original data set includes bridge structural parameters, current change rate, damage location corresponding to the current change rate, and damage degree coefficient corresponding to the current change rate.

[0027] Step 3: Preprocessing the original dataset, including: constructing a feature matrix composed of current change rate; using damage location as classification label, extracting continuous floating-point values ​​of damage degree as regression label, mapping damage degree to loss level to generate auxiliary classification label; using stratified sampling to divide the training set, validation set and test set, with damage location as stratification variable; performing feature standardization on bridge structural parameters and current change rate, and performing regression target standardization on regression label; This involves preprocessing the original dataset, including feature label separation, dataset partitioning, feature standardization, regression target standardization, and damage level mapping. Feature and label separation includes extracting continuous floating-point values ​​of damage level as regression labels from damage location as classification labels, mapping damage level to loss level, and generating auxiliary classification labels.

[0028] Feature and label separation is the core step in data preprocessing. A feature matrix is ​​constructed based on the original data, specifically a feature matrix composed of current change rates, called the current change rate matrix. Simultaneously, three supervisory labels are built: damage location is converted into a 0-based index as the classification label; continuous floating-point values ​​of damage severity are extracted as regression labels; and damage severity is mapped to four loss severity levels based on engineering thresholds, generating auxiliary classification labels. This separation process ensures the correspondence between features and labels, providing a data foundation of multi-supervisory signals for subsequent stratified sampling, standardization, and multi-task collaborative prediction models.

[0029] Damage severity level mapping is a mapping from damage severity to damage level, damage severity As an example, the injury level is defined as mild injury, moderate injury, moderate to severe injury, and severe injury, and the mapping formula is as follows: Based on engineering experience with concrete damage in the anchorage zone of bridge expansion joints, the degree of damage is divided into the above damage levels. The degree of damage is used for fine quantification and provides auxiliary constraints for the damage level. Continuous regression and discrete classification complement each other to prevent the regression results from contradicting the engineering attempts. By dividing the level by equal interval thresholds, it is ensured that there are enough samples in each level interval to avoid serious imbalance in the classification task. This design takes into account both the engineering physical meaning and the sample balance required for training multi-task models.

[0030] The dataset partitioning employs a strategy combining stratified sampling and two partitioning steps. First, all samples are partitioned according to damage location as the stratification variable, and the original dataset is divided into training, validation, and test sets at proportions of 70%, 10%, and 20% respectively. The first partition creates a test set comprising 20% ​​of the original dataset, while the remaining data is partitioned into training and validation sets using the same stratification variable. A fixed random seed ensures reproducibility.

[0031] This partitioning strategy constructs three independent and uniformly distributed subsets for multi-task learning: a training set for parameter updates, a validation set for hyperparameter selection and early stopping, and a test set for final generalization performance evaluation. Stratified sampling effectively mitigates the problem of imbalanced lesion location categories, preventing certain rare locations from being completely missing or having insufficient sample sizes in specific datasets, thus ensuring the reliability of the model's predictions of lesion location, severity, and severity level. The existence of the validation set enables the early stopping mechanism to be implemented. Simultaneously, it supports the Optuna framework in using the validation set for three-fold cross-validation during hyperparameter search, significantly reducing the risk of overfitting, improving the stability and generalization ability of hyperparameter selection, and providing a data foundation for subsequent feature standardization, model training, and performance evaluation.

[0032] Feature standardization uses the Z-score method to transform the data into a distribution with a mean of 0 and a standard deviation of 1, eliminating dimensional differences, preventing features with large values ​​from dominating gradient updates, accelerating gradient descent, and improving model training stability and generalization ability.

[0033] Standardizing the current change rate primarily serves to prevent features with large values ​​from dominating gradient updates, accelerate gradient descent, and improve model training stability and generalization ability. The structural parameters of the bridge are also standardized, mainly to eliminate dimensional differences.

[0034] Updating the current change matrix through feature standardization can be understood as first performing feature standardization on the current change rate data and then constructing the current change rate matrix.

[0035] The standardized formula is as follows: in, This represents the value after standardization. Indicates the value of the original feature or regression target; It represents the mean of a certain feature or regression target in the training set; It represents the standard deviation of a feature or regression target in the training set.

[0036] The regression target standardization also adopts the Z-score method, as shown in the standardization formula above. It is only applied to continuous floating-point values ​​of damage degree. The original damage degree value is small, and direct regression will lead to the gradient scale being too small and overwhelmed by the gradient of the classification task. Standardization balances the gradient scale of the regression branch task with that of the classification task, improves the convergence quality of the regression branch, improves the stability of multi-task joint training, and is independent of feature standardization, without interfering with each other. Each retains its mean and standard deviation, which can achieve lossless reversibility and return the standardized value predicted by the model to the true damage degree.

[0037] Both feature standardization and regression target standardization strictly adhere to the principle of only fitting the mean and standard deviation of the training set before transforming the validation and test sets, thus avoiding information leakage and jointly ensuring the stable training and reliable prediction of multi-task deep learning models.

[0038] Understandably, standardization allows the training, validation, and test sets to be updated automatically. In other words, the data used in step 4 consists of the standardized training set, the standardized validation set, and the standardized test set.

[0039] Step 4: Based on the preprocessed training set, a multi-task MLP model with a shared underlying layer is adopted. Collaborative feature learning is achieved by combining three types of loss functions. Hyperparameter search is performed using Optuna to complete the automatic search of hyperparameters and determine the regression loss function. Combined with three-fold cross-validation and early stopping mechanism, a prediction model that can simultaneously predict the location, degree, and level of concrete damage in the expansion joint anchorage zone is obtained.

[0040] Optuna is an open-source Python hyperparameter optimization library specifically designed for automated intelligent hyperparameter tuning. The multi-task MLP model is a multi-task deep learning model based on a modified multilayer perceptron architecture.

[0041] This embodiment employs a method of sharing underlying network parameters. The multi-task MLP model includes a shared feature extraction layer. The three tasks—damage location classification, damage level classification, and damage severity regression—share this shared feature extraction layer, forcing the model to learn a common damage representation, achieving knowledge transfer and parameter reuse, and improving model training efficiency. Different loss functions are configured according to each task type, using Focal Loss, cross-entropy loss, and adaptive regression loss to complete supervised learning for the corresponding tasks. A weighted fusion strategy is used to jointly optimize multiple losses, and cross-regularization is applied based on the formation of different tasks, effectively compensating for the limitations of single-task feature learning. The complementary constraints of multi-branch outputs are fully utilized: damage level classification provides ordered semantic anchoring information, while the damage severity regression branch outputs continuous quantitative representation results, with the two mutually constraining and calibrating each other. This significantly adapts to complex data scenarios with imbalanced classes and sparse damage samples, significantly improving the stability of model training.

[0042] The multi-task MLP model's input layer receives data such as bridge structural parameters and current change rate. The shared feature extraction layer consists of multiple fully connected blocks, each performing linear transformation, ReLU activation, and dropout operations sequentially. The number of hidden layers and neurons in the shared feature extraction layer are adaptively determined by Optuna. The multi-task MLP model has three independent output heads: one for damage location classification, one for damage level classification, and one for damage severity regression. Each output head is a single fully connected layer, and all three output heads share features extracted from the bottom layer (shared feature extraction layer).

[0043] This embodiment achieves hyperparameter tuning through Optuna's random hyperparameter search combined with three-fold cross-validation. The training set is divided into three subsets. One subset is used as the validation set, and the remaining two subsets are used as the training set. This process is repeated three times, and the performance index of the three validations is calculated as the hyperparameter evaluation score for that group. This score guides Optuna to update the search direction, providing high flexibility. The built-in TPE (Tree-structured Parzen Estimator) Bayesian sampler can adaptively predict better hyperparameter regions based on historical trial results, avoiding blind grid search or random search. Optuna calls a complete tuning task a Study, where the evaluation of each parameter combination is called a Trial. The asynchronous Successive Halving (continuous halving or successive halving) pruning algorithm can terminate trails with no obvious potential in advance during cross-validation, thereby significantly saving computational resources, significantly improving search efficiency, and accelerating the optimization process.

[0044] In this example, the choice of loss function varies depending on the target task. Because there are many lesion location categories and the samples are unevenly distributed, with some lesion location categories having a low sample size, using a conventional cross-entropy loss function would cause the model to favor the majority class and ignore the minority class, resulting in low accuracy in identifying rare locations. Therefore, the Focal Loss function is used, where... It can reduce the weight of easily classified samples and focus on learning difficult-to-classify samples, effectively alleviating the problem of class imbalance.

[0045] The Focal Loss function is used for damage location classification, and the formula is as follows: in, The loss value represents the classification of the damage location. This represents the predicted confidence probability of the true label category of the sample; This represents the class balancing factor, which assigns a greater loss weight to minority class samples. This represents the power loss factor, which adjusts the model's focus on easily classified and difficult-to-classify samples.

[0046] There are four categories of damage levels. The sample distribution is relatively uniform, and there is a natural order relationship between the levels. The convergence is stable. Cross-entropy is adapted to the ordered classification characteristics of the damage level classification task. Combined with the equivalent partitioning method, stable auxiliary constraint learning is achieved.

[0047] The loss function used for classification is the cross-entropy loss function, as shown in the following formula: in, This indicates the loss level classification. This indicates the total number of samples, or the number of samples in a batch. This represents the sample index, with a value ranging from 1 to... ; Indicates sample The true damage level label. The model represents the samples The probability distribution of the predicted damage level.

[0048] For damage severity regression tasks, damage severity itself is a continuous value, and different loss functions have different sensitivities to outliers. Therefore, the damage severity regression loss function is automatically searched using Optuna and combined with three-fold cross-validation to select the optimal loss function from MSE, Huber, and LogCosh based on the characteristics of the current dataset. This allows for the adaptation of prediction needs to different data distributions.

[0049] The formula for the MSE loss function (mean squared error loss function) is as follows: in, This represents the mean squared error loss value. Indicates the number of samples in a batch; Indicates sample The true extent of damage; Indicates sample The predicted degree of damage.

[0050] The Huber loss function (also known as the smoothed average absolute error loss function) is formulated as follows: in, This represents Huber's loss value. This represents the prediction error for a single sample. This represents the threshold parameter, which controls the conversion boundary between squared error and linear error.

[0051] The formula for the LogCosh loss function (log-hyperbolic cosine loss function) is as follows: in, This represents the logarithmic hyperbolic cosine loss value. Represented as a hyperbolic cosine function, This represents the information input to the multi-task MLP model. This represents the prediction error for a single sample.

[0052] In the process of hyperparameter tuning using triple cross-validation, the formula for the weighted combination index of hyperparameter search is: in, This represents a weighted combination index that integrates damage degree regression, damage level classification, and damage location classification tasks under a certain fold cross-validation in the hyperparameter search, used to evaluate the model's performance under that fold. The mean absolute error represents the damage degree regression task, reflecting the model's bias in predicting continuous damage degree; the smaller the value, the higher the regression accuracy. This represents the weighted F1 score for the damage classification task; the closer it is to 1, the better the classification performance. The error rate indicating the classification level; This represents the weighted F1 score for the damage location classification task.

[0053] In this embodiment, the selection of the loss function for the damage severity regression task underwent an adaptive process from exploring multiple types to converging with LogCosh. In the early stage, both Huber and LogCosh performed well, with Huber achieving a better value first. In the middle stage, Huber dominated, but the optimal value of the weighted combined index remained around 0.08. In the later stage, Optuna's sampling strategy gradually leaned towards LogCosh, continuously producing combined indices below 0.07. The target value of the best trial Trial25 (25th evaluation) dropped to 0.064, significantly higher than the initial optimal value. Therefore, LogCosh loss was used as the loss function for damage severity. This process achieved automatic selection of the regression loss function.

[0054] The formula for the total loss of the model is as follows: in, This represents the total loss value of the multi-task MLP model. , , All are weights, respectively , , The weights are determined in the three-fold cross-validation through hyperparameter search.

[0055] The model's loss function is selected based on the characteristics of each task. Through weight balancing, gradient scale matching of the three tasks is achieved, preventing one task from dominating the training. This facilitates the collaborative prediction of lesion location, lesion degree, and lesion level, forming positive cross-regularization, improving the model's generalization ability, and also providing interpretability for engineering applications.

[0056] Figure 2 The loss curves for the training and validation process of one embodiment show that the patterns learned by the model on the training set can be well transferred to the unseen validation data; the model has high learning efficiency and quickly captures the core patterns of the data; the model does not exhibit obvious overfitting and shows good generalization ability.

[0057] Step 5: Obtain the bridge structural parameters of the bridge to be predicted and the current change rate of the expansion joint concrete of the bridge to be predicted, and input them into the prediction model. Use the prediction model to predict the damage of the expansion joint concrete, that is, use the prediction model to predict the damage location, damage degree (i.e. damage degree coefficient) and damage level.

[0058] In this embodiment, the method for predicting concrete damage in expansion joints further includes a step of evaluating the prediction model using multi-dimensional evaluation metrics. In this embodiment, for classification tasks involving damage location and damage level, accuracy, precision, recall, and F1 score are used as evaluation metrics; for regression tasks involving damage degree, mean absolute error, root mean square error, and coefficient of determination are selected as evaluation metrics to achieve a comprehensive representation of the model's performance. Accuracy represents the proportion of correct predictions overall; precision measures the proportion of true positives among the predicted positives; recall represents the proportion of true positives correctly identified. The combination of both reveals the ability to identify rare damage locations and specific levels; the F1 score represents the harmonic mean of precision and recall, providing a balanced metric suitable for scenarios with uneven distribution.

[0059] The evaluation metrics for the classification task in this embodiment are shown in Table 1: Table 1 Classification Evaluation Indicators Mean absolute error (MAE) directly reflects the average absolute error between the predicted and actual values; root mean square error (RMSE) detects deviations from the prediction by squared error; the coefficient of determination (R²) 2 () indicates the proportion of captured data to the total variance.

[0060] The evaluation metrics for the damage severity regression task in this model were MAE 0.022, RMSE 0.038, and R0.038. 2 The value is 0.96. Other examples include the confusion matrix for damage location, the scatter regression plot for damage severity, the residual violin plot, and the confusion matrix for damage grade prediction, as shown below. Figures 3 to 6 As shown. Figure 3 It is the confusion matrix of the 20 locations (damage locations) with the lowest accuracy, showing that the overall prediction performance is excellent. Figure 4 This is a scatter regression plot of the damage level of the test set in this embodiment. It can be seen from this regression scatter plot that there is no systematic overall deviation of the scatter points. Figure 5The violin plot of the residuals of the test set damage in this embodiment shows that the overall residual distribution is highly concentrated, with no systematic bias issues. Figure 6 The confusion matrix diagram for the test set level (damage level) prediction in this embodiment shows that the overall classification performance is excellent.

[0061] See Figure 7 This disclosure provides a multi-task deep learning-based system for predicting concrete damage in expansion joints, including: The data acquisition module is used to collect bridge structural parameters based on bridge design drawings. The finite element model construction module is used to establish a finite element model of the bridge based on the bridge structural parameters, wherein the concrete in the expansion joint anchorage zone is self-sensing concrete; it is used to simulate the piezoresistive effect of the self-sensing concrete to obtain a force-electric coupling finite element model; it is used to set multi-level damage degree coefficients, and for each damage location of each damage degree coefficient, it calculates the current change rate before and after the damage, and obtains the original data set including bridge structural parameters, current change rate, damage location corresponding to the current change rate, and damage degree coefficient corresponding to the current change rate. The data preprocessing module is used to preprocess the data based on the original dataset. Specifically, it is used to: construct a feature matrix composed of the current change rate; use the damage location as a classification label, extract continuous floating-point values ​​of the damage degree as regression labels, and map the damage degree to the loss level to generate auxiliary classification labels; use stratified sampling to divide the training set, validation set, and test set, with the damage location as a stratification variable; perform feature standardization on the bridge structural parameters and current change rate, and perform regression target standardization on the regression labels. The training module is used to train a multi-task MLP model with a shared underlying layer to obtain a prediction model that can simultaneously predict the location, degree, and level of concrete damage in the expansion joint anchorage zone. Specifically, it is used for: feature learning through joint damage degree regression loss, cross-entropy-based loss level classification loss, and focus-based loss damage location classification loss. The function for damage degree regression loss is selected from MSE, Huber, and LogCosh using an automatic hyperparameter search method; hyperparameter tuning is performed using three-fold cross-validation; and hyperparameters are determined using an early stopping mechanism. The prediction module is used to obtain the bridge structural parameters of the bridge to be predicted and the current change rate of the expansion joint concrete, and to predict the damage of the expansion joint concrete using the prediction model.

[0062] In specific implementation, the expansion joint concrete damage prediction system based on multi-task deep learning can refer to the expansion joint concrete damage prediction method based on multi-task deep learning in any of the above embodiments to achieve expansion joint concrete damage prediction. The specific implementation steps will not be repeated.

[0063] 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 expansion joint concrete damage prediction method based on multi-task deep learning according to any of the above embodiments.

[0064] This disclosure also provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the steps of the multi-task deep learning-based method for predicting concrete damage in expansion joints as described in any of the above embodiments.

[0065] The present invention discloses a method and system for predicting damage in expansion joint concrete based on multi-task deep learning. It constructs a force-electric coupling finite element model and builds a raw data set including bridge structural parameters, current change rate, damage location, and damage degree coefficient. The finite element model introduces self-sensing concrete, whose piezoresistive effect causes changes in the local current change rate when micro-cracks appear inside the concrete. Through specific design of data preprocessing and multi-task MLP model training method, including the design of three loss terms acting on a shared bottom layer, the method can collaboratively predict damage location, damage degree, and damage level. The prediction is accurate and can identify early micro-cracks, realize damage early warning, and provide technical support for preventive maintenance.

[0066] Specifically: This disclosure constructs a multi-indicator collaborative prediction structure to achieve collaborative prediction of the location, level, and degree of concrete damage in the expansion joint anchorage zone. The damage level provides discrete semantic constraints, ensuring that the regression conforms to the damage level boundary; the regression branch outputs fine-grained numerical values, assisting in more accurate damage level classification. A shared underlying network forces the feature representations of the three tasks to transfer between each other, enabling the location, level, and degree features to be jointly optimized, improving the generalization ability for small sample tasks. A weighted joint loss function automatically balances the three tasks, allowing each branch to calibrate itself during training, avoiding a single task dominating model performance. The final model can simultaneously output damage location, damage level, and degree of damage, achieving collaborative prediction of multiple damage indicators and significantly improving the comprehensiveness and accuracy of the prediction. This disclosure employs a positive cross-regularization design to enhance model robustness: the three tasks achieve gradient collaborative updates through a shared underlying network. The lesion location classification task focuses on the spatial distribution characteristics of the data, the lesion level classification task strengthens the discrimination boundary of lesion severity, and the lesion degree regression task emphasizes numerical changes. The gradients of the three tasks propagate backward simultaneously, forming a mutually constraining regularization effect. The classification task suppresses overfitting of the regression branch in regions with few samples, while the regression task provides a continuous and smooth loss surface for the classification branch, avoiding oscillations in the decision boundary. Furthermore, when the regression branch outputs data due to noise or other abnormal data interference, the classification result can serve as an independent judgment basis; conversely, when the classification branch has low confidence near the level boundary, the regression value can provide a reference. This effectively improves the model's robustness to noisy data and limited samples, increases the accuracy of predictions, and ensures the reliability of decisions in practical applications. This disclosure achieves intelligent detection and early warning: Its ability to detect early cracks is significantly superior to traditional damage identification methods. Traditional methods rely heavily on visual inspection, which is highly subjective. Displacement gauges and strain gauges are not sensitive to micro-cracks, and damage can only be detected after it reaches a certain level. This disclosure introduces self-sensing concrete. Its piezoresistive effect causes a change in the local current rate of change when micro-cracks appear inside the concrete. Through finite element simulation combined with a multi-task MLP model, the regression branch can output continuous damage values, enabling the identification of early micro-cracks and precise determination of damage location. Compared to traditional inspections or external sensors, this disclosure significantly advances the warning time for loss identification, providing a technical path for preventative maintenance. This disclosure provides accurate damage assessment data for the maintenance of bridge expansion joints and has significant engineering value.

[0067] To improve the model's generalization ability, this disclosure introduces the Optuna framework for adaptive hyperparameter search, jointly optimizing the network structure and task loss weights (three loss terms). Specifically for damage regression tasks, it incorporates MSE loss, Huber loss, and LogCosh loss into the search space, and combines three-fold cross-validation and early stopping mechanisms to verify the automatic selection of the loss function based on combined metrics. This mechanism avoids the subjectivity and blindness of manual hyperparameter tuning, enabling the model to maintain optimal generalization performance under different bridge and sensor configurations, significantly improving transferability. Improving data acquisition efficiency: This publicly available data is generated through finite element simulation, rather than relying on on-site measurements. A refined numerical model of bridge expansion joints is established, and a self-sensing concrete piezoresistive coefficient matrix is ​​input. By setting multi-level damage degree coefficients and different damage locations, force-electric multiphysics coupling analysis is conducted to calculate the current change rate at various locations in the concrete under different working conditions. Data sets including damage location, damage degree, and corresponding current response are generated in batches. This method can include common damage states, ensures absolutely accurate labels, significantly reduces data acquisition time and cost, and provides a complete and reliable training foundation for multi-task deep learning models.

[0068] 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.

[0069] 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 concrete damage in expansion joints based on multi-task deep learning, characterized in that, include: Collect bridge structural parameters based on bridge design drawings; A finite element model of the bridge was established based on the bridge structural parameters, in which self-sensing concrete was used in the expansion joint anchorage zone. The piezoresistive effect of the self-sensing concrete was simulated to obtain a force-electric coupling finite element model. Multi-level damage degree coefficients were set, and the rate of change of current before and after damage was calculated for each damage location of each damage degree coefficient, resulting in an original data set including bridge structural parameters, rate of change of current, damage location corresponding to the rate of change of current, and damage degree coefficient corresponding to the rate of change of current. Data preprocessing was performed on the original dataset, including: constructing a feature matrix composed of current change rate; using damage location as classification label, extracting continuous floating-point values ​​of damage degree as regression label, and mapping damage degree to loss level to generate auxiliary classification label; using stratified sampling to divide the training set, validation set and test set, with damage location as stratified variable; performing feature standardization on bridge structural parameters and current change rate, and performing regression target standardization on regression label; Training employs a shared-bottom-level multi-task MLP model to obtain a predictive model capable of simultaneously predicting the location, extent, and grade of concrete damage in the expansion joint anchorage zone. This model includes: joint damage extent regression loss, cross-entropy-based loss level classification loss, and focus-based loss damage location classification loss to achieve feature learning. The damage extent regression loss function is selected from MSE, Huber, and LogCosh using an automatic hyperparameter search method. Triple-fold cross-validation is used for hyperparameter tuning. An early stopping mechanism is used to determine the hyperparameters. Obtain the bridge structural parameters and the current change rate of the expansion joint concrete of the bridge to be predicted, and use the prediction model to predict the damage of the expansion joint concrete.

2. The method for predicting concrete damage in expansion joints based on multi-task deep learning according to claim 1, characterized in that, The bridge structural parameters include bridge geometry, support type, and expansion joint dimensions.

3. The method for predicting concrete damage in expansion joints based on multi-task deep learning according to claim 1, characterized in that, The multi-task MLP model includes a shared feature extraction layer and three output heads. The shared feature extraction layer is used to extract shared features, and the output heads are used to classify damage location, damage level, or damage degree based on the shared features.

4. The method for predicting concrete damage in expansion joints based on multi-task deep learning according to claim 3, characterized in that, The shared feature extraction layer consists of multiple fully connected blocks, each of which is used to perform linear transformation, ReLU activation, and random activation operations in sequence.

5. The method for predicting concrete damage in expansion joints based on multi-task deep learning according to claim 4, characterized in that, The number of hidden layers and neurons in the shared feature extraction layer are adaptively determined by Optuna; the hyperparameter tuning using tri-fold cross-validation specifically involves: achieving hyperparameter tuning through random search of Optuna hyperparameters combined with tri-fold cross-validation.

6. The method for predicting concrete damage in expansion joints based on multi-task deep learning according to claim 1, characterized in that, In the hyperparameter tuning using triple cross-validation, the formula for the weighted combination index of hyperparameter search is: in, This represents a weighted combination index of the comprehensive damage degree regression, damage level classification, and damage location classification tasks under a certain fold cross-validation in the hyperparameter search. The mean absolute error of the damage severity regression task; Represents the weighted F1 score for the damage severity classification task; This represents the weighted F1 score for the damage location classification task.

7. The method for predicting concrete damage in expansion joints based on multi-task deep learning according to claim 1, characterized in that, LogCosh was selected from MSE, Huber, and LogCosh as the function for the regression loss of damage level using an automatic hyperparameter search method.

8. The method for predicting concrete damage in expansion joints based on multi-task deep learning according to claim 1, characterized in that, The method for predicting concrete damage in expansion joints also includes the step of evaluating the prediction model using multi-dimensional evaluation indicators.

9. The method for predicting concrete damage in expansion joints based on multi-task deep learning according to claim 8, characterized in that, For the prediction of damage location and damage level, accuracy, precision, recall, and F1 score are used as evaluation indicators; for the prediction of damage severity, mean absolute error, root mean square error, and coefficient of determination are used as evaluation indicators.

10. A concrete damage prediction system for expansion joints based on multi-task deep learning, characterized in that, include: The data acquisition module is used to collect bridge structural parameters based on bridge design drawings. The finite element model construction module is used to establish a finite element model of the bridge based on the bridge structural parameters, wherein the concrete in the expansion joint anchorage zone is self-sensing concrete; it is used to simulate the piezoresistive effect of the self-sensing concrete to obtain a force-electric coupling finite element model; it is used to set multi-level damage degree coefficients, and for each damage location of each damage degree coefficient, it calculates the current change rate before and after the damage, and obtains the original data set including bridge structural parameters, current change rate, damage location corresponding to the current change rate, and damage degree coefficient corresponding to the current change rate. The data preprocessing module is used to preprocess the data based on the original dataset. Specifically, it is used to: construct a feature matrix composed of the current change rate; use the damage location as a classification label, extract continuous floating-point values ​​of the damage degree as regression labels, and map the damage degree to the loss level to generate auxiliary classification labels; use stratified sampling to divide the training set, validation set, and test set, with the damage location as a stratification variable; perform feature standardization on the bridge structural parameters and current change rate, and perform regression target standardization on the regression labels. The training module is used to train a multi-task MLP model with a shared underlying layer to obtain a prediction model that can simultaneously predict the location, degree, and level of concrete damage in the expansion joint anchorage zone. Specifically, it is used for: feature learning through joint damage degree regression loss, cross-entropy-based loss level classification loss, and focus-based loss damage location classification loss. The function for damage degree regression loss is selected from MSE, Huber, and LogCosh using an automatic hyperparameter search method; hyperparameter tuning is performed using three-fold cross-validation; and hyperparameters are determined using an early stopping mechanism. The prediction module is used to obtain the bridge structural parameters of the bridge to be predicted and the current change rate of the expansion joint concrete, and to predict the damage of the expansion joint concrete using the prediction model.

Images