Bridge scour damage identification method and system based on deep learning

Abstract

The present invention discloses a method and system for identifying bridge scour damage based on deep learning, which solves the problems of poor universality of current bridge scour damage identification methods and limited efficiency and accuracy in scour damage identification. The method includes the steps of: constructing an original data set; data preprocessing; constructing a sequence-to-sequence multi-parameter model with an encoder-decoder architecture using a two-layer long short-term memory network, and adding a Dropout layer and a LayerNorm layer standardization to the model; defining a loss function for model training and optimization, using the Adam optimization algorithm with L2 regularization and using the early stopping method to train and optimize the Seq2Seq multi-parameter model, and selecting the model with the smallest loss in the validation set as the best prediction model for identifying bridge scour damage. The present invention can quickly identify complex scour damage conditions on multi-span bridges, improve the efficiency and accuracy of scour damage identification, and has stronger universality and higher recognition efficiency.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a bridge scour damage identification method and system based on deep learning. Background Art

[0002] Bridges are inevitably subject to scour damage caused by floods, prolonged water flow, and other factors. Scour is a major factor in bridge failure. Scour weakens the pile-soil restraint, reducing the bearing capacity of the bridge structure and seriously threatening its operational safety. Therefore, developing a rapid identification method for bridge scour damage, which can quickly and accurately identify scour locations and uneven scour damage, is crucial for ensuring the safe operation of bridge structures.

[0003] Currently, physics-based bridge pier scour detection technology relies on underwater operations and on-site testing. These technologies suffer from limitations such as timeliness, low accuracy, high equipment costs, complex installation processes, high risks for testing personnel, and susceptibility of monitoring data to environmental interference. These limitations pose significant challenges to direct scour measurement. In fact, scour directly reduces the buried depth of bridge foundations, altering substructure boundary conditions and subsequently changing the stiffness of the bridge structure, significantly affecting its dynamic characteristics. Consequently, researchers have begun to focus on identifying bridge scour damage based on dynamic characteristics.

[0004] For example, the paper "Bridge Pier Scour State Analysis Based on Dynamic Characteristics Identification" published in the journal "China Journal of Highway and Transport" proposed obtaining a structural flexibility matrix based on the conversion of structural natural frequency and vibration mode, constructing the "structural calculated displacement difference" as a scour damage identification indicator, and realized bridge scour damage identification using a simply supported beam bridge as an example.

[0005] For example, the Chinese patent with publication number CN119272568A, titled "A method for indirect identification of deep-water bridge foundation scour depth," proposes a method for identifying the foundation scour depth of a high-speed railway train-track-cross-sea bridge system. By fitting, a functional relationship between characteristic indicators and scour depth is established, thereby realizing the prediction of foundation scour depth.

[0006] For example, the Chinese patent with publication number CN111353238A, entitled "A method for identifying bridge pier scour depth based on vehicle sensing," proposes a method for identifying the scour depth of bridge piers using the measured fundamental frequency of the bridge. The measured fundamental frequency of the bridge extracted from the vehicle acceleration spectrum is substituted into the functional relationship between the bridge fundamental frequency and the scour depth of the bridge pier to calculate the scour depth of the bridge pier.

[0007] For example, the Chinese patent with publication number CN114036974A, entitled "A Bridge Scour Dynamic Identification Method Based on Health Monitoring Data," proposes a method for dynamically identifying foundation scour depth by analyzing the dynamic characteristics of the structural system. The method uses a health monitoring system to collect the acceleration-time curve of each bridge foundation structure when it vibrates under scour conditions, and calculates the time-frequency characteristics of the reference mode, thereby dynamically identifying the scour depth.

[0008] With the rapid development of computer science and artificial intelligence technology, bridge scour damage identification based on dynamic characteristics and intelligent algorithms has gradually become a research hotspot.

[0009] For example, the paper "Research on Bridge Scour Depth Identification Method Based on Support Vector Machine" published in the Journal of Water Resources and Architectural Engineering uses a feature system implementation algorithm to extract bridge modal parameters from bridge vibration response data under environmental excitation, and uses the support vector machine algorithm to establish a nonlinear mapping relationship between modal parameters and scour depth to realize the scour depth identification of the bridge.

[0010] For example, the paper "Bridge Damage Identification Based on Dynamic Fingerprint and Bayesian Data Fusion" published in the Journal of Northeastern University (Natural Science Edition) integrates Bayesian theory with multi-source monitoring data, and uses modal parameter changes to perform scour damage inversion, with an identification accuracy of over 85%.

[0011] For example, the paper "Bridge Scour Dynamic Assessment Based on Time-Frequency Analysis and Neural Network" published in the Journal of Tianjin University (Natural Science and Engineering Technology Edition) uses wavelet transform to extract the time-frequency characteristics of bridge vibration signals, and uses BP neural network to establish a mapping relationship between scour damage and bridge vibration characteristics, thereby realizing the synchronous assessment of scour damage and location.

[0012] For example, the paper "Bridge Scour Depth Identification Based on Dynamic Characteristics and Improved Particle Swarm Optimization Algorithm" published in the Journal of Jilin University (Engineering Edition) uses the dynamic characteristics of the bridge structure and the static displacement of the pier top under unit force as bridge scour depth identification parameters. With the help of an improved particle swarm optimization algorithm, a quantitative correlation is established between the scour depth identification parameters and the scour damage of different bridge piers, thereby identifying the different scour depths of each bridge pier.

[0013] The above research validates the feasibility of identifying scour damage and its location based on dynamic characteristics, providing a useful reference for the rapid identification of scour damage in complex bridge structures. However, existing methods generally rely on large amounts of measured data or are limited to a single specific bridge scenario, lacking universal applicability. Furthermore, there is limited research on identifying scour locations and uneven scour at different locations on multi-span bridges. Consequently, there is no rapid and accurate prediction method for complex scour scenarios on multi-span bridges. Summary of the Invention

[0014] In view of the problems that the current bridge scour damage identification method lacks universality due to the significant differences in bridge design parameters in actual projects, and the efficiency and accuracy of identifying scour locations and uneven scour damage at each location on multi-span bridges need to be further improved, the present invention provides a bridge scour damage identification method and system based on deep learning. By comprehensively considering the geometric characteristics of bridges, soil parameters and complex scour scenarios of multi-span bridges, combined with deep learning algorithms, a universal bridge scour damage identification method and system are formed, which can achieve rapid and accurate prediction of bridge scour locations and uneven scour damage, and effectively ensure the safe operation of bridge structures.

[0015] In order to solve the above problems, the present invention adopts the following technical solutions:

[0016] A bridge scour identification method based on deep learning, the method comprising the following steps:

[0017] A bridge scour damage identification method based on deep learning includes the following steps:

[0018] Step S100: A sampling algorithm is used to obtain combined working conditions of bridge parameters, soil parameters, scour locations of multi-span bridges, and uneven scour damage, and finite element software is used to calculate scour damage identification parameters under each working condition, thereby constructing an original data set containing bridge parameters, soil parameters, scour locations of multi-span bridges, uneven scour damage, and scour damage identification parameters;

[0019] Step S200: pre-processing the data in the original data set to obtain an input data set for rapid identification of bridge scour damage;

[0020] Step S300: A sequence-to-sequence multi-parameter model with an encoder-decoder architecture is constructed using a two-layer long short-term memory network. Dropout and LayerNorm layers are added to the sequence-to-sequence multi-parameter model for normalization. The encoder receives an input sequence including fixed features and dynamic features, compresses the input sequence into a hidden state and a cell state using a layer of long short-term memory network, and then processes the input sequence. The cell state features are output to the decoder. The decoder generates output features using another layer of long short-term memory network and uses a fully connected layer to map the cell state features output by the encoder into scour damage prediction results.

[0021] Step S400: Define the loss function for model training and optimization, use the Adam optimization algorithm with L2 regularization, and use the early stopping method to train and optimize the Seq2Seq multi-parameter model. Select the model with the smallest loss in the validation set as the best prediction model, which is used to identify bridge scour damage.

[0022] Accordingly, the present invention also proposes a bridge scour damage identification system based on deep learning, which includes:

[0023] The data acquisition module is used to obtain the combined working conditions of bridge parameters, soil parameters, scour locations of multi-span bridges, and uneven scour damage using a sampling algorithm. The scour damage identification parameters under each working condition are calculated using finite element software, and the original data set containing bridge parameters, soil parameters, scour locations of multi-span bridges, uneven scour damage, and scour damage identification parameters is constructed.

[0024] The preprocessing module is used to preprocess the data in the original data set to obtain the input data set for rapid identification of bridge scour damage;

[0025] A model construction module is used to construct a sequence-to-sequence multi-parameter model with an encoder-decoder architecture using a two-layer long short-term memory network. Dropout and LayerNorm layers are added to the sequence-to-sequence multi-parameter model for standardization. The encoder receives an input sequence consisting of fixed and dynamic features, compresses the input sequence into a hidden state and cell state using a layer of long short-term memory network, and then processes the input sequence. The cell state features are then output to the decoder. The decoder generates output features using another layer of long short-term memory network and uses a fully connected layer to map the cell state features output by the encoder into scour damage prediction results.

[0026] The training and optimization module is used to define the loss function for model training and optimization. It uses the Adam optimization algorithm, adds L2 regularization, and utilizes the early stopping method to train and optimize the Seq2Seq multi-parameter model. The model with the smallest loss in the validation set is selected as the optimal prediction model, which is used to identify bridge scour damage.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] (1) Provides a universal bridge scour damage identification method: The present invention comprehensively considers bridge parameters, soil parameters, and complex scour scenarios of multi-span bridges. The parameter types are comprehensively considered and the range is reasonably set. The large sample data generated can include various bridge structures. The rate of change of the first-order natural frequency of the bridge and the first-order mode ratio at the top of the pier are used as scour damage identification parameters, which can reflect the scour location and uneven scour damage of multi-span bridges. Based on large sample data and with the help of deep learning algorithms, the complex implicit relationship between scour damage and various parameters is accurately established, forming a universal scour damage identification method;

[0029] (2) Improved efficiency of scour damage identification: The present invention adopts a deep learning-based method to establish a quantitative mapping relationship between bridge parameters, soil parameters, scour locations of multi-span bridges, and scour damage at each scour location through steps such as data sample acquisition, data preprocessing, initial model construction, training, and optimization. This method can directly identify scour locations and scour depths based on changes in bridge dynamic characteristics, providing an efficient solution for rapid inversion of bridge scour damage.

[0030] (3) Improved accuracy of scour damage identification: The present invention adopts a Seq2Seq multi-parameter model based on a double-layer LSTM, and utilizes its gating mechanism to flexibly process variable-length sequences, effectively capturing temporal dependencies and the complex nonlinear relationship between scour damage and various parameters. It can accurately invert the different scour depths at various locations of the bridge according to changes in the dynamic characteristics of the bridge, and achieve accurate prediction of the scour location and uneven scour damage of multi-span bridges;

[0031] (4) Enhanced model generalization ability: The present invention uses the Adam optimizer to accelerate model convergence and reduce training time. At the same time, the Dropout regularization and early stopping method are applied to effectively prevent overfitting, enhance model robustness and model training efficiency, and further improve the model generalization ability, so that it can be better applied to the bridge scour damage identification task requirements under different conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings:

[0033] Figure 1 is a flow chart of a bridge scour damage identification method in an embodiment of the present invention;

[0034] Figure 2 This is a flow chart of data preprocessing in an embodiment of the present invention;

[0035] Figure 3 This is a prediction effect diagram of the Seq2Seq multi-parameter model in an embodiment of the present invention;

[0036] Figure 4 Graph showing the loss function learning curves of the models in the embodiments of the present invention;

[0037] Figure 5 : is a comparison chart of the residual errors of the prediction results of each model in the embodiment of the present invention; DETAILED DESCRIPTION

[0038] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. It should be understood that the specific embodiments described are only used to explain this application and are not used to limit this application.

[0039] like Figure 1 As shown, this embodiment provides a bridge scour damage identification method based on deep learning, which includes the following steps S100 to S400.

[0040] Step S100: Constructing an original data set. The original data involved in the original data set include bridge parameters, soil parameters, scour locations of multi-span bridges, uneven scour damage, and scour damage identification parameters.

[0041] Step S101, obtaining bridge parameter and soil parameter samples: Due to the large number of water-related bridges in my country and their diverse structural forms, in order to make the identification method of the present invention universal, the parameter type and parameter range must be fully considered when obtaining samples. The bridge parameters selected in this embodiment include the number of bridge spans, ... , span, pier shear span ratio, pile slenderness ratio and substructure elastic modulus, soil parameters include internal friction angle and soil density. Bridge parameters and soil parameters fluctuate within a certain range. The fluctuation range of each parameter is determined according to relevant standards such as "Highway Bridge Seismic Design Code" and "Highway Bridge and Culvert Foundation and Foundation Design Code". In this embodiment, the number of bridge spans is 100. The range of the span is 2 to 5, the span follows a log-normal distribution with a mean of 33.325 m and a coefficient of variation of 27.5%, the shear span ratio of the pier ranges from 2.5 to 10, the slenderness ratio of the pile ranges from 20 to 40, the elastic modulus of the substructure follows a normal distribution with a mean of 26.9 GPa and a coefficient of variation of 9.17%, the internal friction angle of the soil follows a normal distribution with a mean of 30° and a coefficient of variation of 10%, and the soil weight follows a normal distribution with a mean of 18.8 kN / m 3 The lognormal distribution has a coefficient of variation of 2.5%. To reduce computational costs for multiple parameter combinations, the Latin Hypercube Sampling (LHS) algorithm was used to sample bridge and soil parameters. This generated a large number of bridge model parameter combinations, ensuring that the parameter combinations evenly covered the entire variable space. A total of 100 sets of bridge and soil parameter samples were generated.

[0042] Step S102: Obtain a combined sample of scour locations and uneven scour damage on a multi-span bridge: The number of bridge spans selected in this embodiment is The number of scour damage combinations ranges from 2 to 5, so the sample includes bridges with two to five spans, with each span having a different number of piers. To fully account for the uneven depths at different locations on each span, the scour damage combinations for bridges with different spans are assumed to be independent. The LHS algorithm is used to independently sample the scour damage (scour depth) for bridges with two, three, four, and five spans. For example, 10 scour damage combination conditions are sampled for the three piers of a two-span bridge, 15 scour damage combination conditions are sampled for the four piers of a three-span bridge, 20 scour damage combination conditions are sampled for the five piers of a four-span bridge, and 25 scour damage combination conditions are sampled for the six piers of a five-span bridge.

[0043] Then, the 100 sets of bridge parameter and soil parameter samples obtained in step S101 are combined with the scour damage combination conditions of bridges with different spans to generate a combination condition including bridge parameters, soil parameters, scour locations of multi-span bridges, and scour damage, forming a total of 1,850 sets of condition samples.

[0044] Step S103, obtaining scour damage identification parameter samples: In order to achieve accurate positioning of the scour location and accurate identification of scour damage, this embodiment uses the bridge first-order natural frequency change rate Ratio of the first-order vibration mode to the pier top position As the scour damage identification parameter, the superscript Indicates the A bridge pier.

[0045] The rate of change of the first-order natural frequency of a bridge is easy to obtain, has universal significance, and can reflect the decrease in bridge stiffness caused by scouring. Its calculation formula is:

[0046] ;

[0047] in, is the first-order natural frequency of the bridge after scour damage; is the first-order natural frequency of the bridge without scour damage.

[0048] The first-order mode ratio at the pier top is the ratio of the vibration amplitude of each pier top to the maximum amplitude of the pier top in the overall vibration mode of the bridge structure. It can effectively reflect the relative stiffness changes caused by the inconsistent scour damage of each pier. Assume that the vibration amplitudes of the first-order mode at the top of each pier of the N-span bridge are , the maximum amplitude of the pier top is defined as , then The first-order mode ratio at the top of each pier can be expressed as:

[0049] ;

[0050] in, For the The first-order vibration mode ratio at the top of each pier is: For the The vibration amplitude of the first-order vibration mode at the top of each pier.

[0051] A parametric bridge numerical analysis model accounting for pile-soil interaction was established using OpenSees finite element software. Scour damage identification parameters were calculated for each combined operating condition based on this parametric bridge numerical analysis model. For the 1,850 operating condition samples generated in this example, encompassing bridge parameters, soil parameters, scour locations, and scour damage, the corresponding first-order natural frequency change rate and first-order mode ratio at the pier top were calculated for each corresponding operating condition. These values ​​were then compared with the natural frequency change rate and mode ratios at each pier top location in the first-order mode under the non-scour condition to obtain scour damage identification parameter samples. The samples of bridge and soil parameters, the combined samples of scour locations and uneven scour damage across multiple spans, and the scour damage identification parameter samples together constituted a large-scale original dataset.

[0052] Step S200: Data preprocessing. The data samples in the original dataset constructed in step S100 are preprocessed. The preprocessing process includes data loading, missing value filling, feature normalization, dataset partitioning, and data conversion. After processing, the input dataset for rapid identification of bridge scour damage is obtained.

[0053] The 1850 sets of raw data in this embodiment are obtained through step S100, which include both fixed feature data (such as the number of spans, span diameter, shear span ratio of piers, etc.) and dynamic feature data (such as the first-order mode ratio at the pier top). In addition, due to the different number of spans of bridges, the dimension of the target variable (i.e., scour depth) is inconsistent. Therefore, before scour damage identification, the raw data samples need to be processed and converted into a format suitable for model input. The specific data preprocessing process is as follows: Figure 2 As shown, the process includes the following steps S201 to S205.

[0054] Step S201, data loading: bridge parameters (number of spans, span diameter, pier shear span ratio, pile slenderness ratio, substructure elastic modulus), soil parameters (internal friction angle, soil density), bridge first-order natural frequency change rate as fixed features, and the first-order mode ratio at the pier top As a dynamic feature, get the maximum number of bridge spans in the original data set , for the maximum span of the bridge The corresponding bridge generates the dynamic target variable scour depth of each pier of the bridge , used to train the Seq2Seq multi-parameter model. .

[0055] Step S202, missing value filling: Since bridges with different spans have different numbers of piers, the amount of uneven scour damage and the first-order mode ratio at the pier top will be different. To ensure that the training and prediction of the model are not affected by the different dimensions of the data, the missing values ​​of the uneven scour damage and the first-order mode ratio at the pier top in different dimensions in the original data set are filled with 0.

[0056] Step S203, feature normalization: In order to eliminate the dimensional differences between different features, the fixed features and dynamic features are normalized separately. Since the feature scales involved vary greatly, the Z-score standard normalization method is used to eliminate the dimensional differences, that is, the data in the original data set is converted to a standard normal distribution with zero mean and unit variance to accelerate the convergence process of gradient descent. The formula of the Z-score standard normalization method is:

[0057] ;

[0058] in, is the normalized value; is the original data value; is the mean of the feature; is the standard deviation of the feature.

[0059] Step S204, data set division: First, the data set is divided into a training set and a test set. The training set accounts for 85% of the original data set and is used for model training; the test set accounts for 15% of the original data set and is used to evaluate model performance; then 20% of the training set is divided as a validation set, that is, a validation set of 17% of the total data. The validation set is used to monitor the model in real time.

[0060] Step S205, data conversion: convert the features of the training set and the test set and the generated dynamic target variable flushing depth into PyTorch tensors.

[0061] Step S300: Seq2Seq Multi-Parameter Model Construction. This step uses a two-layer long short-term memory (LSTM) network to build a sequence-to-sequence (Seq2Seq) multi-parameter model with an encoder-decoder architecture to handle the variable-length sequences caused by the varying number of spans in multi-span bridges. Dropout and LayerNorm layers are added to the Seq2Seq multi-parameter model for normalization.

[0062] Encoder: The encoder receives an input sequence. In this embodiment, the input sequence contains fixed features and dynamic features. Among them, the number of spans, span diameter, pier shear span ratio, pile slenderness ratio, substructure elastic modulus, internal friction angle, soil weight and first-order natural frequency change rate are fixed features with a dimension of 8; the first-order vibration mode ratio at the pier top is 1. is a dynamic feature with a dimension of Therefore, the dimension of the input sequence received by the encoder is Then, a layer of LSTM network is used to compress the input sequence into a hidden state and cell state, and then process it, outputting the cell state features to the decoder.

[0063] Decoder: Receives the cell state features output by the encoder and generates output features with the help of another LSTM network layer. Then, a fully connected layer is used to map the cell state features output by the encoder into the prediction results of scour damage. Taking the example of a bridge with 2 to 5 spans in this embodiment, the scour damage prediction result of the 2-span bridge is and , the scour damage prediction results of the 5-span bridge are and .

[0064] Dropout is a commonly used regularization technique that prevents overfitting in neural networks by randomly discarding some neurons, thereby enhancing the model's generalization ability. In this example, a dropout layer was added to both layers of the LSTM network, with a dropout rate set to 0.2. This randomly discards 20% of the neuron outputs during training (setting their outputs to 0), forcing the network to learn more robust features.

[0065] Adding LayerNorm normalization to the Seq2Seq multi-parameter model addresses distribution variations across LSTM network layers. This standardizes the feature dimension of individual samples, stabilizing the input distribution across layers. This allows for a larger learning rate and prevents vanishing or exploding gradients. Furthermore, the use of two LSTM layers in this example can cause the activation values ​​to scale gradually. Adding LayerNorm normalization can prevent this cumulative scale drift.

[0066] The Seq2Seq multi-parameter model based on a two-layer LSTM performs a comprehensive analysis of data such as bridge parameters, soil parameters, complex scour scenarios for multi-span bridges, and the dynamic characteristics of bridge structures. It extracts underlying patterns and regularities from this data, enabling accurate prediction of scour locations and uneven scour damage at each location on multi-span bridges. This deep learning-based rapid bridge scour damage identification method offers advantages such as high automation, efficient processing, and high prediction accuracy. It can simultaneously identify scour damage and its location with high precision, providing real-time, reliable data support for bridge management departments and significantly improving bridge maintenance efficiency.

[0067] Step S400: Model training and optimization. The Seq2Seq multi-parameter model established in step S300 is trained and optimized, and then validated on a validation set. The model with the smallest validation set loss is selected as the optimal prediction model, which is used to accurately identify bridge scour damage.

[0068] Definition of loss function: This embodiment uses mean squared error (MSE) as the loss function, and its calculation formula is:

[0069] ;

[0070] in, is the true value; is the predicted value; is the sample size.

[0071] The Adam optimization algorithm is used to train the Seq2Seq multi-parameter model established in step S300. This algorithm combines the advantages of the momentum method and the adaptive learning rate and can effectively accelerate the convergence of the model. In this embodiment, the learning rate of the Adam optimization algorithm is set to 0.001.

[0072] Add L2 regularization: The Seq2Seq multi-parameter model established in step S300 contains two LSTM layers and a fully connected layer. It has a large number of parameters and is prone to overfitting. Therefore, L2 regularization is added to prevent overfitting of the model, improve generalization ability, smooth the loss function, and improve numerical stability. L2 regularization is achieved by adding the sum of squares of weights to the loss function. Assume that the original loss function is ,in Represents the model parameters, then the loss function after L2 regularization is added for:

[0073] ;

[0074] in, is the original loss function; is the regularization coefficient; is the sum of squares of all model weights, are the parameters such as weights and biases that need to be learned and optimized in the model. is the number of parameters in the model.

[0075] Early stopping is a method that monitors the validation set loss to determine whether to terminate training early. By monitoring the validation set loss and stopping training if the validation set loss does not decrease within consecutive epochs, it effectively prevents model overfitting and improves the model's generalization ability.

[0076] Furthermore, after step S400, the following steps are further included:

[0077] Step S500: Evaluate the best prediction model using evaluation indicators, where the evaluation indicators include but are not limited to mean absolute error (MAE), mean square error (MSE), and coefficient of determination (R²).

[0078] MAE represents the mean absolute difference between the predicted value and the true value, that is, the average error of the result. MSE calculates the squared average of the prediction error, which is more sensitive to large errors and can amplify the impact of outliers. 2 This measure measures how well the model fits the data, typically ranging from 0 to 1. The prediction results for each evaluation metric on the test set using the method in this embodiment were: MAE of 0.0079, MSE of 0.0001, and R² of 0.9983. The evaluation results demonstrate that the model exhibits high prediction accuracy across all evaluation metrics.

[0079] Then, the predicted values ​​of the test set are compared with the actual values, and the results are as follows: Figure 3 As shown in Figure 2, the actual scour depth and the predicted scour depth are distributed diagonally, indicating that the model can well predict the scour damage of all test set samples with a small error.

[0080] To further verify the accuracy and superiority of the Seq2Seq multi-parameter model in this embodiment, the Seq2Seq multi-parameter model was compared with the Transformer model, convolutional neural network (CNN), random forest (RandomForest) and XGBoost traditional machine learning models on the training set and validation set. First, the loss function learning curves of the Transformer model, CNN model and the Seq2Seq multi-parameter model in this embodiment were plotted, as shown in Figure 2. Figure 4As shown in the figure, it can be seen that the Seq2Seq multi-parameter model shows a trend of rapid decline and then stabilization, and the loss function of the training set and validation set is significantly lower than that of other models, which proves the superiority of the model. Then, by calculating MSE, MAE and R 2 The prediction performance of each model was evaluated using indicators such as LSTM and CNN. The results are shown in Table 1. It can be seen that the Seq2Seq multi-parameter model using a two-layer LSTM encoder-decoder architecture significantly outperforms other comparison models in all evaluation indicators.

[0081] Table 1 Comparison of prediction results of various algorithms

[0082]

[0083] In addition, in order to intuitively compare the prediction effects of each model, this embodiment calculates the residuals between the prediction results of each model and the actual value and plots them as violin plots, as shown in Figure 5 As shown by Figure 5 It can be seen that the Seq2Seq multi-parameter model has the smallest residual, that is, the deviation between the prediction result and the theoretical value is small, and the residual distribution is more concentrated, which proves that its prediction accuracy is the highest.

[0084] Another embodiment of the present invention provides a bridge scour damage identification system based on deep learning, which specifically includes:

[0085] The data acquisition module is used to obtain the combined working conditions of bridge parameters, soil parameters, scour locations of multi-span bridges, and uneven scour damage using a sampling algorithm. The scour damage identification parameters under each working condition are calculated using finite element software, and the original data set containing bridge parameters, soil parameters, scour locations of multi-span bridges, uneven scour damage, and scour damage identification parameters is constructed.

[0086] The preprocessing module is used to preprocess the data in the original data set to obtain the input data set for rapid identification of bridge scour damage;

[0087] A model construction module is used to construct a sequence-to-sequence multi-parameter model with an encoder-decoder architecture using a two-layer long short-term memory network. Dropout and LayerNorm layers are added to the sequence-to-sequence multi-parameter model for standardization. The encoder receives an input sequence consisting of fixed and dynamic features, compresses the input sequence into a hidden state and cell state using a layer of long short-term memory network, and then processes the input sequence. The cell state features are then output to the decoder. The decoder generates output features using another layer of long short-term memory network and uses a fully connected layer to map the cell state features output by the encoder into scour damage prediction results.

[0088] The training and optimization module is used to define the loss function for model training and optimization. It uses the Adam optimization algorithm, adds L2 regularization, and utilizes the early stopping method to train and optimize the Seq2Seq multi-parameter model. The model with the smallest loss in the validation set is selected as the optimal prediction model, which is used to identify bridge scour damage.

[0089] The implementation method of the specific functions of each module in the bridge scour damage identification system of this embodiment can refer to the implementation method described in the above-mentioned bridge scour damage identification method embodiment, and will not be repeated here.

[0090] In summary, the present invention provides a method and system for rapid identification of bridge scour damage based on deep learning. Through the steps of data sample acquisition, data preprocessing, initial model construction, training and optimization, a reliable bridge scour damage identification model is constructed, which enables accurate identification of complex scour scenarios in multi-span bridges and provides reliable data support for bridge management departments. The present invention adopts advanced technologies such as deep learning, combines a variety of data processing methods and model building techniques, and can simultaneously identify scour damage and location, thereby improving the efficiency and accuracy of scour damage identification and significantly improving bridge maintenance efficiency. Compared with traditional scour identification methods, the bridge scour damage identification method provided by the present invention has faster recognition efficiency and higher recognition accuracy, and has good engineering applicability and economy.

[0091] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned 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.

[0092] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A bridge scour damage identification method based on deep learning is characterized by: The following steps are involved: Step S100: A sampling algorithm is used to obtain combined working conditions of bridge parameters, soil parameters, scour locations of multi-span bridges, and uneven scour damage, and finite element software is used to calculate scour damage identification parameters under each working condition, thereby constructing an original data set containing bridge parameters, soil parameters, scour locations of multi-span bridges, uneven scour damage, and scour damage identification parameters; Step S200: pre-processing the data in the original data set to obtain an input data set for rapid identification of bridge scour damage; Step S300: A sequence-to-sequence multi-parameter model with an encoder-decoder architecture is constructed using a two-layer long short-term memory network. Dropout and LayerNorm layers are added to the sequence-to-sequence multi-parameter model for normalization. The encoder receives an input sequence including fixed features and dynamic features, compresses the input sequence into a hidden state and a cell state using a layer of long short-term memory network, and then processes the input sequence. The cell state features are output to the decoder. The decoder generates output features using another layer of long short-term memory network and uses a fully connected layer to map the cell state features output by the encoder into scour damage prediction results. Step S400: Define the loss function for model training and optimization, use the Adam optimization algorithm with L2 regularization, and use the early stopping method to train and optimize the Seq2Seq multi-parameter model. Select the model with the smallest loss in the validation set as the best prediction model, which is used to identify bridge scour damage.

2. The bridge scour damage identification method based on deep learning according to claim 1 is characterized in that: Step S100 includes the following steps: Step S101: Bridge parameters include the number of spans, span diameter, pier shear span ratio, pile slenderness ratio, and substructure elastic modulus; soil parameters include the internal friction angle and soil density. A sampling algorithm is used to sample the bridge parameters and soil parameters to generate bridge parameter and soil parameter samples. Step S102: Using the same sampling algorithm as step S101, the non-uniform scour damage of bridges with different spans is independently sampled to generate scour damage combination conditions. The bridge parameter and soil parameter samples are then combined with the different scour damage combination conditions to generate a combination condition including bridge parameters, soil parameters, scour locations of multi-span bridges, and scour damage. Step S103: Use finite element software to establish a parameterized bridge numerical analysis model that considers pile-soil interaction. Calculate the scour damage identification parameters for each combined working condition obtained in step S102 based on the parameterized bridge numerical analysis model. The scour damage identification parameters include the rate of change of the first-order natural frequency of the bridge and the first-order mode ratio at the pier top.

3. The bridge scour damage identification method based on deep learning according to claim 2 is characterized in that: Change rate of the first-order natural frequency of the bridge The calculation formula is: ； in, is the first-order natural frequency of the bridge after scour damage; is the first-order natural frequency of the bridge without scour damage; No. The first-order vibration mode ratio at the top of the pier The calculation formula is: ； in, For the The vibration amplitude of the first-order vibration mode at the top of each pier, is the number of bridge spans.

4. The bridge scour damage identification method based on deep learning according to claim 2 is characterized in that: The number of bridge spans ranges from 2 to 5, and the span follows a log-normal distribution with a mean of 33.325 m and a coefficient of variation of 27.5%. The shear span ratio of the piers ranges from 2.5 to 10, and the slenderness ratio of the piles ranges from 20 to 40. The elastic modulus of the substructure follows a normal distribution with a mean of 26.9 GPa and a coefficient of variation of 9.17%. The internal friction angle follows a normal distribution with a mean of 30° and a coefficient of variation of 10%. The soil density follows a normal distribution with a mean of 18.8 kN / m. 3 The data were lognormally distributed with a coefficient of variation of 2.5%.

5. The bridge scour damage identification method based on deep learning according to claim 2 is characterized in that: The sampling algorithm adopts Latin hypercube sampling algorithm, and the finite element software adopts OpenSees finite element software.

6. The bridge scour damage identification method based on deep learning according to any one of claims 1 to 5, characterized in that: Step S200 includes the following steps: Step S201: Using the bridge parameters, soil parameters, and the rate of change of the first-order natural frequency of the bridge as fixed features, and the first-order mode ratio at the pier top as a dynamic feature, the dynamic target variable scour depth of each pier is generated for the bridge corresponding to the maximum number of bridge spans in the original data set. ; Step S202: Fill missing values ​​of different parts of the dimension of the uneven scour damage and the first-order mode ratio at the pier top in the original data set with 0; Step S203: using the Z-score standard normalization method to convert the data in the original data set into a standard normal distribution with zero mean and unit variance; Step S204: Divide the normalized dataset into a training set for model training and a test set for evaluating model performance, with a ratio of 85% and 15% respectively. Separately, 20% of the training set is used as a validation set for real-time monitoring of the model. Step S205: Convert the features of the training set and the test set and the generated dynamic target variable flush depth into PyTorch tensors.

7. The bridge scour damage identification method based on deep learning according to any one of claims 1 to 5, characterized in that: The input sequence received by the encoder has the dimension ,in, is the maximum number of bridge spans in the original dataset; the Dropout rate is set to 0.

2.

8. The bridge scour damage identification method based on deep learning according to any one of claims 1 to 5, characterized in that: In step S400, the mean square error is used as the loss function; the learning rate of the Adam optimization algorithm is set to 0.001; the loss function after adding L2 regularization The calculation formula is: ； in, is the original loss function; is the regularization coefficient; is the sum of squares of all model weights, are the parameters that need to be learned and optimized in the model; is the number of parameters in the model.

9. The bridge scour damage identification method based on deep learning according to any one of claims 1 to 5, characterized in that: After step S400, the following steps are also included: Step S500: Evaluate the best prediction model using three evaluation indicators: mean absolute error, mean square error, and coefficient of determination.

10. The bridge scour damage identification system based on deep learning is characterized by: include: The data acquisition module is used to obtain the combined working conditions of bridge parameters, soil parameters, scour locations of multi-span bridges, and uneven scour damage using a sampling algorithm. The scour damage identification parameters under each working condition are calculated using finite element software, and the original data set containing bridge parameters, soil parameters, scour locations of multi-span bridges, uneven scour damage, and scour damage identification parameters is constructed. The preprocessing module is used to preprocess the data in the original data set to obtain the input data set for rapid identification of bridge scour damage; A model construction module is used to construct a sequence-to-sequence multi-parameter model with an encoder-decoder architecture using a two-layer long short-term memory network. Dropout and LayerNorm layers are added to the sequence-to-sequence multi-parameter model for standardization. The encoder receives an input sequence consisting of fixed and dynamic features, compresses the input sequence into a hidden state and cell state using a layer of long short-term memory network, and then processes the input sequence. The cell state features are then output to the decoder. The decoder generates output features using another layer of long short-term memory network and uses a fully connected layer to map the cell state features output by the encoder into scour damage prediction results. The training and optimization module is used to define the loss function for model training and optimization. It uses the Adam optimization algorithm, adds L2 regularization, and utilizes the early stopping method to train and optimize the Seq2Seq multi-parameter model. The model with the smallest loss in the validation set is selected as the optimal prediction model, which is used to identify bridge scour damage.

Images