Structural damage identification method based on data fusion and one-dimensional convolutional neural network

Abstract

The application discloses a structural damage identification method based on data fusion and a one-dimensional convolutional neural network, S1: obtaining vibration signals of a bridge structure under different damage states, and establishing an original data set, and performing normalization processing; S2: performing data enhancement processing in a sliding window mode; S3: establishing an improved one-dimensional convolutional neural network model, and performing training and testing to obtain the trained improved one-dimensional convolutional neural network model; S4: using the trained improved one-dimensional convolutional neural network model to perform structural damage identification on the bridge structure to obtain a plurality of decisions, and performing decision fusion processing on the plurality of decisions by using a data fusion technology to obtain a comprehensive result of structural damage detection. The application can fully mine damage information of the bridge structure from the vibration signal data by adding the data fusion technology, and solves the problem that the bridge structure damage detection result is affected under the influence of environmental interference, sensor instability and other factors.

Description

Technical Field

[0001] This invention relates to the field of structural damage identification technology, and in particular to a structural damage identification method based on data fusion and a one-dimensional convolutional neural network. Background Technology

[0002] With the rise of deep learning, new methods and hopes have been provided for efficiently processing big data. Among them, Convolutional Neural Networks (CNNs), as a type of deep neural network, have demonstrated superior capabilities in image classification and recognition, target localization and detection, and damage state identification. Damage identification methods combining CNNs with vibration signals have also been widely studied. In recent years, related research has also shown its excellent capabilities in the field of structural health monitoring, playing an important role in the health monitoring technologies of important structures such as bridges, high-rise buildings, offshore wind turbines, high-pile wharves, and steel frame structures.

[0003] During the identification phase, new response data of the structure is collected, and a trained neural network model is used to predict structural damage. However, even highly sophisticated neural network models cannot completely and accurately identify structural damage and will always have certain limitations. Under the influence of environmental interference, sensor instability, and other factors, the basic characteristics of some vibration signals can be easily altered, thus leading to the possibility of misjudging structural damage. Summary of the Invention

[0004] This invention provides a structural damage identification method based on data fusion and a one-dimensional convolutional neural network to overcome the technical problem that the basic characteristics of some vibration signals are easily altered under the influence of environmental interference, sensor instability and other factors, thus leading to misjudgment of structural damage.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] The structural damage identification method based on data fusion and one-dimensional convolutional neural networks includes the following steps:

[0007] S1: Vibration signals of the bridge structure under different damage states are acquired through several sensors, and an original dataset is established. The original dataset is then normalized to obtain a processed dataset.

[0008] S2: Divide the processed dataset into a training set and a test set, and perform data augmentation processing on the training set and the test set using a sliding window method;

[0009] S3: Establish an improved one-dimensional convolutional neural network model, train it based on the data-augmented training set, and test it based on the data-augmented test set to obtain the trained improved one-dimensional convolutional neural network model.

[0010] The improved one-dimensional convolutional neural network model includes a feature extraction module and a feature mapping module. The feature extraction module is used to extract features representing damage in the dataset after data augmentation and output them to the feature mapping module. The feature mapping module is used to establish a mapping relationship between the features representing damage and the damage state.

[0011] S4: Using the improved one-dimensional convolutional neural network model after training, structural damage identification is performed on the bridge structure to obtain several decisions. The several decisions are then fused using data fusion technology to obtain a comprehensive result of structural damage detection.

[0012] Specifically, in S2, the process of performing data augmentation on the training and test sets using a sliding window is as follows:

[0013] Several windows of size h×w are set up, where h represents the window length (time dimension) and w represents the window width (sensor dimension). The sliding window time step is set to s, and adjacent windows overlap. Then, each complete long-period data point in the training and test sets is divided into N identical sub-data samples, calculated using the following formula:

[0014]

[0015] In the formula, L is the length of the multidimensional time series, and f is the sampling frequency.

[0016] 3. The structural damage identification method based on data fusion and one-dimensional convolutional neural network according to claim 2, characterized in that, in S1, the calculation formula for normalizing the original dataset is:

[0017]

[0018] Where L0 is the original dataset, σ and σ(L0) are the mean and standard deviation of the original data, respectively. It is the processed dataset.

[0019] Specifically, in S3, the feature extraction module includes three feature extraction blocks with the same structure connected in sequence; each feature extraction block includes a one-dimensional convolutional layer, a batch normalization layer and a pooling layer. The one-dimensional convolutional layer is used to extract features representing damage in the input data and transmit them to the batch normalization layer. The batch normalization layer is used to perform batch normalization on the input data and transmit it to the pooling layer. The pooling layer is used to reduce the dimensionality of the input data.

[0020] The one-dimensional convolutional layer comprises K neurons, each performing filtering processing on the input data, as shown in the formula:

[0021] h k =f'(con1D(w k ,X)+b k )

[0022]

[0023] Among them, h k This is the result of the k-th convolution step; b k Indicates the offset coefficient; x i ,x i+1 ,…,x i+j Let represent the input data from step i, step i+1 to step i+h; h is the time dimension; f' is the non-linear activation function of the feature extraction module; X represents the input data of the one-dimensional convolutional layer, with an input dimension of (k′, w′, 1), where k′ is the number of data points contained in a single sensor dimension, w′ is the number of sensors deployed, and 1 is the number of channels; w k is the kernel weight vector of a one-dimensional convolutional layer; conv1D(·) is a one-dimensional convolution operator; This represents a vector multiplication operation;

[0024] The output data of the one-dimensional convolutional layer is normalized by a Batch Normalization (BN) layer. The calculation formula for the BN layer is as follows:

[0025]

[0026] x i ′=γh i +β≡BN γ,β (h i )

[0027] in, h represents the normalized feature data. i Represents the original feature data; μ B σ represents the average value of small batch B. B The standard deviation of the small batch B is represented by γ, the scaling factor by β, the offset factor by BN, and the normalization layer by x. i ′ represents the adjusted normalized data; ε is a constant.

[0028] Specifically, in S3, the feature mapping module includes a first fully connected layer, a ReLU activation function layer, a dropout layer, a second fully connected layer, and a softmax activation function layer connected in sequence.

[0029] The general calculation formulas for the first fully connected layer and the second fully connected layer are:

[0030] h j+1 =f"(∑h j ×wj+1 +b j+1 )

[0031] Among them, h j For the input data of the fully connected layer; w j+1 and b j+1 These represent the weights and additional bias terms, respectively; f” is the non-linear activation function of the feature mapping module; h j+1 This represents the output data of the fully connected layer, and j represents the j-th step.

[0032] Specifically, in S3, the improved one-dimensional convolutional neural network model determines the structural hyperparameters and training hyperparameters through a Bayesian optimization algorithm, including;

[0033] 1) Based on the structure of the improved one-dimensional convolutional neural network model, a set x containing several hyperparameter elements is obtained, including the number of convolutional layers, the number of filters in each layer, kernel size, stride, number of pooling layers, size, stride, padding, number of fully connected layers, number of neurons, normalization technique, Dropout layer parameters, activation function, optimizer, learning rate, batch size, and ratio of training set to test set;

[0034] 2) The search space of all parameters in set x is determined through experiments. The improved one-dimensional convolutional neural network model is initialized to obtain the initial parameter set D. It is then assumed that the initial parameter set D satisfies the established surrogate model M. Based on the prior distribution and the initial parameter set D, the posterior distribution is calculated. The acquisition function S is constructed using the posterior distribution. The next set of hyperparameters x1 is selected through the acquisition function S. The hyperparameter set x1 is substituted into the improved one-dimensional convolutional neural network model for training to obtain the network model evaluation value y1. After multiple iterations, when the loss function converges, the optimal hyperparameter combination of the improved one-dimensional convolutional neural network model is obtained.

[0035] Specifically, in S4, the process of obtaining the comprehensive result of structural damage detection by performing decision fusion processing on several decisions through data fusion technology includes:

[0036] When identifying the location of damage to a bridge structure, a voting method is used to calculate the comprehensive result of the structural damage detection. The calculation formula is as follows:

[0037]

[0038] Where H(x) represents the comprehensive structural damage detection result obtained using the voting method; T represents the number of classifier models, and the decision output of each classifier model is {h1, h2, ..., h...} T}; l is the number of category labels, and each category label is {c1, c2, ..., c l};C argmaxThe function to return the index of the maximum value. For the 1st, 2nd...lth output decision;

[0039] When identifying the location and extent of damage to a bridge structure, the mean value method is used to calculate the comprehensive result of the structural damage detection. The calculation formula is as follows:

[0040]

[0041] Where H(x)' represents the comprehensive structural damage detection result obtained using the mean method, T' represents the number of regression prediction models, and the decision output of each regression prediction model is {h1′,h2′,…,h T Let n' be the number of position change degree labels, and let the position change degree labels be {d1, d2, ..., dn'}. n}; This is the 1st, 2nd, ..., nth output decision.

[0042] Specifically, the voting method is any one of the following: absolute majority voting, relative majority voting, weighted voting, and soft voting.

[0043] Beneficial effects: This invention employs an improved one-dimensional convolutional neural network (CNN) model to extract spatial structure dimension features from multi-dimensional time-series data. Its computational efficiency surpasses that of a two-dimensional CNN model, making it more suitable for processing large volumes of time-series data. Furthermore, a Bayesian optimization algorithm is embedded during the training process of the improved CNN model to optimize and obtain a suitable hyperparameter combination for structural damage identification. This improves the prediction accuracy and robustness of the improved CNN model, enhances its generalization ability in structural damage diagnosis, and reduces the probability of misjudgment. Based on the improved CNN model, data fusion technology is added, enabling the full extraction of bridge structure damage information from vibration signal data. This addresses the problem that environmental interference and sensor instability can alter some vibration signals, thus affecting the bridge structure damage detection results. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of the structural damage recognition method based on data fusion and one-dimensional convolutional neural network in this invention;

[0046] Figure 2 This is a schematic diagram of the data augmentation process in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the feature extraction module in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the damage identification process of a bridge structure in an embodiment of the present invention;

[0049] Figure 5 This is a diagram showing the structural damage identification results for scenario 16 in an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] This embodiment provides a structural damage recognition method based on data fusion and a one-dimensional convolutional neural network, such as Figure 1 As shown, the specific steps include:

[0052] S1: Multiple damage units are set on the bridge structure. Each damage unit contains different damage locations and damage degrees. An external excitation is applied to the bridge structure. Vibration signals of the bridge structure under different damage states are acquired through several sensors. An original dataset is established. The original dataset is normalized to obtain a processed dataset.

[0053] Specifically, the sensor is an accelerometer or a cable force sensor. In this embodiment, several accelerometers collect and monitor the acceleration data of the bridge structure under different damage states. The acceleration data is then one-hot encoded according to the corresponding structural damage state, establishing a data set that establishes the correspondence between the acceleration data and different structural damage states. The advantages of one-hot encoding are its simplicity and effectiveness, and it can solve the problem that discrete features cannot be directly used in machine learning algorithms. It is also easy for network languages ​​to read. The label is designed as a vector consisting of w′ elements (w′ equals the unit number of all monitoring locations of the bridge structure), where each element value represents the degree of damage at the monitoring location. All the acceleration data and the label encoding constitute the original dataset.

[0054] Specifically, in this embodiment, the data in the original dataset are all multidimensional time series of length L, and the sampling frequency is set to f Hz. In order to accelerate the network training convergence speed and generalization ability, the original dataset is normalized, and the calculation formula is as follows:

[0055]

[0056] Where L0 is the original dataset, σ and σ(L0) are the mean and standard deviation of the original data, respectively. It is the processed dataset.

[0057] S2: Divide the processed dataset into a training set and a test set, and use the data slicing method to cut the data in the training set and the test set into multiple data segments at the same time. Then, perform data augmentation processing on the training set and the test set in the manner of sliding window.

[0058] S3: Establish an improved one-dimensional convolutional neural network model, train it based on the data-augmented training set, and test it based on the data-augmented test set to obtain the trained improved one-dimensional convolutional neural network model.

[0059] Specifically, the training set is input into the improved one-dimensional convolutional neural network model. Based on the training set, the model learns and understands the features of the data and predicts the target variable from these features. The output data vector is a predicted label encoding of the damage state. A loss function is used to estimate the difference between the predicted label encoding and the true label encoding, and the gradient value is used to update the neural network. A test set is used to evaluate the performance of the final model. In this embodiment, a validation set is also set up to evaluate the model's performance and adjust the parameters. Based on the validation set, the optimal model parameters are found to enable the improved one-dimensional convolutional neural network model to achieve the best performance.

[0060] The improved one-dimensional convolutional neural network model includes a feature extraction module and a feature mapping module. The feature extraction module is used to extract features representing damage in the data augmented dataset and output them to the feature mapping module. The feature mapping module is used to establish a mapping relationship between the features representing damage and the damage type.

[0061] S4: Using the improved one-dimensional convolutional neural network model after training, structural damage identification is performed on the bridge structure to obtain several decisions. The several decisions are then fused using data fusion technology to obtain a comprehensive result of structural damage detection.

[0062] In a specific embodiment, in S2, since the training process of a deep learning network requires a certain batch of independent data samples of the same size, rather than a complete time series, a sliding window method is needed for data augmentation. The sliding window is a common method for signal data decomposition. The process of performing data augmentation on the training and test sets using the sliding window method is as follows:

[0063] like Figure 2 As shown, several windows of size h×w are set, where h represents the window length (time dimension) and w represents the window width (sensor dimension). The window sliding time step is set to s. Adjacent windows overlap. Then, each complete long-period data point in the training and test sets is divided into N identical sub-data samples to achieve data augmentation. The calculation formula is:

[0064]

[0065] In the formula, L is the length of the multidimensional time series, and f is the sampling frequency.

[0066] Specifically, in order to improve computational efficiency and reduce data redundancy, this embodiment adopts an overlapping method to set the sliding window. That is, each time the window is slid, the edge of the window overlaps with the edge of the previous window. This can effectively reduce the amount of computation and achieve the effect of data augmentation by overlapping adjacent windows.

[0067] In a specific embodiment, in S3, such as Figure 3 As shown, the feature extraction module includes three structurally identical first feature extraction blocks Block_1, second feature extraction block Block_2, and third feature extraction block Block_3 connected in sequence. Each feature extraction block includes a one-dimensional convolutional layer 1D-Conv, a batch normalization layer, and a pooling layer Maxpooling1D. The one-dimensional convolutional layer is used to extract features representing impairments in the input data and transmit them to the batch normalization layer. The batch normalization layer is used to perform batch normalization on the input data and transmit it to the pooling layer. The pooling layer is used to reduce the dimensionality of the input data.

[0068] The one-dimensional convolutional layer comprises K neurons, each performing filtering processing on the input data, as shown in the formula:

[0069] h k =f'(con1D(w k ,X)+b k )

[0070]

[0071] Among them, h k This is the result after the k-th convolution; bk Indicates the offset coefficient; x i ,x i+1 ,…,x i+h Let represent the input data from step i, step i+1 to step i+h; h is the time dimension; f' is the non-linear activation function of the feature extraction module; X represents the input data of the one-dimensional convolutional layer, with an input dimension of (k′, w′, 1), where k′ is the number of data points contained in a single sensor dimension, w′ is the number of sensors deployed, and 1 is the number of channels; w k is the kernel weight vector of a one-dimensional convolutional layer; conv1D(·) is a one-dimensional convolution operator; This represents a vector multiplication operation;

[0072] The output data of the one-dimensional convolutional layer is normalized by a Batch Normalization (BN) layer. The calculation formula for the BN layer is as follows:

[0073]

[0074] x i =γh i +β≡BN γ,β (h i )

[0075] in, h represents the normalized feature data. i Represents the original feature data; μ B σ represents the average value of small batch B. B The standard deviation of the small batch B is represented by γ, the scaling factor by β, the offset factor by BN, and the normalization layer by x. i This represents the adjusted and normalized data; ε is a single value to prevent the denominator from becoming zero.

[0076] Specifically, during training, a one-dimensional convolutional neural network model learns the features of the input data layer by layer. As the feature extraction layers deepen, learning instability may occur when performing backpropagation, such as gradient vanishing or gradient exploding. Batch normalization is a method to prevent internal covariance shift, which is to normalize by performing normalization in mini-batch units. The batch normalization algorithm calculates the mean and variance of each batch of training data, then shifts and scales the original data to zero mean and unit variance, and then assigns the shifted data to weights and biases to enhance its effect. Reparameterization using BN layers helps alleviate the problem of coordinated updates through layers in a neural network.

[0077] Specifically, after a one-dimensional convolutional layer, multiple feature maps are generated. Often, many feature maps contain some useless feature information. Therefore, a pooling layer is set up. The pooling operation works by collecting the maximum value of a specific region or averaging the values ​​obtained from a specific region. It can emphasize the role of specific features without changing the characteristics of the feature map, thus increasing the robustness of the network and compressing the dimensionality of the feature map output.

[0078] Specifically, the feature extraction module adopts a one-dimensional convolutional neural network model, which has compact computational characteristics, low computational requirements, and is suitable for real-time data monitoring and identification using low-cost hardware.

[0079] In a specific embodiment, the feature mapping module includes a first fully connected layer FC layers 1, a ReLU activation function layer, a dropout layer, a second fully connected layer FC layers 2, and a Softmax activation function layer connected in sequence. The fully connected layers form a classifier for structural damage identification and establish a mapping relationship between the features extracted by the feature extraction module and the structural damage state.

[0080] The general calculation formulas for the first fully connected layer and the second fully connected layer are:

[0081] h j+1 =f"(∑h j ×w j+1 +b j+1 )

[0082] Among them, h j For the input data of the fully connected layer; w j+1 and b j+1 These represent the weights and additional bias terms, respectively; f” is the non-linear activation function of the feature mapping module; h j+1 This represents the output data of the fully connected layer, and j represents the j-th step.

[0083] Specifically, such as Figure 3 As shown, a Flatten layer is provided between the feature extraction module and the feature mapping module. The Flatten layer is used to convert the multidimensional data output by the feature extraction module into one-dimensional data and transmit it to the feature mapping module.

[0084] Specifically, the second fully connected layer is the output layer. Damage recognition belongs to multi-label classification. The activation function of the second fully connected layer is the softmax function. The dimension of the output vector is equal to the number of detection locations, and the value of each element in the output vector is related to the degree of damage. Dropout is one of the techniques used to solve the overfitting problem of neural networks. It generates more effective training examples by reducing co-adaptation between neurons. Overfitting can be determined as long as the training loss is much smaller than the test loss.

[0085] In a specific embodiment, in S3, the improved one-dimensional convolutional neural network model determines the structural hyperparameters and training hyperparameters using a Bayesian optimization algorithm. For the problem of identifying different structural damages, datasets with different structures belong to multi-source domains and exhibit significant differences. No deep neural network structure can adequately solve all structural damage identification problems. Even network structures and parameters that perform exceptionally well for a specific problem face the primary and most important challenge when applying deep convolutional neural networks to specific structural damage identification problems: reconstructing the deep convolutional neural network structure and configuring optimized parameters, or adjusting an existing excellent network structure. Both processes require extensive experience and specialized knowledge, undoubtedly making them extremely difficult. Previous research has shown that the Bayesian optimization algorithm is more suitable for deep learning hyperparameter optimization than network search, random search, and other optimization algorithms. Bayesian optimization can efficiently find the optimal hyperparameters with fewer iterations. Therefore, in this embodiment, the structural hyperparameters and training hyperparameters are determined using a Bayesian optimization algorithm, including:

[0086] 1) Based on the improved one-dimensional convolutional neural network model, a set x containing several hyperparameter elements is obtained, mainly including structural hyperparameters and training hyperparameters. Structural hyperparameters include the number of convolutional layers, the number of filters in each layer, kernel size, stride, number of pooling layers, size, stride, padding, number of fully connected layers, number of neurons, normalization technique, Dropout layer parameters, and activation function. Training hyperparameters include optimizer, learning rate, batch size, and ratio of training set to test set.

[0087] 2) The search space of all parameters in set x is determined through experiments. The improved one-dimensional convolutional neural network model is initialized to obtain the initial parameter set D. It is then assumed that the initial parameter set D satisfies the established surrogate model M. Based on the prior distribution and the initial parameter set D, the posterior distribution p(y|x,D) is calculated, where x is the parameter, y is the variable, and p is the posterior distribution probability. The acquisition function S is constructed using the posterior distribution. The next set of hyperparameters x1 is selected through the acquisition function S. The hyperparameter set x1 is then substituted into the improved one-dimensional convolutional neural network model for training to obtain the network model evaluation value y1. After multiple iterations, when the loss function converges, the optimal hyperparameter combination of the improved one-dimensional convolutional neural network model is obtained.

[0088] In this embodiment, in order to determine the search space of each hyperparameter element, research from different fields was referenced and multiple experiments were conducted to specify the search range of each hyperparameter. The specific values ​​are shown in Table 1.

[0089] Table 1:

[0090]

[0091] In a specific embodiment, S4, the process of obtaining the comprehensive result of structural damage detection by performing decision fusion processing on several decisions through data fusion technology includes:

[0092] When identifying the location of damage to a bridge structure, a voting method is used to calculate the comprehensive result of the structural damage detection. The calculation formula is as follows:

[0093]

[0094] Where H(x) represents the comprehensive structural damage detection result obtained using the voting method; T represents the number of classifier models, and the decision output of each classifier model is {h1, h2, ..., h...} T}; l is the number of category labels, and each category label is {c1, c2, ..., c l};C argmax The function to return the index of the maximum value. For the 1st, 2nd...lth output decision;

[0095] When identifying the location and extent of damage to a bridge structure, the mean value method is used to calculate the comprehensive result of the structural damage detection. The calculation formula is as follows:

[0096]

[0097] Where H(x)' represents the comprehensive structural damage detection result obtained using the mean method, T' represents the number of regression prediction models, and the decision output of each regression prediction model is {h1′,h2′,…,h T Let n' be the number of position change degree labels, and let the position change degree labels be {d1, d2, ..., dn'}. n}; This is the 1st, 2nd, ..., nth output decision.

[0098] This embodiment proposes a data fusion technology decision, which refers to using a certain damage identification method to process multiple sets of sensor data within the same short time period to perform damage identification, such as... Figure 4 As shown, the data collected by each sensor is processed to obtain data 1, 2, 3, ..., n. The prediction results are obtained through this single model to obtain decision 1, 2, 3, ..., n, namely Damage_1, Damage_2, Damage_3, ..., Damage_n. Then, the prediction results are fused to obtain the final comprehensive decision.

[0099] In specific embodiments, the voting method includes absolute majority voting, relative majority voting, weighted voting, and soft voting, and any one of these methods can be used.

[0100] To demonstrate the effectiveness of this embodiment, a simply supported beam vibration test was conducted. A 1.5m long stainless steel simply supported beam was selected as the research object. The beam has a rectangular cross-section, a width of 0.1m, and a height of 0.02m. Rolling bearings were used to fix and constrain the beam at both ends to simulate simply supported supports. The simply supported beam was divided into 30 equal micro-beam elements. The material properties of the beam are: Young's modulus of 206 GPa and density of 7900 kg / m³. 3 To simulate the random vibration of structures in a real-world environment, a vibrator was used as the external excitation source, and a ZS2020 signal generator was used as the signal generator. The vibrator was controlled to vibrate in a random waveform, thereby exciting the simply supported beam to vibrate. In this way, the vibration response of the simply supported beam under actual environmental loads can be approximately simulated, and the corresponding acceleration data can be recorded for subsequent analysis and research.

[0101] In this embodiment, in order to collect structural vibration acceleration data of a simply supported beam structure under different damage states, accelerometers are installed at the measuring points of the simply supported steel beam. A JM5937 dynamic signal acquisition instrument is used to analyze the signals returned by the accelerometers and transmit them to a computer terminal.

[0102] To verify the performance of the proposed structural damage identification method, acceleration data samples of a simply supported beam structure in both normal and damaged states are required. This embodiment employs the method of adding mass blocks to the beam structure to increase the local structural stiffness and mass of the simply supported beam, simulating negative structural damage. This allows verification of the sensitivity of the proposed method to changes in the stiffness and mass of actual structural micro-units. One or more units are randomly selected on the beam structure for mass adjustment. This embodiment uses homogeneous steel plates as the added mass blocks, each plate measuring 100mm × 50mm × 5mm and weighing approximately 196g. The degree of damage depends on the number of plates connected to the beam. Three levels of damage are set by adding 1, 2, and 3 layers of plates, with these added masses representing approximately 1.45% (Level 1), 2.90% (Level 2), and 4.35% (Level 3) of the total beam mass, respectively. The experiment considered both single-location mass change and multi-location mass change scenarios. Five units were randomly selected from different regions of the beam structure for adjustment. The first type of experiment only considered single-location mass change, as shown in Table 2, and designed 15 local mass change scenarios (from scenario 1 to scenario 15). In addition, scenario 0 was a scenario with no structural change. The second type of experiment considered scenarios where different degrees of additional mass were applied to two locations simultaneously, as shown in the table, and designed 10 different combination scenarios.

[0103] Table 2

[0104]

[0105] Table 3

[0106]

[0107] Using the method proposed in this embodiment, a structural pattern recognition test was conducted using a one-dimensional convolutional neural network model with optimal parameter combinations. The damage recognition capability of the improved one-dimensional convolutional neural network model was further analyzed from two aspects: damage location and damage degree. First, its recognition effect on local single-change scenarios of the structure was verified. Data samples of 16 single-change scenarios of simply supported beams were used, and a total of 1280 validation samples were obtained using the sliding window technique. The improved one-dimensional convolutional neural network model and data fusion technique could identify the structural state mapped to each validation sample with 100% accuracy. Next, for 10 composite change scenarios from scenarios 16 to 25, a regression prediction method was used. For each scenario, 80 validation sample data were used, and the recognition results were statistically analyzed using box plots. The horizontal axis represents the unit point number, and the vertical axis represents the degree of change. For locations with structural damage, outliers will be displayed in the plot, indicating that the data at that unit number location has a large degree of change. Figure 5 The image shows the identification results for scenario 16, where damage conditions were set at points 4 and 10. As can be seen from the image, the box plot shows changes at cell numbers 4 and 10. This indicates that the proposed method can accurately identify and locate the damage location for each scenario, and the regression prediction of the degree of change also exhibits a relatively high level of accuracy. Overall, the method proposed in this embodiment significantly improves the accuracy of structural damage identification.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A structural damage identification method based on data fusion and one-dimensional convolutional neural networks, characterized in that, The specific steps include: S1: Vibration signals of the bridge structure under different damage states are acquired through several sensors, and an original dataset is established. The original dataset is then normalized to obtain a processed dataset. S2: Divide the processed dataset into a training set and a test set, and perform data augmentation processing on the training set and the test set using a sliding window method; S3: Establish an improved one-dimensional convolutional neural network model, train it based on the data-augmented training set, and test it based on the data-augmented test set to obtain the trained improved one-dimensional convolutional neural network model; the improved one-dimensional convolutional neural network model includes a feature extraction module and a feature mapping module. The feature extraction module is used to extract features representing damage in the data-augmented dataset and output them to the feature mapping module. The feature mapping module is used to establish a mapping relationship between the features representing damage and the damage state. The feature extraction module includes three structurally identical feature extraction blocks connected in sequence; each feature extraction block includes a one-dimensional convolutional layer, a batch normalization layer, and a pooling layer. The one-dimensional convolutional layer is used to extract features representing damage in the input data and transmit them to the batch normalization layer. The batch normalization layer is used to perform batch normalization on the input data and transmit it to the pooling layer. The pooling layer is used to reduce the dimensionality of the input data. The one-dimensional convolutional layer comprises K neurons, each performing filtering processing on the input data, as shown in the formula: in, For the first k The result of step convolution; Indicates the offset coefficient; Indicates the first i Step, the first i+ Step 1 to the i+h Step input data; h The time dimension; This is the non-linear activation function for the feature extraction module; This represents the input data for a one-dimensional convolutional layer, with input dimensions of ( , w ,1), This refers to the number of data points contained in the input data within a single sensor dimension. w The number of sensors deployed, where 1 represents the number of channels; This is the kernel weight vector of a one-dimensional convolutional layer; conv1D (·) is a one-dimensional convolution operator; This represents a vector multiplication operation; The output data of the one-dimensional convolutional layer is normalized by a Batch Normalization (BN) layer. The calculation formula for the BN layer is as follows: in, This represents the normalized feature data. Represents the original feature data; µ B Indicates small batch B The average value, Indicates small batch B The standard deviation, γ Indicates the scaling factor; β Indicates the offset factor; BN Indicates the normalization layer; This represents the adjusted and normalized data; ε is a constant. The feature mapping module includes a first fully connected layer, a ReLU activation function layer, a dropout layer, a second fully connected layer, and a softmax activation function layer connected in sequence. The general calculation formulas for the first fully connected layer and the second fully connected layer are: in, This is the input data for the fully connected layer; and These represent the weights and additional bias terms, respectively. is the nonlinear activation function for the feature mapping module; This represents the output data of the fully connected layer. j Indicates the first j step; The improved one-dimensional convolutional neural network model determines its structural hyperparameters and training hyperparameters using a Bayesian optimization algorithm, including: 1) Based on the structure of the improved one-dimensional convolutional neural network model, a set x containing several hyperparameter elements is obtained, including the number of convolutional layers, the number of filters in each layer, kernel size, stride, number of pooling layers, size, stride, padding, number of fully connected layers, number of neurons, normalization technique, Dropout layer parameters, activation function, optimizer, learning rate, batch size, and ratio of training set to test set; 2) The search space of all parameters in set x is determined through experiments. The improved one-dimensional convolutional neural network model is initialized to obtain the initial parameter set D. It is then assumed that the initial parameter set D satisfies the established surrogate model M. Based on the prior distribution and the initial parameter set D, the posterior distribution is calculated. The acquisition function S is constructed using the posterior distribution. The next set of hyperparameters x1 is selected through the acquisition function S. The hyperparameter set x1 is substituted into the improved one-dimensional convolutional neural network model for training to obtain the network model evaluation value y1. After multiple iterations, when the loss function converges, the optimal hyperparameter combination of the improved one-dimensional convolutional neural network model is obtained. S4: Using the improved one-dimensional convolutional neural network model after training, structural damage identification is performed on the bridge structure to obtain several decisions. The several decisions are then fused using data fusion technology to obtain a comprehensive result of structural damage detection. In S4, the process of obtaining the comprehensive result of structural damage detection by performing decision fusion processing on several decisions through data fusion technology includes: When identifying the location of damage to a bridge structure, a voting method is used to calculate the comprehensive result of the structural damage detection. The calculation formula is as follows: in, H(x) The results of structural damage detection obtained using the voting method are as follows; T Let be the number of classifier models, and let the decision output by each classifier model be { h 1 ,h 2 ,…,h T }; l The number of category labels, where each category label is { c 1 , c 2 ,…,c l }; The function to return the index of the maximum value. … For the 1st, 2nd… l One output decision; When identifying the location and extent of damage to a bridge structure, the mean value method is used to calculate the comprehensive result of the structural damage detection. The calculation formula is as follows: in, The overall results of structural damage detection obtained using the mean value method are as follows: Let be the number of regression prediction models, and let the decision output by each regression prediction model be { h 1 ,h 2 ,…,h T }, n The number of labels indicating the degree of positional change, where the labels are { d 1 ,d 2 ,…,d n }; … For the 1st, 2nd...nth Each output decision.

2. The structural damage identification method based on data fusion and a one-dimensional convolutional neural network according to claim 1, characterized in that, In S2, the process of performing data augmentation on the training and test sets using a sliding window is as follows: Set several sizes as h × w The window, h The window length represents the time dimension. w Let the window width be the sensor dimension, and the window sliding time step be s. The sliding window is set by overlapping adjacent windows. Then, each complete long-period data point in the training and test sets is divided into... N For each identical sub-data sample, the calculation formula is: In the formula, L The length of the multidimensional time series. f The sampling frequency.

3. The structural damage identification method based on data fusion and a one-dimensional convolutional neural network according to claim 2, characterized in that, In S1, the calculation formula for normalizing the original dataset is as follows: in, It is the original dataset. and These are the mean and standard deviation of the original data, respectively. It is the processed dataset.

4. The structural damage identification method based on data fusion and a one-dimensional convolutional neural network according to claim 1, characterized in that, The voting method can be any one of the following: absolute majority voting, relative majority voting, weighted voting, and soft voting.

Images