Structural damage identification method based on deep reconstruction network and multi-dimensional feature fusion

Abstract

The application belongs to the technical field of bridge health monitoring, and particularly relates to a structure damage identification method based on a deep reconstruction network and multi-dimensional feature fusion, which comprises the following steps: obtaining an acceleration signal of a target bridge structure; inputting the acceleration signal into a Res-UNet-AE autoencoder to obtain a reconstructed signal, and calculating a reconstruction error and a signal-to-noise ratio according to the reconstructed signal; inputting the acceleration signal and the reconstructed signal into a perception autoencoder to obtain a perception index; fusing the reconstruction error, the signal-to-noise ratio and the perception index to obtain a three-dimensional damage feature space; and performing unsupervised clustering and identification according to the three-dimensional damage feature space to obtain a damage category identification result of the target bridge. The application can realize accurate differentiation of different working conditions of a structure without relying on any damage label.

Description

Technical Field

[0001] This invention relates to the field of bridge health monitoring technology, and in particular to a structural damage identification method based on deep reconstruction networks and multidimensional feature fusion. Background Technology

[0002] With the accelerating pace of modern urbanization, bridge structures, as a crucial component of infrastructure construction, play a vital role in modern transportation systems. However, during long-term service, bridge structures can suffer damage due to factors such as loads, environmental corrosion, fatigue effects, and unforeseen events (e.g., earthquakes, fires), severely impacting their safety and reliability. Therefore, conducting research on structural health monitoring (SHM) for bridge structures is of paramount importance for ensuring the safe operation of engineering structures, extending their service life, and reducing maintenance costs. Although a large amount of health monitoring data on bridge structures has been accumulated, it remains difficult to accurately identify structural damage simply by visually observing these raw response signals.

[0003] With the rapid development of artificial intelligence technology, machine learning algorithms have been widely applied in the field of structural damage detection. Currently, commonly used algorithms mainly include deep learning models such as Principal Component Analysis (PCA), Artificial Neural Networks (ANN), and Convolutional Neural Networks (CNN). Compared with traditional signal processing methods, these algorithms show significant advantages in recognition accuracy and can effectively capture the complex nonlinear mapping relationship between structural response and damage state. However, these methods essentially rely on supervised learning, requiring a large number of labeled training samples with damage information to achieve effective modeling. In real-world engineering environments, structures are in a healthy state most of the time, and real damage events are extremely rare and difficult to simulate artificially, making obtaining sufficient damage samples highly challenging both technically and economically.

[0004] As an effective solution to the aforementioned problems, unsupervised learning methods exhibit unique advantages due to their independence from damage labels. However, existing unsupervised learning methods still face numerous challenges in practical applications. First, most current methods use single metrics such as Mean Squared Error (MSE) or Mean Absolute Error (MAE) to evaluate reconstruction quality. These metrics have limited representational power and struggle to accurately identify response changes caused by high-noise backgrounds or subtle damage, thus affecting the accuracy of damage detection. Second, some models, due to shallow network structures or overly simple feature encoding mechanisms, limit their ability to extract deep features from complex temporal signals, resulting in poor robustness and generalization ability in complex engineering environments. More critically, most existing algorithms remain within the scope of anomaly detection, only able to determine whether anomalies exist in the structure, but struggling to further identify the type and extent of damage, thus limiting their practicality in fine-grained structural evaluation and maintenance strategy formulation. Summary of the Invention

[0005] The purpose of this invention is to provide a structural damage identification method based on deep reconstruction network and multi-dimensional feature fusion, which aims to accurately distinguish different working conditions of structures by constructing a feature space with richer information and greater sensitivity to damage without relying on any damage labels.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] Structural damage identification methods based on deep reconstruction networks and multidimensional feature fusion include:

[0008] Acquire the acceleration signal of the target bridge structure;

[0009] The acceleration signal is input into the Res-UNet-AE autoencoder to obtain the reconstructed signal. The reconstruction error and signal-to-noise ratio are calculated based on the reconstructed signal. The Res-UNet-AE autoencoder integrates the UNet model, the ResNet model and the convolutional autoencoder CAE, and is obtained by training with healthy acceleration signals.

[0010] The acceleration signal and the reconstructed signal are input into the perceptual autoencoder to obtain perceptual indicators, wherein the perceptual autoencoder is trained using an acceleration dataset that includes healthy and damaged states.

[0011] The reconstruction error, signal-to-noise ratio, and perception indicators are fused to obtain a three-dimensional damage feature space;

[0012] Unsupervised clustering and identification are performed based on the three-dimensional damage feature space to obtain the damage category identification results of the target bridge.

[0013] Optionally, the Res-UNet-AE autoencoder adopts a symmetrical encoder-decoder structure;

[0014] The encoder consists of an initial feature extraction layer and three coding layers. Each layer contains two residual blocks and extracts abstract features step by step through downsampling. The bottleneck layer processes the highest-level features and integrates global information. The decoder contains three decoding layers, which restore the temporal resolution through upsampling and fuse with the skip connection features of the corresponding layer of the encoder to finally output the reconstructed signal.

[0015] Optionally, training the Res-UNet-AE autoencoder using the health acceleration signal includes:

[0016] The healthy acceleration signal is input into the initial Res-UNet-AE autoencoder, and the reconstructed signal corresponding to the healthy acceleration signal is output. The reconstruction loss is calculated based on the reconstructed signal.

[0017] The health acceleration signal and the corresponding reconstruction signal are input into the sensing autoencoder to calculate the sensing loss.

[0018] The Res-UNet-AE autoencoder is trained by obtaining a joint loss function based on the reconstruction loss and the perception loss.

[0019] Optionally, the joint loss function is:

[0020] ;

[0021] ;

[0022] ;

[0023] in, For the joint loss function, To reconstruct the loss, Let λ be the perceptual loss, and λ be the weighting coefficient that balances the reconstruction loss and the perceptual loss. (·) represents the first _th ... Layer feature extraction function, For the first The number of elements in the feature map corresponding to the layer. The original signal, To reconstruct the signal, It is an L1 norm.

[0024] Optionally, calculating the reconstruction error and signal-to-noise ratio based on the reconstructed signal includes:

[0025] ;

[0026] ;

[0027] in, For reconstruction error, Where S is the signal-to-noise ratio, and S is the total number of signal sample points. The original signal corresponding to the s-th signal sample point. Let be the reconstructed signal corresponding to the s-th signal sample point.

[0028] Optionally, the perception metrics to be acquired include:

[0029] ;

[0030] in, For perception indicators, For the first The weight coefficients of the feature map corresponding to the layer. (·) represents the first _th ... Layer feature extraction function.

[0031] Optionally, unsupervised clustering and identification based on the three-dimensional damage feature space includes:

[0032] Based on the Gaussian mixture model, samples in the three-dimensional damage feature space are assigned to the Gaussian component with the highest posterior probability, thereby achieving automatic clustering of different loss states and obtaining cluster labels.

[0033] The clustering labels and real labels are optimally matched using the Hungarian algorithm to obtain the damage category identification results.

[0034] Optionally, the optimal matching of the cluster labels and the real labels using the Hungarian algorithm includes: constructing a confusion matrix between the real labels and the cluster labels, transforming the problem of the optimal correspondence between the cluster labels and the real labels into a maximum weight matching problem, solving the optimal label mapping using the Hungarian algorithm, and obtaining the damage category identification result.

[0035] The beneficial effects of this invention are as follows: This invention proposes a structural damage identification method based on deep reconstruction networks and multi-dimensional feature fusion. It aims to accurately distinguish different structural conditions without relying on any damage labels by constructing a feature space that is richer in information and more sensitive to damage. First, it overcomes the limitations of traditional reconstruction errors by measuring the distance between the original signal and the reconstructed signal in the deep feature space, thus uncovering key damage information ignored by traditional methods. Second, it constructs a three-dimensional feature space composed of Mean Absolute Error (MAE), Signal-to-Noise Ratio (SNR), and Perceptual Loss (PL), and adopts a deep Res-UNet-AE as the core reconstruction model. This model significantly enhances the representation and reconstruction capabilities of complex time-series signals through residual learning and multi-scale skip connection structures within the model. Finally, it combines a pre-trained perceptual feature encoder to provide high-level semantic supervision and applies a Gaussian Mixture Model (GMM) in the three-dimensional feature space for automatic condition identification, forming a complete end-to-end unsupervised damage classification solution. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of a structural damage identification method based on deep reconstruction network and multidimensional feature fusion according to an embodiment of the present invention;

[0038] Figure 2 This is a diagram of the Res-UNet-AE architecture according to an embodiment of the present invention;

[0039] Figure 3 This is a diagram of the perceptual autoencoder architecture according to an embodiment of the present invention;

[0040] Figure 4 This is a layout diagram of the target bridge acceleration sensor according to an embodiment of the present invention;

[0041] Figure 5 This is a comparison chart of signal reconstruction results under the healthy state (INT, all components are intact) of an embodiment of the present invention;

[0042] Figure 6The above are comparison results of signal reconstruction under loss state one and damage state two in an embodiment of the present invention. (a) is a comparison of signal reconstruction under damage state one (DMG1, the cross section of the mid-span vertical member is reduced by 50%), and (b) is a comparison of signal reconstruction under damage state two (DMG2, the mid-span vertical member is completely cut off).

[0043] Figure 7 The above are comparison results of signal reconstruction in the repair state and the third damage state of the present invention. (a) is a comparison of signal reconstruction in the repair state (RCV, welding repair of the cut vertical truss members), and (b) is a comparison of signal reconstruction in the third damage state (DMG3, complete cutting of the 5 / 8 span vertical members).

[0044] Figure 8 Box plots of three damage indicators under five working conditions in this embodiment of the invention are shown, wherein (a) is a box plot of reconstruction error under five working conditions, (b) is a box plot of signal-to-noise ratio under five working conditions, and (c) is a box plot of sensing indicator under five working conditions.

[0045] Figure 9 The following are two-dimensional and three-dimensional index distribution diagrams of embodiments of the present invention, wherein (a) is a two-dimensional index distribution diagram of signal-to-noise ratio and reconstruction error, (b) is a two-dimensional index distribution diagram of sensing index and reconstruction error, (c) is a two-dimensional index distribution diagram of sensing index and signal-to-noise ratio, and (d) is a three-dimensional index distribution diagram of sensing index, signal-to-noise ratio and reconstruction error.

[0046] Figure 10 This is a three-dimensional feature space clustering result diagram based on the GMM clustering results and labeled with the mean values, according to an embodiment of the present invention.

[0047] Figure 11 This is a schematic diagram of the confusion matrix of GMM clustering in an embodiment of the present invention;

[0048] Figure 12 This is a distribution diagram of the Z24 bridge sensors according to an embodiment of the present invention;

[0049] Figure 13 This is a clustering result diagram of the Z24 bridge three-dimensional feature space based on the GMM clustering results, according to an embodiment of the present invention.

[0050] Figure 14 This is a confusion matrix diagram of Z24 bridge GMM clustering in an embodiment of the present invention. Detailed Implementation

[0051] 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, and 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.

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] This embodiment proposes a structural damage identification method based on deep reconstruction networks and multi-dimensional feature fusion, including:

[0054] Acquire the acceleration signal of the target bridge structure;

[0055] The acceleration signal is input into the Res-UNet-AE autoencoder to obtain the reconstructed signal. The reconstruction error and signal-to-noise ratio are calculated based on the reconstructed signal. The Res-UNet-AE autoencoder integrates the UNet model, the ResNet model and the convolutional autoencoder CAE, and is trained using healthy acceleration signals.

[0056] The acceleration signal and the reconstructed signal are input into the perceptual autoencoder to obtain the perceptual index. The perceptual autoencoder is trained using an acceleration dataset that includes both healthy and damaged states.

[0057] A three-dimensional damage feature space is obtained by fusing reconstruction error, signal-to-noise ratio and perception indicators.

[0058] Unsupervised clustering and identification are performed based on the three-dimensional damage feature space to obtain the damage category identification results of the target bridge.

[0059] Specifically, such as Figure 1 As shown, firstly, a Res-UNet-AE autoencoder is constructed as the core reconstruction model. This model integrates the multi-scale feature processing capabilities of UNet, the deep learning advantages of ResNet, and the unsupervised learning mechanism of the convolutional autoencoder. Secondly, a perceptual autoencoder is designed to provide deep semantic supervision, and a pre-trained perceptual feature encoder provides high-level constraints for training the main model. Thirdly, a three-dimensional damage feature space is constructed based on the trained model, including three damage-sensitive features: reconstruction error, signal-to-noise ratio, and perceptual indicators. Finally, a Gaussian mixture model is used to perform unsupervised clustering on the three-dimensional feature space to achieve automatic identification and classification of structural damage states.

[0060] Furthermore, the Res-UNet-AE autoencoder adopts a symmetrical encoder-decoder structure;

[0061] The encoder consists of an initial feature extraction layer and three coding layers. Each layer contains two residual blocks and extracts abstract features step by step through downsampling. The bottleneck layer processes the highest-level features and integrates global information. The decoder contains three decoding layers, which restore the temporal resolution through upsampling and fuse with the skip connection features of the corresponding layer of the encoder to finally output the reconstructed signal.

[0062] Specifically, this embodiment constructs a Res-UNet-AE autoencoder, organically integrating the multi-scale feature processing framework of UNet, the residual learning mechanism of ResNet, and the unsupervised reconstruction capability of convolutional autoencoder CAE. The Res-UNet-AE architecture is as follows: Figure 2 As shown in Table 1, the configuration information of Res-U-Net-AE is as follows. ResNet: Uses "residual blocks" as the core building block, introducing a residual learning mechanism. UNet: Employs a "skip connection" architecture to achieve multi-scale feature fusion. Convolutional Autoencoder (AE): Follows its overall unsupervised reconstruction paradigm of "encoder-bottleneck layer-decoder".

[0063] Table 1

[0064]

[0065] The ResNet-UNet fusion architecture proposed in this embodiment has 73.3M parameters, which is increased compared to traditional methods. However, through reasonable architecture design and optimized training strategies, a good balance between parameter efficiency and performance is achieved. The increased number of parameters compared to the standard UNet is mainly used to enhance feature extraction and perceptual loss calculation, which results in a significant improvement in anomaly detection performance. Meanwhile, strategies such as staged training and adaptive learning rates effectively control the training time to around 20 minutes, making the model highly feasible for practical engineering applications.

[0066] Furthermore, training the Res-UNet-AE autoencoder using healthy acceleration signals includes:

[0067] The healthy acceleration signal is input into the initial Res-UNet-AE autoencoder, and the reconstructed signal corresponding to the healthy acceleration signal is output. The reconstruction loss is calculated based on the reconstructed signal.

[0068] The health acceleration signal and the corresponding reconstructed signal are input into the sensing autoencoder to calculate the sensing loss.

[0069] Based on the reconstruction loss and perception loss, a joint loss function is obtained to train the Res-UNet-AE autoencoder.

[0070] Specifically, the perceptual autoencoder is an independent model used to provide deep semantic supervision. Its configuration information and architecture are shown in Table 2. Figure 3The perceptual autoencoder shown consists of a perceptual feature encoder and a decoder. The encoder extracts multi-scale features through residual blocks and downsampling, and compresses the feature map into a fixed-length vector through global average pooling and fully connected layers. The decoder expands the dimension using fully connected layers, and gradually recovers the temporal resolution and number of channels of the signal through upsampling and convolution. In the damage recognition process, the first stage involves independent pre-training on a large amount of vibration data, including healthy states and various damage states, to learn robust and discriminative feature representations. The second stage retains and fixes only the encoder part, providing high-level semantic supervision during Res-UNet-AE training. Specifically, in each forward propagation of Res-UNet-AE, the original healthy acceleration signal and its corresponding reconstructed signal are simultaneously input into the fixed perceptual encoder. The perceptual loss is obtained by calculating the output difference between the two in the encoder's multi-layer feature space. This perceptual loss, together with the reconstruction loss of Res-UNet-AE itself, constitutes a joint loss function, and this gradient is used for backpropagation to update only the network weights of Res-UNet-AE.

[0071] This supervision not only includes point-by-point reconstruction accuracy but also provides semantic constraints in the feature space. This is achieved by calculating the difference between the original signal and the Res-UNet-AE reconstructed signal in the intermediate layers (perceptual loss). The model can optimize reconstruction while maintaining semantic and structural consistency, making the deviation of the damage signal more obvious in the feature space, thereby improving the sensitivity and robustness to subtle damage.

[0072]

[0073] Furthermore, the joint loss function is:

[0074] ;

[0075] The Res-UNet-AE autoencoder is trained using only healthy vibrational signals, and the optimization objective is to minimize the reconstruction loss.

[0076] ;

[0077] To improve sensitivity to subtle damage, deep semantic supervision is introduced into the perceptual feature encoder. The perceptual loss is defined as:

[0078] ;

[0079] in, For the joint loss function, To reconstruct the loss, The perceptual loss is denoted by λ, which is a weighting coefficient balancing the reconstruction loss and the perceptual loss. Its specific value is determined by minimizing the reconstruction error on an independent validation set containing only healthy state samples using a grid search algorithm. (·) represents the first _th ... Layer feature extraction function, For the first The number of elements in the feature map corresponding to the layer. The original signal, To reconstruct the signal, It is the L1 norm, which represents the sum of the absolute values ​​of the elements in a vector or matrix.

[0080] This design allows the model to maintain consistency with healthy signals in both pixel-level accuracy and high-level semantics, significantly enhancing its sensitivity to and robustness to abnormal signals.

[0081] Furthermore, this embodiment comprehensively assesses whether the structure is damaged and determines the specific type of damage by using three damage-sensitive features.

[0082] The first metric is Reconstruction Error. This metric directly measures the difference between the original signal and the reconstructed signal at each point. This embodiment uses Mean Absolute Error (MAE) as the reconstruction error, which is insensitive to outliers and robustly reflects the overall reconstruction fidelity. When the structure is healthy, the model can reconstruct the signal well, thus obtaining... Low. When the structure is abnormal, the model cannot accurately reproduce its pattern. It will increase significantly. The formula for calculating the reconstruction error is as follows:

[0083] ;

[0084] in, The reconstruction error is represented by S, where S is the total number of signal sample points. The original signal corresponding to the s-th signal sample point. Let be the reconstructed signal corresponding to the s-th signal sample point.

[0085] The second metric is the signal-to-noise ratio (SNR). SNRSignal-to-noise ratio (SNR) assesses reconstruction quality from the perspective of relative energy. It is the ratio of the energy of the original signal (Signal Power) to the energy of the reconstruction error (Noise Power). For normal signals, the energy should be much greater than the error energy, resulting in a high SNR; when the structure is abnormal, the proportion of reconstruction error energy increases, leading to a decrease in SNR. The formula for calculating SNR (in decibels, dB) is as follows:

[0086] ;

[0087] in, This refers to the signal-to-noise ratio.

[0088] The third indicator is the perceptual loss indicator. This metric utilizes a pre-trained perceptual feature encoder to measure the difference between the original and reconstructed signals in the deep feature space. This differs from metrics that only focus on point-by-point differences. different, This metric aims to capture higher-level structural and semantic biases. Even if the MAE of two signals is small, their feature representations at different levels of the perceptual encoder may differ significantly. This metric can identify anomalous patterns that the model cannot successfully mimic at the perceptual level. Its calculation method is as follows:

[0089] ;

[0090] in, For perception indicators, For the first The weight coefficients of the feature map corresponding to the layer are taken as the values ​​of the first layer. The reciprocal of the number of elements in the feature map corresponding to the layer. (·) represents the first _th ... Layer feature extraction function, The original signal, For reconstructing the signal.

[0091] Through the above process, each signal sample in the test set is transformed into a unique three-dimensional damage feature vector. , , The fusion of these three indicators creates an information-rich feature space.

[0092] In constructing the 3D damage feature space, the intrinsic relationships between various features were analyzed in depth. A strong negative correlation exists between the reconstruction error (MAE) and the signal-to-noise ratio (SNR), while the perceptual index (PL) provides unique and complementary damage information from a deep semantic level. This complementarity among features allows the 3D feature space to more comprehensively represent the structural state from multiple dimensions, including numerical fidelity, energy distribution, and semantic features. To ensure stable training of the model on the correlated feature space, a hierarchical weight initialization and phased training strategy was adopted, effectively avoiding gradient conflicts and demonstrating stable convergence during training. Furthermore, by combining regularization techniques such as Dropout and weight decay with an early stopping mechanism, the performance difference between the model on the training and validation sets was controlled within 5%, demonstrating its excellent generalization ability and robustness.

[0093] Furthermore, unsupervised clustering and identification based on the three-dimensional damage feature space includes:

[0094] Based on the Gaussian mixture model, samples in the three-dimensional damage feature space are assigned to the Gaussian component with the highest posterior probability, thereby achieving automatic clustering of different loss states and obtaining cluster labels.

[0095] The damage category identification results are obtained by using the Hungarian algorithm to optimally match the cluster labels with the real labels.

[0096] Furthermore, the optimal matching of cluster labels and real labels using the Hungarian algorithm includes: constructing a confusion matrix between real labels and cluster labels, transforming the problem of finding the optimal correspondence between cluster labels and real labels into a maximum weight matching problem, solving for the optimal label mapping using the Hungarian algorithm, and obtaining the damage category identification results.

[0097] Specifically, to achieve automatic identification of structural damage states, an unsupervised clustering method based on the Gaussian Mixture Model (GMM) is employed based on this three-dimensional damage feature space. GMM is a commonly used probabilistic generative model that can effectively model and cluster complex data distributions in this feature space. Its core idea is to treat the feature vector of each sample as a mixture of multiple Gaussian distributions, fit the parameters of each Gaussian component by maximizing likelihood estimation, and then softly assign samples accordingly. Its probability density function... It can be represented as:

[0098] ;

[0099] in, These are the indices of the Gaussian components. It is the total number of Gaussian components (i.e., the preset number of clusters, corresponding to the total number of damage conditions to be identified). It is the mixing coefficient of the k-th Gaussian component. and Then they are respectively the first The mean vector and covariance matrix of each Gaussian component. It is a sample point in the three-dimensional damage feature space (i.e., a feature vector composed of reconstruction error, signal-to-noise ratio and perception index). It is the probability density function of a multivariate Gaussian distribution.

[0100] The model parameters are iteratively optimized using the Expectation-Maximization (EM) algorithm to maximize the log-likelihood function of the observed data. .

[0101] ;

[0102] Where N is the total number of sample points in the three-dimensional damage feature space. The first in the three-dimensional damage feature space The feature vector of each sample point.

[0103] In step E, the posterior probability of each sample belonging to each Gaussian component is calculated. (Level of responsibility):

[0104] ;

[0105] in, , , They represent the first The mixing coefficients, mean vector, and covariance matrix of each Gaussian component. This indicates that the nth sample point belongs to the nth sample point. The posterior probability of each Gaussian component It is the probability density function of a multivariate Gaussian distribution.

[0106] In step M, update the mixing coefficients, mean, and covariance matrix:

[0107] ;

[0108] ;

[0109] ;

[0110] Here, the superscript T is the matrix transpose operator.

[0111] Finally, each sample is assigned to the Gaussian component with the highest posterior probability, achieving automatic clustering for different damage states. Since the labels obtained from unsupervised clustering are independent of the true label order, the Hungarian Algorithm is used to find the optimal label correspondence on the confusion matrix to maximize clustering accuracy.

[0112] First, construct the confusion matrix between the real labels and the cluster labels:

[0113] ;

[0114] in, Represents the true category Clustering categories The number of overlapping samples is the confusion matrix. For the sample The true damage category label function, For the sample Clustering label function obtained after GMM clustering.

[0115] Then, the problem of finding the optimal correspondence between cluster labels and true labels is transformed into a maximum weight matching problem: each true class and each cluster class are treated as two vertices of a bipartite graph to confuse the matrix elements. As a connection to the real category Clustering categories The weights of the edges are determined. The optimal label mapping can be obtained by solving for the maximum weight perfect matching on this bipartite graph using the Hungarian algorithm. This mapping is for each real category. Assign a unique cluster category This maximizes the total weight of the matching edges. Based on this optimal mapping, the clustering accuracy is calculated as follows:

[0116] ;

[0117] in, Indicates optimal label mapping Below, the real category Correctly classified into cluster categories The number of samples, This is the final calculated clustering accuracy evaluation metric. This represents the total number of sample points in the three-dimensional damage feature space.

[0118] This method can effectively evaluate the correspondence between clustering results and actual damage states. By modeling the multidimensional feature space using the GMM model, the differences in feature distribution among different damage states can be fully explored, enabling automatic differentiation of structural health status. Combining the Hungarian algorithm for optimal matching between cluster labels and actual labels not only solves the problem of uncertain label order in unsupervised clustering but also quantifies the consistency between clustering results and actual damage states.

[0119] The unsupervised structural damage identification method proposed in this embodiment will be validated based on field measurement data of the target bridge to evaluate the applicability and performance of the method in a real engineering environment.

[0120] The target bridge is a single-span simply supported steel truss bridge. This embodiment designs five progressive damage conditions to simulate the typical damage evolution process that may occur during the service of the truss bridge. The settings for each condition are shown in Table 3.

[0121]

[0122] Eight single-axis accelerometers are strategically arranged along the vertical direction of the bridge deck, such as... Figure 4 As shown in the diagram, five sensors were positioned on one side of the damaged truss member, and three sensors were installed on the opposite side to ensure comprehensive capture of local and global response changes caused by the damage. Environmental vibration tests and vehicle-induced vibration tests were conducted under each working condition. The vehicle test involved a standard loaded vehicle passing over the bridge at a constant speed of 40 km / h. All sensors simultaneously acquired 35 seconds of data at a sampling rate of 200 Hz, resulting in a total of 7,000 continuous time-series data points.

[0123] Representative data files from multiple repeated experiments were selected for analysis, containing synchronously acquired data from eight sensor channels. Continuous time-series signals were processed using non-overlapping segmentation, with each sample having a fixed length of 100 data points. At a sampling frequency of 200 Hz, the corresponding vibration response time window was 0.5 seconds. This time window was set primarily considering two factors: firstly, it should cover the complete decay process of the structural dynamic response in the time domain, fully capturing transient and steady-state information; secondly, it should include 5-15 complete cycles of the dominant frequency vibration in the frequency domain, providing a frequency resolution of 2 Hz, which helps in identifying low-frequency modes and long-period response characteristics.

[0124] Through multi-sensor synchronous acquisition and a reasonable signal segmentation strategy, 70 samples were finally constructed for each damage condition, and the sample division is shown in Table 4. Each sample has a dimension of (100, 8), where 100 represents the time step and 8 represents the number of sensor channels.

[0125]

[0126] This embodiment employs an unsupervised anomaly detection training strategy. The Res-UNet-AE autoencoder is trained using only 44 training samples under the healthy state (INT), reflecting the fundamental idea of ​​model learning using only normal state data. By learning the inherent vibration characteristics of healthy structures, the model generates significant reconstruction errors when inputting data from damaged conditions, thereby achieving automatic identification of the damage state. The validation and test sets contain samples from all conditions and are used for hyperparameter tuning and final performance evaluation, respectively. The test set is further used to calculate three-dimensional damage characterization indices, including reconstruction error (MAE), signal-to-noise ratio (SNR), and perception index (PL), and unsupervised clustering analysis is performed based on these indices.

[0127] To visually demonstrate the effectiveness of unsupervised anomaly detection, this embodiment selects a typical sample from the first sensor channel and compares and analyzes the original signal and the reconstructed signal under five different operating conditions. Figure 5 , Figure 6 (a)- Figure 6 (b) Figure 7 (a)- Figure 7 (b) Reconstruction results of the Res-UNet-AE autoencoder under healthy and various damaged conditions. The continuous signal shown has been smoothed using Savitzky-Golay filtering. Figure 5 It can be seen that the original signal and the reconstructed signal overlap well in the healthy state (INT), and the reconstruction effect is good. This is because the training dataset only contains data from the healthy state, allowing the Res-UNet-AE autoencoder to remember this pattern during training and encode and decode the data accordingly. However, when impairment occurs, the dynamic characteristics of the structure may change, leading to differences in Res-UNet-AE encoding and decoding. After smoothing, unnecessary high-frequency fluctuations in the reconstructed signal are effectively suppressed, but the essential feature differences caused by impairment are preserved, resulting in significant differences between the original data and the reconstructed data. Crucially, all impairment conditions produce reconstruction errors greater than those in the healthy state, clearly revealing their anomaly nature and proving that reconstruction error is a sensitive indicator for distinguishing between healthy and impairment states.

[0128] Utilizing the relationship between health and damage states in reconstruction error, this embodiment further achieves damage detection by tracking the change of reconstruction error over time. When the reconstruction error exceeds a threshold, the system triggers an anomaly warning. Table 5 shows the damage detection accuracy under different reconstruction error thresholds. As shown in Table 5, the recognition accuracy reaches its optimal level of 100% when the reconstruction error threshold is set between 0.07 and 0.2. This paper uses the Weibull extreme value distribution theory to automatically estimate the threshold for unsupervised damage detection. As shown in Table 5, the reconstruction error threshold determined by this method is 0.094, which is within the threshold range. Therefore, the unsupervised damage detection method can achieve 100% detection accuracy.

[0129]

[0130] The proposed Res-UNet-AE autoencoder is used to obtain reconstruction error, and 100% damage detection is achieved by tracking changes in the reconstruction error. However, it cannot distinguish between different damage conditions. Building upon this, Res-UNet-AE is used to obtain reconstruction error and signal-to-noise ratio. Simultaneously, a perceptual model is used to obtain perceptual indicators. By combining three damage-sensitive features and using GMM for unsupervised clustering of these three features under different damage conditions, effective classification of various damage conditions is ultimately achieved.

[0131] First, statistical analysis and visualization are performed on three impairment-sensitive features: reconstruction error, signal-to-noise ratio, and perception index. Figure 8 (a)- Figure 8 (c) shows the error distribution of three damage-sensitive features under five operating conditions. Compared to other damage conditions, the healthy state has the lowest median value and the smallest error in the reconstruction error and perception index. This is because both Res-UNet-AE and the perception model effectively learned the vibration mode of the structure in the healthy state during training, resulting in smaller reconstruction errors and perception indices. Simultaneously, the healthy condition has the highest median signal-to-noise ratio (SNR), which represents the ratio of the energy of the original signal to the energy of the reconstruction error. For normal signals, the energy should be much greater than the error energy, resulting in a high SNR. Figure 8 Further investigation revealed that as the degree of damage increased, the median values ​​of reconstruction error and perception index showed a significant upward trend, while the signal-to-noise ratio showed a downward trend, reflecting a monotonically increasing or decreasing relationship between the three damage sensitivity features and the degree of damage.

[0132] Figure 9 (a)- Figure 9 (d) illustrates the correlations among three impairment-sensitive features: reconstruction error, signal-to-noise ratio, and perception metrics. From Figure 9 (a)- Figure 9(d) It can be seen that both pairwise indicators and combinations of all three indicators can effectively distinguish between healthy and damaged states. However, it is worth noting that... Figure 9 (a)- Figure 9 (d) cannot distinguish between partially damaged operating conditions. For example, in Figure 9 (a)- Figure 9 In (c), the data points corresponding to the two operating conditions, DMG1 and RCV, are relatively close to each other, even within... Figure 9 (b) shows an overlap. Therefore, DMG1 and RCV represent two different scenarios: a 50% reduction in the cross-section of the mid-span vertical truss member and welding repair of the cut member, respectively. Since RCV, although repairing the cut member, cannot completely restore it to its pre-cut state, it can also be considered an anomaly. For these two scenarios, DMG1 and RCV, representing different degrees of damage at the same location, the data points corresponding to the two damage scenarios are close together in the two-dimensional scatter plot, making them difficult to distinguish. Figure 9 In (d), most damage conditions can be distinguished by combining the three damage identification features, but DMG1, RCV and DMG3 are close to each other and are still difficult to distinguish.

[0133] Based on the above analysis, a three-dimensional feature space distribution was constructed using three damage sensitivity indicators, and the Growing Mixture Model (GMM) was used to perform unsupervised clustering on the test set data to distinguish different damage conditions. Table 6 shows the results of various evaluation indicators after GMM clustering. In the unsupervised clustering evaluation, the silhouette score reached 0.6584, and a value greater than 0.5 is generally considered to indicate good clustering results. The Calinski-Harabasz index (CH index) was 674.9998, and a higher value indicates better clustering results; its high value further verifies the characteristics of high separation between clusters and high cohesion within clusters. The Davies-Bouldin index (DB index) was 0.5516, and a lower value indicates better clustering results; its relatively low value confirms that the clustering results have excellent separation and compactness. In the supervised evaluation, GMM achieved a perfect score of 1.000 on all key indicators, achieving 100% accurate classification of the test samples. This result means that the clustering partitioning of GMM is completely consistent with the true labels of the samples, which fully verifies the effectiveness of the constructed three-dimensional feature space and the excellent performance of the unsupervised clustering method in structural damage identification.

[0134]

[0135] Figure 10 This displays the spatial clustering results for three sensitive features after GMM processing. Figure 12It can be seen that GMM successfully divided the data points corresponding to different working conditions into five close and separate clusters, and the center of each cluster is highly consistent with the actual data distribution of different damage working conditions. Figure 11 The confusion matrix further demonstrates the perfect classification effect, with all samples correctly classified and all five working conditions accurately identified.

[0136] To further verify the universality and effectiveness of the unsupervised damage identification method proposed in this paper, the effectiveness of the method in this embodiment is verified using the publicly available Z24 bridge benchmark dataset.

[0137] Bridge Z24 was a post-tensioned prestressed concrete box girder bridge, which underwent a month-long progressive damage test before its demolition in 1998. This example analyzes three representative conditions with clear damage gradients: intact condition (UD), pier settlement of 40 mm (D1), and pier settlement of 95 mm (D2). Experimental data were collected using 34 accelerometers evenly distributed throughout the bridge structure, with a sampling frequency of 100 Hz. The sensor arrangement is as follows... Figure 12 As shown. For the data from each sensor, 100 samples were extracted from the intact state as a training set to build an unsupervised learning model; at the same time, 50 samples were extracted from the intact state and the two damaged states respectively, forming a test set of a total of 150 samples to comprehensively evaluate the damage detection and trend capture capabilities of the method.

[0138] In the first stage of damage identification, the reconstruction error is obtained through the proposed Res-UNet autoencoder. When the reconstruction error exceeds a threshold, the system displays an anomaly. Table 7 shows the damage detection accuracy of the algorithm under different thresholds. This embodiment uses Weibull extreme value distribution theory to automatically estimate the threshold for unsupervised damage detection. As shown in Table 7, the reconstruction error threshold determined by this method is 0.007, which is within the prior threshold range. Therefore, the unsupervised damage detection method can achieve 100% detection accuracy.

[0139]

[0140] In the second stage of damage identification, this embodiment obtains three damage-sensitive features through Res-UNet-AE and a perceptual model: reconstruction error, signal-to-noise ratio, and perceptual index. Furthermore, these three indices are unsupervised clustered using GMM to distinguish different damage conditions. Figure 13 The results of GMM clustering are presented, showing that GMM successfully achieved high-precision classification of three working conditions, and the three working conditions formed relatively independent clusters in the feature space; the cluster boundaries between different working conditions are clear and the separation is good, demonstrating excellent inter-class separability. Figure 14The confusion matrix is ​​shown for GMM clustering of bridge Z24. The confusion matrix results show that the method achieved a classification accuracy of 98%: the intact state (UD) and the pier settlement of 95 mm (D2) were both 100% accurately identified, the pier settlement of 40 mm (D1) had a recognition rate of 94%, and only 3 samples were misclassified as pier settlement of 95 mm, showing excellent overall performance.

[0141] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A structural damage identification method based on deep reconstruction networks and multidimensional feature fusion, characterized in that, include: Acquire the acceleration signal of the target bridge structure; The acceleration signal is input into the Res-UNet-AE autoencoder to obtain the reconstructed signal. The reconstruction error and signal-to-noise ratio are calculated based on the reconstructed signal. The Res-UNet-AE autoencoder integrates the UNet model, the ResNet model and the convolutional autoencoder CAE, and is obtained by training with healthy acceleration signals. The Res-UNet-AE autoencoder adopts a symmetrical encoder-decoder structure; The encoder consists of an initial feature extraction layer and three coding layers. Each layer contains two residual blocks and extracts abstract features step by step through downsampling. The bottleneck layer processes the highest-level features and integrates global information. The decoder contains three decoding layers, restores the temporal resolution through upsampling, and fuses the jump connection features of the corresponding layer of the encoder to finally output the reconstructed signal. Training the Res-UNet-AE autoencoder using the health acceleration signal includes: The healthy acceleration signal is input into the initial Res-UNet-AE autoencoder, and the reconstructed signal corresponding to the healthy acceleration signal is output. The reconstruction loss is calculated based on the reconstructed signal. The health acceleration signal and the corresponding reconstruction signal are input into the sensing autoencoder to calculate the sensing loss. Based on the reconstruction loss and the perception loss, a joint loss function is obtained to train the Res-UNet-AE autoencoder; The acceleration signal and the reconstructed signal are input into the perceptual autoencoder to obtain perceptual indicators, wherein the perceptual autoencoder is trained using an acceleration dataset that includes healthy and damaged states. The reconstruction error, signal-to-noise ratio, and perception indicators are fused to obtain a three-dimensional damage feature space; Unsupervised clustering and identification are performed based on the three-dimensional damage feature space to obtain the damage category identification results of the target bridge, including: Based on the Gaussian mixture model, samples in the three-dimensional damage feature space are assigned to the Gaussian component with the highest posterior probability, thereby achieving automatic clustering of different loss states and obtaining cluster labels. The clustering labels and real labels are optimally matched using the Hungarian algorithm to obtain the damage category identification results.

2. The structural damage identification method based on deep reconstruction network and multidimensional feature fusion according to claim 1, characterized in that, The joint loss function is: ； ； ； in, For the joint loss function, To reconstruct the loss, Let λ be the perceptual loss, and λ be the weighting coefficient that balances the reconstruction loss and the perceptual loss. (·) represents the first _th ... Layer feature extraction function, For the first The number of elements in the feature map corresponding to the layer. The original signal, To reconstruct the signal, It is an L1 norm.

3. The structural damage identification method based on deep reconstruction network and multidimensional feature fusion according to claim 2, characterized in that, Calculating the reconstruction error and signal-to-noise ratio based on the reconstructed signal includes: ； ； in, For reconstruction error, Where S is the signal-to-noise ratio, and S is the total number of signal sample points. The original signal corresponding to the s-th signal sample point. Let be the reconstructed signal corresponding to the s-th signal sample point.

4. The structural damage identification method based on deep reconstruction network and multidimensional feature fusion according to claim 3, characterized in that, The perception metrics to be acquired include: ； in, For perception indicators, For the first The weight coefficients of the feature map corresponding to the layer. (·) represents the first _th ... Layer feature extraction function.

5. The structural damage identification method based on deep reconstruction network and multidimensional feature fusion according to claim 1, characterized in that, The optimal matching of the cluster labels and the real labels using the Hungarian algorithm includes: constructing a confusion matrix between the real labels and the cluster labels; transforming the problem of finding the optimal correspondence between the cluster labels and the real labels into a maximum weight matching problem; solving for the optimal label mapping using the Hungarian algorithm; and obtaining the damage category identification result.

Images