A method and system for identifying cable-stayed cable abnormalities based on signal decomposition and deep learning fusion

Abstract

The application relates to the technical field of stay cables and discloses a stay cable anomaly identification method and system based on signal decomposition and deep learning fusion, which comprises the following steps: acquiring stay cable vibration related original data, wherein the related original data comprises stay cable vibration acceleration data, stay cable force data and environmental influence data; performing signal decomposition preprocessing and feature extraction on the original data to obtain to-be-trained feature data; inputting the to-be-trained feature data into a deep learning model for training to establish a mapping relationship between stay cable vibration signals and health states; calculating a signal reconstruction error based on the trained deep learning model to obtain an anomaly judgment threshold; inputting a test signal into the trained deep learning model to obtain a test signal reconstruction error; and identifying the abnormal state of the stay cable based on the comparison result of the test signal reconstruction error and the anomaly judgment threshold to obtain a stay cable anomaly identification result.

Description

Technical Field

[0001] This invention relates to the field of cable-stayed bridge technology, specifically to a method and system for anomaly identification of cable-stayed bridges based on the fusion of signal decomposition and deep learning. Background Technology

[0002] As the core load-bearing component of cable-stayed bridges, the cable stress state is a crucial indicator of hidden internal damage such as wire breakage and anchor loosening. Achieving high-precision monitoring and early anomaly identification of cable stress in cable-stayed bridges has become a core requirement for ensuring the long-term safety of bridges.

[0003] Under complex environmental excitation, cable vibration signals often exhibit strong non-stationary characteristics. Traditional Fourier analysis methods are prone to fundamental frequency identification errors and frequency aliasing when processing such signals, making it difficult to meet practical requirements for cable force calculation accuracy. In addition, current anomaly identification methods rely heavily on inspection results and are not sensitive enough to subtle changes in dynamic characteristics caused by early weak anomalies within the cable, making them unsuitable for the operation and maintenance management needs of modern bridges. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a method and system for anomaly identification of cable-stayed bridges based on the fusion of signal decomposition and deep learning, in order to solve the aforementioned technical problems.

[0005] Firstly, a method for anomaly identification of cable-stayed bridges based on the fusion of signal decomposition and deep learning is provided, including: Obtain raw data related to the vibration of the cable-stayed bridge, including cable vibration acceleration data, cable force data, and environmental impact data. The original data is subjected to signal decomposition preprocessing and feature extraction to obtain the feature data to be trained; The feature data to be trained is input into a deep learning model for training to establish a mapping relationship between the vibration signal and health status of the cable-stayed bridge. The signal reconstruction error is calculated based on the trained deep learning model to obtain the anomaly judgment threshold. The test signal is input into the trained deep learning model to obtain the test signal reconstruction error; Based on the comparison between the test signal reconstruction error and the anomaly judgment threshold, the abnormal state of the cable is identified, and the cable anomaly identification result is obtained.

[0006] Furthermore, after acquiring the raw data related to the vibration of the stay cables, the process also includes three-dimensional processing of the raw data: The collected cable vibration acceleration data were reconstructed in three dimensions to obtain the three-dimensional vibration data of the cable. The three-dimensional vibration data of the cable-stayed bridge is segmented to obtain several three-dimensional sub-block data. The three-dimensional sub-block data is converted using the maximum intensity projection method to obtain two-dimensional image data; The two-dimensional image data is divided using a fast random partitioning algorithm to obtain multiple datasets. Feature filtering is performed on multiple datasets to obtain multiple training feature data.

[0007] Furthermore, the original data undergoes signal decomposition preprocessing and feature extraction to obtain the training feature data, including: The original data is adaptively decomposed using an adaptive noise complete set empirical mode decomposition algorithm to obtain several mode components and residual components. Analyze the energy proportion of each modal component to select the modal components that carry the core energy of the signal; The filtered modal components are reconstructed to obtain the feature-enhanced signal.

[0008] Furthermore, the original data is adaptively decomposed using an adaptive noise complete set empirical mode decomposition algorithm to obtain several modal components and residual components, including: Add Gaussian white noise with a preset standard deviation to the original data to obtain a noisy signal; Empirical mode decomposition is performed on the noisy signal to calculate the local average value of the signal, thereby obtaining the first-order mode component and the first-order residual value; Repeat the above steps to obtain the modal components and corresponding residual values ​​of each stage in sequence until the decomposition meets the preset termination condition, and obtain all modal components and residual components.

[0009] Furthermore, the deep learning model is a one-dimensional convolutional autoencoder, and constructing the one-dimensional convolutional autoencoder includes: An encoding layer is constructed, which is equipped with several convolutional filters to extract features from the input training feature data through convolution calculation; A pooling layer is set after the convolutional layer, and pooling operation is performed using a pooling kernel of a preset size; A decoding layer is constructed, which is equipped with several deconvolution filters. The signal is reconstructed from the feature vector output by the coding layer through upsampling and deconvolution calculations. A fully connected layer is set after the decoding layer, and a loss function based on the difference between the reconstructed signal and the original signal is set to obtain a complete one-dimensional convolutional autoencoder.

[0010] Furthermore, the feature data to be trained is input into the deep learning model for training, including: Multiple datasets obtained through a fast random partitioning algorithm are allocated into training and validation sets according to a preset ratio. The training set is input into the encoding layer of a one-dimensional convolutional autoencoder, and deep features of the signal are extracted through convolution and pooling operations to obtain feature vectors. The feature vector is input into the decoding layer, and the signal is reconstructed through upsampling and deconvolution operations. The error between the reconstructed signal and the original signal in the training set is calculated based on the loss function. The parameters and bias values ​​of the convolutional filter and deconvolutional filter in the model are updated through the backpropagation algorithm. The process is iterated repeatedly until the model loss value tends to stabilize. The validation set is input into the trained one-dimensional convolutional autoencoder to calculate the reconstruction error of the validation set signal; The model's generalization ability is evaluated by comparing the difference between the reconstruction error on the validation set and the reconstruction error on the training set. If the difference exceeds the preset range, adjust the parameters and bias values ​​of the convolutional filter and deconvolutional filter in the model, and iterate the training again. If the difference is within a preset range, the model training is complete.

[0011] Furthermore, based on the trained deep learning model, the signal reconstruction error is calculated to obtain the anomaly judgment threshold, including: The training set signal and validation set signal obtained after training are input into the trained deep learning model to obtain the corresponding training set reconstruction error and validation set reconstruction error. The overall mean and standard deviation of the reconstruction errors of the training set and the validation set are calculated. Based on the population mean and population standard deviation, and combined with the statistical principle of normal distribution, the anomaly judgment threshold is obtained.

[0012] Furthermore, the test signal is input into the trained deep learning model to obtain the test signal reconstruction error, including: The test signal is preprocessed; The preprocessed test signal is input into the trained deep learning model, and the reconstructed signal corresponding to the test signal is obtained by extracting features from the encoding layer and reconstructing the signal from the decoding layer. The difference between the test signal, the original signal, and the reconstructed signal is calculated to obtain the test signal reconstruction error.

[0013] Furthermore, based on the comparison between the test signal reconstruction error and the anomaly judgment threshold, the abnormal state of the stay cable is identified, and the anomaly identification result of the stay cable is obtained, including: If the reconstruction error of the test signal exceeds the anomaly judgment threshold, then the corresponding cable is determined to be abnormal. If the reconstruction error of the test signal does not exceed the anomaly judgment threshold, the corresponding cable is determined to be in a healthy state.

[0014] Secondly, a cable-stayed bridge anomaly identification system based on the fusion of signal decomposition and deep learning is provided, based on any one of the cable-stayed bridge anomaly identification methods based on the fusion of signal decomposition and deep learning described above, including: The acquisition module is configured to acquire raw data related to the vibration of the stay cables. The preprocessing module is configured to perform signal decomposition preprocessing and feature extraction on the original data to obtain the feature data to be trained. The training module is configured to input the feature data to be trained into a deep learning model for training, and to establish a mapping relationship between the vibration signal of the cable-stayed bridge and its health status. The calculation module is configured to calculate the signal reconstruction error based on the trained deep learning model and obtain the anomaly judgment threshold. The input module is configured to input the test signal into the trained deep learning model to obtain the test signal reconstruction error. The comparison module is configured to identify the abnormal state of the cable-stayed bridge based on the comparison result between the reconstruction error of the test signal and the anomaly judgment threshold, and obtain the anomaly identification result of the cable-stayed bridge.

[0015] The invention employing the above technical solution has the following advantages: This invention utilizes ICEEMDAN to adaptively decompose the original cable-stayed bridge vibration signal. By filtering key intrinsic mode function components containing core information, environmental noise is effectively removed, achieving signal feature enhancement and reconstruction. Furthermore, the enhanced signal is input into a 1D-CAE network to mine sensitive features related to internal anomalies in the cable, constructing a mapping relationship from vibration signal to structural health status. This invention aims to achieve intelligent identification of early-stage, hidden anomalies in cable-stayed bridges without damaging prior labels, providing technical support for the digital and intelligent operation and maintenance of cable-stayed bridges. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention, the accompanying drawings used in the specific embodiments will be briefly described below. In all the drawings, the elements or parts are not necessarily drawn to scale.

[0017] Figure 1 The flowchart of the present invention is a method for anomaly identification of cable-stayed bridges based on the fusion of signal decomposition and deep learning. Figure 1 ; Figure 2 This invention provides a method for anomaly identification of cable-stayed bridges based on the fusion of signal decomposition and deep learning. Figure 2 ; Figure 3This is a schematic diagram of the arrangement of vibration acceleration sensors for cable-stayed bridge cables in a method for identifying anomalies based on signal decomposition and deep learning fusion, as described in this invention. Figure 4 This invention presents a method for identifying anomalies in cable-stayed bridges based on the fusion of signal decomposition and deep learning, and waveforms of the first nine IMF components obtained by ICEEMDAN decomposition in the system. Figure 5 This invention provides a method for identifying anomalies in cable-stayed bridges based on the fusion of signal decomposition and deep learning, and compares the individual waveforms of each order of IMF components in the system. Figure 6 This invention presents a method for anomaly identification of cable-stayed bridges based on the fusion of signal decomposition and deep learning, and curves showing the variation of the loss function of the 1D-CAE model training set and validation set with the number of iterations in the system. Figure 7 This invention relates to a method for anomaly identification of cable-stayed bridges based on the fusion of signal decomposition and deep learning, and a histogram of the reconstruction error distribution of the training set in the system. Figure 8 This invention relates to a method for identifying anomalies in cable-stayed bridges based on the fusion of signal decomposition and deep learning, and a histogram of the reconstruction error distribution of the validation set in the system. Figure 9 This invention presents a method for anomaly identification of cable-stayed bridges based on the fusion of signal decomposition and deep learning, and a diagram showing the distribution of reconstruction error and anomaly identification results of the test set in the system. Figure 10 This invention relates to a method for identifying anomalies in cable-stayed bridges based on the fusion of signal decomposition and deep learning, and the results of anomaly identification under simulated working conditions in the system. Detailed Implementation

[0018] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0019] like Figures 1-10As shown, this invention discloses a cable-stayed bridge anomaly identification method based on signal decomposition and deep learning fusion. An example is based on the cable health monitoring of a double-tower, single-cable-stayed bridge. The main span of the bridge is (125+23+125) m, and eight cable vibration acceleration sensors are installed on the bridge with a sampling frequency of 50 Hz (data sources and monitoring system parameters are derived from actual engineering monitoring requirements). This method achieves unsupervised anomaly identification of cable-stayed bridges through a complete process of "raw data acquisition → data preprocessing (3D / signal decomposition) → deep learning model (1D-CAE) training → anomaly judgment threshold determination → test signal anomaly identification." The core of this method lies in using Adaptive Noise Complete Ensemble Empirical Mode Decomposition (ICEEMDAN) to remove environmental noise and enhance signal features, and then using a one-dimensional convolutional autoencoder (1D-CAE) to mine the mapping relationship between vibration signals and health status. This allows for the identification of early, hidden anomalies without prior damage labels, including: Obtain raw data related to the vibration of the cable-stayed bridge, including cable vibration acceleration data, cable force data, and environmental impact data. The original data is preprocessed by signal decomposition and feature extraction to obtain the feature data to be trained; The feature data to be trained is input into the deep learning model for training, and a mapping relationship between the vibration signal of the cable-stayed bridge and its health status is established. The signal reconstruction error is calculated based on the trained deep learning model to obtain the anomaly judgment threshold. The test signal is input into the trained deep learning model to obtain the test signal reconstruction error; Based on the comparison between the test signal reconstruction error and the anomaly judgment threshold, the abnormal state of the stay cable is identified, and the anomaly identification result of the stay cable is obtained.

[0020] Specifically, the bridge health monitoring system collects three types of raw data: ① cable vibration acceleration data (collected by vibration acceleration sensors, reflecting the vibration state of the cable); ② cable force data (collected by cable force monitoring equipment, directly characterizing the load-bearing state of the cable, and is the core indicator for damage assessment); ③ environmental impact data (including wind speed, temperature, and traffic load, used to subsequently eliminate signal fluctuations caused by non-damage factors).

[0021] This embodiment collected a total of 7.09 million data points (corresponding to acceleration data during continuous monitoring periods. To facilitate subsequent processing, 3,000 data points within each 60-second period were divided into a basic sample unit to ensure the temporal continuity and data integrity of the samples).

[0022] The principle behind this step is that damage to the stay cable (such as wire breakage or anchor loosening) directly leads to changes in vibration characteristics (acceleration) and cable force. Environmental factors can interfere with the signal. Therefore, collecting multi-dimensional data can avoid the limitations of single data and lay the foundation for subsequent extraction of damage characteristics.

[0023] In this embodiment, after acquiring the raw data related to the vibration of the stay cable, the method further includes performing three-dimensional processing on the raw data: The collected cable vibration acceleration data were reconstructed in three dimensions to obtain the three-dimensional vibration data of the cable. The three-dimensional vibration data of the cable-stayed bridge is segmented to obtain several three-dimensional sub-block data. The maximum intensity projection method is used to transform the three-dimensional sub-block data to obtain two-dimensional image data; The two-dimensional image data is divided into multiple datasets using a fast random partitioning algorithm. Feature filtering is performed on multiple datasets to obtain multiple training feature data.

[0024] Specifically, considering the temporal characteristics of the cable-stayed bridge vibration acceleration data, one-dimensional time-series data is transformed into three-dimensional structural data through three-dimensional processing, which facilitates the intuitive extraction of joint features in the spatial and temporal domains. The specific steps are as follows: 3D Reconstruction: The collected cable vibration acceleration data is reconstructed in three dimensions, with time series as the first dimension, acceleration amplitude as the second dimension, and sensor acquisition channels as the third dimension. This transforms the discrete time-series acceleration data into three-dimensional cable vibration data with spatial structure, making the dynamic changes of cable vibration easier to capture by feature extraction algorithms.

[0025] Three-dimensional sub-block segmentation: The three-dimensional vibration data of the cable stay is segmented according to a preset size to obtain several independent three-dimensional sub-block data. Each sub-block corresponds to the complete vibration information within a time segment, avoiding the inefficiency of subsequent processing due to excessive data volume.

[0026] Two-dimensional transformation and dataset partitioning: The maximum intensity projection method is used to project and transform the data of each three-dimensional sub-block along the sensor channel dimension to obtain two-dimensional image data. This method can retain the key features with the largest vibration amplitude in the three-dimensional sub-block and filter redundant low-amplitude noise signals. Then, the two-dimensional image data is divided into multiple datasets through a fast random partitioning algorithm. The partitioning process ensures that the distribution characteristics of each dataset are consistent.

[0027] Feature selection: Feature selection is performed on the multiple datasets after partitioning to remove redundant features that are irrelevant to the body's health status, resulting in multiple feature data to be trained and used as input for subsequent deep learning models.

[0028] The principle behind this step is that 3D processing combines the temporal and amplitude features of time-series data, maximum intensity projection highlights key vibration information, and rapid random partitioning ensures the representativeness of the dataset. The three work together to achieve data dimensionality reduction and feature focusing, providing high-quality input for model training.

[0029] In this embodiment, the original data undergoes signal decomposition preprocessing and feature extraction to obtain the feature data to be trained, including: The original data is adaptively decomposed using the adaptive noise complete set empirical mode decomposition algorithm to obtain several mode components and residual components. Analyze the energy proportion of each modal component and screen out the modal components that carry the core energy of the signal; The filtered modal components are reconstructed to obtain the feature-enhanced signal.

[0030] In this embodiment, an adaptive noise complete set empirical mode decomposition algorithm is used to adaptively decompose the original data, obtaining several mode components and residual components, including: Add Gaussian white noise with a preset standard deviation to the original data to obtain a noisy signal; Empirical mode decomposition is performed on the noisy signal to calculate the local average value of the signal, thereby obtaining the first-order mode component and the first-order residual value; Repeat the above steps to obtain the modal components and corresponding residual values ​​of each stage in sequence until the decomposition meets the preset termination condition, and obtain all modal components and residual components.

[0031] Specifically, to address the strong non-stationary characteristics of vibration signals under complex environments, the ICEEMDAN algorithm is used to adaptively decompose the original data (including acceleration and cable force data), remove environmental noise, and extract core damage features. The specific steps are as follows: Adaptive decomposition: Gaussian white noise with a preset standard deviation is added to the original data (the standard deviation is determined according to the noise level of the original data. In this embodiment, the standard deviation of the first decomposition step is matched with the signal amplitude to avoid excessively strong or weak noise) to obtain a noisy signal; empirical mode decomposition is performed on the noisy signal. By screening extreme points and fitting the envelope, the local average value of the signal is calculated, and then the first-order mode component and the first-order residual value are separated. The first-order mode classification mainly carries the high-frequency environmental noise and the subtle vibration characteristics of the cable. The first-order residual value is the remaining part of the signal after removing the first-order mode component.

[0032] Iterative acquisition of all components: Repeatedly add Gaussian white noise with the corresponding standard deviation to the forward residual value → EMD decomposition → calculate local average value → obtain the current order modal component and new residual value: When the number of decompositions is K, add Gaussian white noise with the corresponding standard deviation to the K-1 order residual value. After EMD decomposition, the K-order modal classification and the K-order residual value are obtained; until the residual value meets the preset condition (in this embodiment, the termination condition is "the residual value has no obvious fluctuation and can no longer be decomposed into effective modal components"), and finally all modal components and 1 residual component are obtained.

[0033] Core mode selection and signal reconstruction: Analyze the energy proportion of each module component and select the modal components that carry the core signal capability. In this embodiment, the first 4 modal components carry a total of 99.44% of the signal energy and contain the core characteristics of cable vibration. The first 4 modal components after selection are reconstructed to obtain the signal with enhanced features. This signal serves as the core component of the feature data to be trained.

[0034] The Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) is an evolution of CEEMDAN. Unlike CEEMDAN, the ICEEMDAN algorithm selects the 3rd digit of the EMD decomposition after adding white noise. The decomposition of the modal components effectively solves the problems of residual noise and pseudo-modes in CEEMDAN. The specific decomposition process is as follows: definition The signal to be decomposed It is the first one obtained by applying the EMD algorithm. One modal component, It is the local average value of the original signal generated by EMD.

[0035] Step 1: Add a set of Gaussian white noise to the original signal Receive new signals for: (1) In the formula: The standard deviation of the Gaussian white noise added during the first signal decomposition is given.

[0036] Step 2: When At that time, the first residual value is obtained. and the first modal component : (2) (3) Step 3: When At that time, the calculation was performed after adding Gaussian white noise through EMD. The local mean is used to obtain the second residual value. and the second modal component : (4) (5) Step 4: Similarly, when At that time, the first individual residuals and the Modal components : (6) (7) In the formula: For the first The standard deviation of the Gaussian white noise added during the sub-decomposition of the signal.

[0037] Step 5: Repeat step 4 until the decomposition is complete, and obtain all modal components and residual components.

[0038] 3.1D-CAE Algorithm The Convolutional Autoencoder (1D-ACE) fuses the convolutional filtering of CNNs with an unsupervised autoencoder. Based on the fundamental idea of ​​AE, the denoising CAE is divided into an encoding layer and a decoding layer. Convolution and pooling operations are introduced in the encoding layer to achieve feature invariance extraction. The encoder performs encoding operations according to equation (1), expressed as follows: (8) In the formula: hk represents k feature maps; ∂ represents the activation function; ∗ represents the convolution calculation; P represents the input; Wn represents the nth convolution filter of the encoder; and b represents the encoder bias value.

[0039] A pooling layer is introduced after the convolutional layer to achieve high-level feature extraction. This paper uses MaxPooling in the pooling layer, with a kernel size of 2 × 2 and a stride of 2, located after the second convolutional layer of the encoder. Through repeated convolution and pooling, a complete encoding process is constructed.

[0040] After encoding, the noise in the original data is reduced, resulting in feature vectors that better reflect its characteristics. In the decoding stage, upsampling and deconvolution operations are used, and decoding is performed according to equation (2), expressed as follows: (9) In the formula: z is the reconstructed signal; W nis the nth convolutional filter of the decoder; b is the bias of the decoder; the feature map h is used as the input of the coding layer.

[0041] Finally, after processing in the fully connected layer, a CAE network for deep feature extraction is implemented. The loss function of this network serves as an adjustment index for the reconstructed signal, and its expression is: (10) Similar to CNN, CAE uses the backpropagation algorithm to obtain the gradient of each parameter, and then updates the parameters until the CAE model finally converges.

[0042] The principle of this step is as follows: The ICEEMDAN algorithm solves the modal aliasing and residual noise problems of traditional EMD and CEEMDAN by "adding adaptive white noise", so that the decomposed modal components have clear physical meaning; core modes are screened and reconstructed to focus on the vibration characteristics of the cable itself, eliminating environmental noise interference, and providing a clean signal basis for the subsequent model to extract damage-sensitive features.

[0043] In this embodiment, the deep learning model is a one-dimensional convolutional autoencoder. Constructing a one-dimensional convolutional autoencoder includes: An encoding layer is constructed, and several convolutional filters are set in the encoding layer. Features are extracted from the input feature data to be trained through convolution calculation. A pooling layer is set after the convolutional layer, and pooling operations are performed using a pooling kernel of a preset size; A decoding layer is constructed, which is equipped with several deconvolution filters. The signal is reconstructed from the feature vector output by the coding layer through upsampling and deconvolution calculations. A fully connected layer is set after the decoding layer, and a loss function based on the difference between the reconstructed signal and the original signal is set to obtain a complete one-dimensional convolutional autoencoder.

[0044] Specifically, this embodiment uses 1D-CAE as the deep learning model. Its core is to integrate the feature extraction capability of convolutional neural networks with the unsupervised reconstruction capability of autoencoders to construct a mapping relationship of "vibration signal → health state". The specific construction steps are as follows: Building the encoding layer: The core function of the encoding layer is to extract deep damage-sensitive features from the feature data to be trained. Several convolutional filters are set (in this embodiment, the number of filters is determined according to the feature dimension to ensure coverage of vibration features in different frequency bands). Features are extracted from the feature data to be trained through convolution calculation. After the convolutional layer, a pooling layer is set with a 2×2 max pooling kernel and a stride of 2 to perform pooling operation on the convolutional feature map. Max pooling can retain the key features with the largest amplitude in the feature map, while reducing the data dimension and avoiding model overfitting.

[0045] Building the decoding layer: The core function of the decoding layer is to reconstruct the original signal from the feature vector output by the encoding layer. Several deconvolution filters are set, and the signal is reconstructed from the feature vector output by the encoding layer through upsampling and deconvolution calculations to obtain the reconstructed signal.

[0046] Building a fully connected layer and setting a loss function: A fully connected layer is set after the decoding layer to map the reconstructed signal output by the decoding layer into a vector with the same dimension as the original data; at the same time, a loss function is set based on the difference between the reconstructed signal and the original signal. This loss function is used to adjust the model parameters so that the reconstructed signal is as close as possible to the original signal, ensuring that the model can learn the signal characteristics in a healthy state.

[0047] The principle of this step is as follows: the encoding layer extracts deep abstract features of the signal through convolution and pooling, the decoding layer verifies the effectiveness of feature extraction through upsampling and deconvolution, and the loss function provides the optimal direction for updating model parameters. Together, these three elements constitute an unsupervised model that can learn the laws of healthy signals.

[0048] In this embodiment, the feature data to be trained is input into the deep learning model for training, including: Multiple datasets obtained through a fast random partitioning algorithm are allocated into training and validation sets according to a preset ratio. The training set is input into the encoding layer of a one-dimensional convolutional autoencoder, and deep features of the signal are extracted through convolution and pooling operations to obtain feature vectors. The feature vectors are input into the decoding layer, and the signal is reconstructed through upsampling and deconvolution operations. The error between the reconstructed signal and the original signal in the training set is calculated based on the loss function. The parameters and bias values ​​of the convolutional filter and deconvolutional filter in the model are updated through the backpropagation algorithm. The process is iterated repeatedly until the model loss value tends to stabilize. The validation set is input into the trained one-dimensional convolutional autoencoder to calculate the reconstruction error of the validation set signal. By comparing the reconstruction error on the validation set with the reconstruction error on the training set, the generalization ability of the model can be evaluated. If the difference exceeds the preset range, adjust the parameters and bias values ​​of the convolutional filter and deconvolutional filter in the model, and iterate the training again. If the difference is within the preset range, the model training is complete.

[0049] Specifically, the core of model training is to enable the 1D-CAE to learn the inherent laws of the vibration signal of the cable-stayed bridge under healthy conditions through iterative optimization, and to establish a stable mapping relationship between vibration characteristics and healthy conditions. The specific steps are as follows: Dataset partitioning: The multiple feature data to be trained obtained in the above steps are allocated into training set and validation set according to a preset ratio (training set: validation set = 8:2 in this embodiment). The training set is used for the model to learn feature patterns, and the validation set is used to evaluate the model's generalization ability.

[0050] Iterative training on the training set: The training set is input into the encoding layer of the 1D-CAE, and local features of the signal are extracted through convolution operations. Then, the dimensionality is reduced by pooling operations to obtain deep feature vectors. The feature vectors are input into the decoding layer, and the signals corresponding to the training set are reconstructed through upsampling and deconvolution operations. The error between the reconstructed signal and the original signal of the training set is calculated based on the loss function. The parameters and bias values ​​of the convolutional filter and deconvolutional filter in the model are updated through the backpropagation algorithm. The above process is iterated repeatedly until the model loss value tends to stabilize (in this embodiment, after 1000 iterations, the training set loss value stabilizes at around 0.468), and the initial model training is completed.

[0051] Validation set optimization model: Input the validation set into the initially trained 1D-CAE model and calculate the reconstruction error of the validation set signal; compare the difference between the validation set reconstruction error and the training set reconstruction error (in this embodiment, the difference is only 0.007, and the generalization error is extremely small) to evaluate the model's generalization ability. If the difference exceeds the preset range, adjust the model parameters and iterate the training again; if the difference is within the preset range, confirm that the model training is complete, and at this time the model has learned the signal feature patterns in a healthy state.

[0052] The principle behind this step is that iterative training allows the model to learn the characteristic patterns of health signals, while optimization of the validation set avoids overfitting and ensures that the model can be stably transferred to the actual monitored test signals, providing a reliable model foundation for subsequent anomaly identification.

[0053] In this embodiment, the signal reconstruction error is calculated based on the trained deep learning model to obtain the anomaly judgment threshold, including: The training set signal and validation set signal obtained after training are input into the trained deep learning model to obtain the corresponding training set reconstruction error and validation set reconstruction error. The population mean and population standard deviation of the reconstruction error of the training set and the reconstruction error of the validation set are statistically analyzed. Based on the population mean and population standard deviation, and combined with the statistical principle of normal distribution, the anomaly judgment threshold is obtained.

[0054] Specifically, the anomaly detection threshold is the core criterion for distinguishing between healthy and abnormal signals. It is determined by calculating the signal reconstruction error distribution under healthy conditions based on the trained model. The specific steps are as follows: Calculate the health state reconstruction error: Input the training set signal and validation set signal obtained after training into the trained 1D-CAE model respectively, and obtain the corresponding training set reconstruction error and validation set reconstruction error through the encoding-decoding process. This error reflects the reconstruction accuracy of the model in the health state and is the basis for threshold setting.

[0055] Statistical error distribution parameters: the overall mean and overall standard deviation of the reconstruction error of the statistical training set and the reconstruction error of the validation set. In this embodiment, the overall mean is 0.00219 and the overall standard deviation is 0.00083. Both reflect the central tendency and dispersion of the reconstruction error under healthy conditions.

[0056] Determine the anomaly judgment threshold: Based on the population mean and population standard deviation, and combined with the internal statistical principle of normal distribution, the threshold is determined. In this embodiment, the normal distribution 3σ principle is adopted, and the anomaly judgment threshold is calculated to be μ+3σ, which ensures that the threshold has a high statistical guarantee rate and reduces the probability of healthy signals being misjudged as abnormal.

[0057] The principle behind this step is that the signal reconstruction error in a healthy state follows a normal distribution. Setting a threshold based on this distribution can maximize the coverage of the error range of healthy signals, while enabling abnormal signals to be effectively identified, thus balancing identification accuracy and false alarm rate.

[0058] In this embodiment, the test signal is input into the trained deep learning model to obtain the test signal reconstruction error, including: Preprocess the test signal; The preprocessed test signal is input into the trained deep learning model, and the reconstructed signal corresponding to the test signal is obtained by extracting features from the encoding layer and reconstructing the signal from the decoding layer. The difference between the test signal, the original signal, and the reconstructed signal is calculated to obtain the test signal reconstruction error.

[0059] In this embodiment, based on the comparison between the test signal reconstruction error and the anomaly judgment threshold, the abnormal state of the stay cable is identified, and the anomaly identification result of the stay cable is obtained, including: If the reconstruction error of the test signal exceeds the anomaly judgment threshold, the corresponding cable is judged to be abnormal; If the reconstruction error of the test signal does not exceed the anomaly judgment threshold, the corresponding cable is determined to be in a healthy state.

[0060] Specifically, the actual monitored test signals are input into the trained model, and anomaly identification is achieved by comparing the reconstruction error with a threshold. The specific steps are as follows: Test signal preprocessing: The test signal is preprocessed in the same way as the preprocessing of the training feature data in the previous steps. The principle is to ensure that the distribution characteristics of the test signal and the training data are consistent, and to avoid reconstruction error distortion caused by differences in preprocessing.

[0061] Calculate the test signal reconstruction error: Input the preprocessed test signal into the trained 1D-CAE model, extract features from the encoding layer and reconstruct the signal from the decoding layer to obtain the reconstructed signal corresponding to the test signal; calculate the difference between the original test signal and the reconstructed signal to obtain the test signal reconstruction error.

[0062] Abnormal state determination: The reconstruction error of the test signal is compared with the abnormal judgment threshold determined in the previous steps. If the reconstruction error of the test signal exceeds the abnormal judgment threshold, it indicates that the characteristics of the test signal are significantly different from the healthy characteristics learned by the model, and the corresponding cable is determined to be abnormal. If the reconstruction error of the test signal does not exceed the abnormal judgment threshold, it indicates that the test signal conforms to the healthy characteristic pattern, and the corresponding cable is determined to be in a healthy state.

[0063] The principle behind this step is as follows: When a cable stays cable is damaged, the characteristics of its vibration signal will deviate from those of a healthy state. The model cannot accurately reconstruct the signal, resulting in a significant increase in reconstruction error that exceeds the threshold. Conversely, the characteristics of a healthy signal are consistent with the learning pattern of the model, and the reconstruction error is smaller and within the threshold range. This logic enables unsupervised anomaly identification.

[0064] We selected measured sample cells of cable acceleration for ICEEMDON decomposition to obtain the characteristic frequencies of each IMF component. The waveforms of the first nine IMFs are shown. It can be seen that the IMFs are arranged sequentially from high frequency to low frequency, and the modal components are complete and continuous. The high-frequency component waveforms are complex, containing details of the structural dynamic response; the low-frequency component waveforms tend to be smoother, reflecting the gradual change components of the signal.

[0065] The Hilbert marginal spectrum corresponding to the IMF component reflects the energy distribution of each mode in the frequency domain.

[0066] Depend on Figure 4 It can be seen that the main peak frequency distribution of each IMF is clear. The marginal spectrum overlap between adjacent IMFs is low, with an average overlap of 0.62, indicating that the decomposition process successfully alleviates the mode aliasing phenomenon. The characteristic statistics of the IMFs are shown in the table below. The asymptotic MSE of the IMFs shows a monotonically decreasing trend with increasing order, eventually achieving complete reconstruction. Energy distribution analysis shows that the first four IMF components cumulatively carry 99.44% of the signal energy. The Hurst exponent of the IMFs shows a monotonically increasing trend with increasing order, revealing the multi-scale dynamic characteristics of the signal from anti-persistence to strong persistence.

[0067] Table 1. IMF Characteristic Statistics Table 2 shows the decomposition results of ICEEMDAN and CEEMDAN for acceleration samples. As can be seen from Table 2, ICEEMDAN reduces the reconstruction error by 4.6%, improves the signal-to-noise ratio by 0.22 dB, and improves the orthogonality index by 11.0% compared with CEEMDAN. This proves that ICEEMDAN further improves the decomposition accuracy and mode separation effect through the improved adaptive noise injection mechanism, and is more suitable for the accurate analysis and feature extraction of cable-stayed bridge vibration signals.

[0068] Table 2 Comparison results of ICEEMDAN and CEEMDAN Based on the above conclusions, the first four IMF components obtained from the measured cable force acceleration signal through ICEEMDAN decomposition were reconstructed using an ensemble, and the reconstructed signal was then input into a one-dimensional convolutional autoencoder (1D-CAE) for training. The changes in the loss functions of the training and validation sets with the number of iterations during training are shown below. Figure 6 As shown.

[0069] It can be seen that both loss curves decrease with the increase of the number of iterations and gradually tend to stabilize, indicating that the 1D-CAE model effectively learns the intrinsic characteristics of the cable vibration signal during the training process.

[0070] The performance metrics of the loss values ​​are shown in Table 3. The final loss values ​​of the training set and the validation set are basically the same, the overfitting rate is close to 1, and the generalization error is extremely small, indicating that the model has not experienced overfitting.

[0071] The results show that the patterns learned by the model in the training set can be well transferred to the validation set, reflecting that the model captures the universal essential features in the data, rather than noise or random patterns specific to the training set.

[0072] Table 3 Performance metrics of loss values Table 4 presents the statistical analysis parameters of the reconstruction errors based on the training and validation sets of health status monitoring data. As shown in Table 4, the values ​​of both are close and of the same order of magnitude, indicating that the model performs stably on different datasets. The overall standard deviation is 0.00083, which is relatively small and reflects the concentration of the reconstruction error distribution under normal operating conditions.

[0073] These parameters accurately characterize the vibration features of the stay cables under healthy conditions. Figure 7 , Figure 8 It can be seen that the reconstruction errors of the training set and validation set basically conform to a normal distribution. According to the normal distribution 3 In principle, the numerical distribution is in The probability between them is 0.9974, therefore... The obtained value of 0.00469 was set as the threshold for anomaly detection and applied to anomaly identification on the test set.

[0074] Table 4 Statistical Analysis Parameters for Reconstruction Errors of Training and Validation Sets Table 5. KS Test Results for Reconstruction Error The test set reconstruction error distribution and anomaly identification results are as follows: Figure 9 As shown. By Figure 9 As can be seen, the reconstruction errors of most samples are densely distributed in the lower numerical range and are completely below the threshold, demonstrating the stability of the model. There are two outlier sample points with reconstruction errors exceeding the threshold, and their deviations are significant, confirming the model's ability to sensitively identify real anomalies.

[0075] Based on the preceding analysis, this threshold is set based on a statistical guarantee rate of 99.74%, theoretically allowing for a very small number of false positives in healthy samples. The actual false positive rate was only 0.56%, consistent with statistical expectations.

[0076] To verify the anomaly identification capability of the proposed ICEEMDAN-1D-CAE method, three simulated abnormal working conditions were constructed: wire breakage, anchor loosening, and cable force attenuation. For each working condition, 60 abnormal samples were set and randomly mixed into the healthy test set to achieve coverage verification of different damage modes, as shown in Table 6.

[0077] The final recognition performance results are shown below. Figure 10 As shown in Table 7, the model achieves an accuracy rate exceeding 95% for identifying three typical anomalies: wire breakage, anchor loosening, and cable tension decay, with wire breakage exhibiting the best performance. The model has a low false alarm rate and can accurately identify early-stage, hidden anomalies in stay cables without requiring damage to prior labels.

[0078] Table 6. Abnormal Operating Condition Simulation Methods Table 7 Abnormal Signal Identification Results This paper proposes an unsupervised anomaly identification method for cable-stayed bridges that integrates ICEEMDAN and 1D-CAE. Through theoretical analysis and verification with experimental data, the following conclusions are drawn: The first four IMF components obtained from ICEEMDAN decomposition carry a cumulative 99.44% of the signal energy. Compared with CEEMDAN, ICEEMDAN has better decomposition performance: the reconstruction error is reduced by 4.6 percentage points, the signal-to-noise ratio is improved by 0.22 dB, and the orthogonality index is improved by 11.0%. The 1D-CAE model training process is stable, with final losses of 0.468 and 0.475 for the training and validation sets, respectively, showing no overfitting. Its reconstruction error conforms to a normal distribution, with an overall mean of 0.00219 and a standard deviation of 0.00083. The threshold of 0.00469, set based on the 3σ principle, can effectively distinguish between healthy and abnormal states.

[0079] This combined method has high recognition accuracy for three typical anomalies: wire breakage, anchor loosening, and cable tension decay. The recognition accuracy for all three anomalies exceeds 95%, with a low false alarm rate, and it does not rely on prior damage labels.

[0080] In other embodiments, a cable-stayed bridge anomaly identification system based on the fusion of signal decomposition and deep learning is provided, and a cable-stayed bridge anomaly identification method based on the fusion of signal decomposition and deep learning, based on any one of the preceding embodiments, includes: The acquisition module is configured to acquire raw data related to the vibration of the stay cables. The preprocessing module is configured to perform signal decomposition preprocessing and feature extraction on the raw data to obtain the feature data to be trained. The training module is configured to input the feature data to be trained into the deep learning model for training, and establish a mapping relationship between the vibration signal of the cable-stayed bridge and its health status. The calculation module is configured to calculate the signal reconstruction error based on the trained deep learning model and obtain the anomaly judgment threshold. The input module is configured to input the test signal into the trained deep learning model to obtain the test signal reconstruction error. The comparison module is configured to identify the abnormal state of the cable-stayed bridge based on the comparison result between the test signal reconstruction error and the anomaly judgment threshold, and obtain the cable-stayed bridge anomaly identification result.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 therein. Such 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, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for anomaly identification of cable-stayed bridges based on the fusion of signal decomposition and deep learning, characterized in that, include: Obtain raw data related to the vibration of the cable-stayed bridge, including cable vibration acceleration data, cable force data, and environmental impact data. The original data is subjected to signal decomposition preprocessing and feature extraction to obtain the feature data to be trained; The feature data to be trained is input into a deep learning model for training to establish a mapping relationship between the vibration signal and health status of the cable-stayed bridge. The signal reconstruction error is calculated based on the trained deep learning model to obtain the anomaly judgment threshold. The test signal is input into the trained deep learning model to obtain the test signal reconstruction error; Based on the comparison between the test signal reconstruction error and the anomaly judgment threshold, the abnormal state of the cable is identified, and the cable anomaly identification result is obtained.

2. The method for anomaly identification of cable-stayed bridges based on signal decomposition and deep learning fusion as described in claim 1, characterized in that, After obtaining the raw data related to the vibration of the cable-stayed bridge, the process also includes three-dimensional processing of the raw data: The collected cable vibration acceleration data were reconstructed in three dimensions to obtain the three-dimensional vibration data of the cable. The three-dimensional vibration data of the cable-stayed bridge is segmented to obtain several three-dimensional sub-block data. The three-dimensional sub-block data is converted using the maximum intensity projection method to obtain two-dimensional image data; The two-dimensional image data is divided using a fast random partitioning algorithm to obtain multiple datasets. Feature filtering is performed on multiple datasets to obtain multiple training feature data.

3. The method for anomaly identification of cable-stayed bridges based on signal decomposition and deep learning fusion as described in claim 1, characterized in that, The original data undergoes signal decomposition preprocessing and feature extraction to obtain the training feature data, including: The original data is adaptively decomposed using an adaptive noise complete set empirical mode decomposition algorithm to obtain several mode components and residual components. Analyze the energy proportion of each modal component to select the modal components that carry the core energy of the signal; The filtered modal components are reconstructed to obtain the feature-enhanced signal.

4. The method for anomaly identification of cable-stayed bridges based on signal decomposition and deep learning fusion as described in claim 3, characterized in that, The original data is adaptively decomposed using an adaptive noise complete set empirical mode decomposition algorithm to obtain several mode components and residual components, including: Add Gaussian white noise with a preset standard deviation to the original data to obtain a noisy signal; Empirical mode decomposition is performed on the noisy signal to calculate the local average value of the signal, thereby obtaining the first-order mode component and the first-order residual value; Repeat the above steps to obtain the modal components and corresponding residual values ​​of each stage in sequence until the decomposition meets the preset termination condition, and obtain all modal components and residual components.

5. The method for anomaly identification of cable-stayed bridges based on signal decomposition and deep learning fusion as described in claim 1, characterized in that, The deep learning model is a one-dimensional convolutional autoencoder, and constructing the one-dimensional convolutional autoencoder includes: An encoding layer is constructed, which is equipped with several convolutional filters to extract features from the input training feature data through convolution calculation; A pooling layer is set after the convolutional layer, and pooling operation is performed using a pooling kernel of a preset size; A decoding layer is constructed, which is equipped with several deconvolution filters. The signal is reconstructed from the feature vector output by the coding layer through upsampling and deconvolution calculations. A fully connected layer is set after the decoding layer, and a loss function based on the difference between the reconstructed signal and the original signal is set to obtain a complete one-dimensional convolutional autoencoder.

6. The method for anomaly identification of cable-stayed bridges based on signal decomposition and deep learning fusion as described in claim 1, characterized in that, The training involves inputting the feature data to be trained into a deep learning model for training, including: Multiple datasets obtained through a fast random partitioning algorithm are allocated into training and validation sets according to a preset ratio. The training set is input into the encoding layer of a one-dimensional convolutional autoencoder, and deep features of the signal are extracted through convolution and pooling operations to obtain feature vectors. The feature vector is input into the decoding layer, and the signal is reconstructed through upsampling and deconvolution operations. The error between the reconstructed signal and the original signal in the training set is calculated based on the loss function. The parameters and bias values ​​of the convolutional filter and deconvolutional filter in the model are updated through the backpropagation algorithm. The process is iterated repeatedly until the model loss value tends to stabilize. The validation set is input into the trained one-dimensional convolutional autoencoder to calculate the reconstruction error of the validation set signal; The model's generalization ability is evaluated by comparing the difference between the reconstruction error on the validation set and the reconstruction error on the training set. If the difference exceeds the preset range, adjust the parameters and bias values ​​of the convolutional filter and deconvolutional filter in the model, and iterate the training again. If the difference is within a preset range, the model training is complete.

7. The method for anomaly identification of cable-stayed bridges based on signal decomposition and deep learning fusion as described in claim 1, characterized in that, The signal reconstruction error is calculated based on the trained deep learning model to obtain the anomaly detection threshold, including: The training set signal and validation set signal obtained after training are input into the trained deep learning model to obtain the corresponding training set reconstruction error and validation set reconstruction error. The overall mean and standard deviation of the reconstruction errors of the training set and the validation set are calculated. Based on the population mean and population standard deviation, and combined with the statistical principle of normal distribution, the anomaly judgment threshold is obtained.

8. The method for anomaly identification of cable-stayed bridges based on signal decomposition and deep learning fusion as described in claim 1, characterized in that, The test signal is input into the trained deep learning model to obtain the test signal reconstruction error, including: The test signal is preprocessed; The preprocessed test signal is input into the trained deep learning model, and the reconstructed signal corresponding to the test signal is obtained by extracting features from the encoding layer and reconstructing the signal from the decoding layer. The difference between the test signal, the original signal, and the reconstructed signal is calculated to obtain the test signal reconstruction error.

9. The method for anomaly identification of cable-stayed bridges based on signal decomposition and deep learning fusion as described in claim 1, characterized in that, Based on the comparison between the test signal reconstruction error and the anomaly judgment threshold, the abnormal state of the stay cable is identified, and the anomaly identification result of the stay cable is obtained, including: If the reconstruction error of the test signal exceeds the anomaly judgment threshold, then the corresponding cable is determined to be abnormal. If the reconstruction error of the test signal does not exceed the anomaly judgment threshold, the corresponding cable is determined to be in a healthy state.

10. A cable-stayed bridge anomaly identification system based on the fusion of signal decomposition and deep learning, characterized in that, A method for anomaly identification of cable-stayed bridges based on the fusion of signal decomposition and deep learning, according to any one of claims 1 to 9, includes: The acquisition module is configured to acquire raw data related to the vibration of the stay cables. The preprocessing module is configured to perform signal decomposition preprocessing and feature extraction on the original data to obtain the feature data to be trained. The training module is configured to input the feature data to be trained into a deep learning model for training, and to establish a mapping relationship between the vibration signal of the cable-stayed bridge and its health status. The calculation module is configured to calculate the signal reconstruction error based on the trained deep learning model and obtain the anomaly judgment threshold. The input module is configured to input the test signal into the trained deep learning model to obtain the test signal reconstruction error. The comparison module is configured to identify the abnormal state of the cable-stayed bridge based on the comparison result between the reconstruction error of the test signal and the anomaly judgment threshold, and obtain the anomaly identification result of the cable-stayed bridge.

Images