Stay cable damage identification method and system and medium

Abstract

The invention relates to the field of stayed cable health monitoring of a cable-stayed bridge, and particularly discloses a stayed cable damage identification method and system and a medium, and the method comprises the steps: obtaining historical piezoelectric signals, collected by a flexible piezoelectric sensor, of a stayed cable in various damage states under the action of an external load, and carrying out the preprocessing of the historical piezoelectric signals; the preprocessed piezoelectric signal data is converted into Gramb angle field image data by adopting a Gramb angle field algorithm; expanding a damaged sample data set by using a conditional variation generative adversarial network; constructing a CNN (Convolutional Neural Network) and Mama network coupling model; training the coupling model by using the expanded damage sample data set to obtain a stay cable damage identification model; and after the piezoelectric signals collected in real time are preprocessed and converted through the Grubrum angle field algorithm, the piezoelectric signals are input into the stay cable damage recognition model, and a stay cable damage recognition result is output. The method has the advantages of stable signal, accurate identification, high adaptability and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of cable-stayed bridge cable health monitoring, and particularly relates to a cable-stayed cable damage identification method and system based on flexible piezoelectric sensing and deep learning, and a medium. BACKGROUND

[0002] In modern transportation networks, bridges are an indispensable part and often occupy a strategic position, being the throat of the entire transportation network. Among many bridge types, cable-stayed bridges are favored due to their large span capacity, excellent mechanical properties, modern and beautiful appearance, and many other advantages. Steel strand cable systems are widely used in cable-stayed bridge cables due to their flexibility and convenience in installation, excellent corrosion resistance and durability, excellent detectability, maintainability and replaceability, and excellent economy.

[0003] As the core load-bearing component of cable-stayed bridges, cable-stayed cables act as the "lifeline" connecting the bridge tower and the bridge deck. Their core function is to efficiently transfer most of the dead load (such as bridge self-weight) and live load (such as vehicles and pedestrians) borne by the bridge deck in tension to the tower, and then to the foundation. The fracture of cable-stayed cables will cause the mechanical system of cable-stayed bridges to be unbalanced, leading to structural instability and overall deformation, causing traffic disruption, seriously affecting regional economy and people's travel, and even causing significant casualties and incalculable social impact.

[0004] During the service period of cable-stayed bridges, cable-stayed cables are subjected to repeated dynamic loads such as vehicle and wind-induced vibration. Microscopic fatigue cracks may occur at surface defects or stress concentration points due to fatigue. Meanwhile, when the HDPE sheath on the outside of the cable is damaged due to aging, moisture in the air will cause corrosion of the steel cable and eventually cause cable wire breakage, seriously threatening the safety of cable-stayed bridges. In order to discover and solve problems in a timely manner, real-time state monitoring of the cable is required.

[0005] Traditional cable-stayed bridge cable damage monitoring techniques have some varying degrees of shortcomings: Visual inspection: can monitor macroscopic surface defects, but cannot detect small defects and cannot understand internal damage and defects of cable-stayed cables; Fiber Bragg grating monitoring: can achieve absolute measurement, is resistant to electromagnetic interference and corrosion, but is not sensitive to small and early local damage, and maintenance and replacement of sensors is relatively difficult; Ultrasonic monitoring: can infer the extent and location of damage, but has poor long-distance damage detection accuracy and cannot penetrate the HDPE film; Magnetostrictive guided wave monitoring: strong excitation energy, long propagation distance, but relatively low sensitivity to early small damage and monitoring results are easily affected by boundary conditions; Radiographic monitoring: high detection accuracy and intuitiveness, providing intuitive imaging capability for cable health monitoring, but its equipment is bulky and expensive, detection speed is slow, and it is harmful to the human body; Magnetic flux leakage monitoring: sensitive to broken wire defects of steel strand cable-stayed cables, high reliability, but it relies on sensor scanning by section, poor real-time performance, not sensitive to early stress concentration and micro-cracks; Metal magnetic memory monitoring: can indirectly identify stress concentration area and achieve early warning, but the monitoring results are easily affected by external magnetic fields; Compared with the above cable-stayed bridge cable damage monitoring technologies, piezoelectric sensing detection technology as a dynamic non-destructive testing technology is sensitive to high-frequency, dynamic local damage (such as micro-crack impact), can capture micro dynamic signals such as stress waves, a single sensor can monitor the damage state in a long distance range, and can capture the dynamic process of damage initiation and expansion in real time and online.

[0006] Traditional rigid piezoelectric sensors (represented by PZT ceramics) are hard, stable in performance, and have strong excitation capability, but their inherent brittleness leads to the coupling of flexible cables depending on adhesives, forming a "hard-soft" contact. This point coupling is easy to fail due to aging of the adhesive layer under long-term wind and rain vibration, temperature cycling, leading to signal attenuation or even sensor falling off, and the sparse distribution mode makes it difficult to achieve accurate damage positioning.

[0007] Traditional signal processing methods (Fourier transform, wavelet transform) fail to fully combine the physical mechanism of structural damage, and have many shortcomings in damage diagnosis: Fourier transform: when processing time-frequency information, the time domain information will be completely lost, and it is not sensitive to transient signal mutations, and the method requires that the signal be stationary, while the signal of the engineering structure is essentially non-stationary during the damage occurrence and development process; Wavelet transform: the effect of wavelet analysis is highly dependent on the selection of wavelet basis functions, and its characterization ability for complex damage patterns is limited.

[0008] Deep learning technology has strong ability in time series data modeling, but its application in structural health monitoring still faces the following challenges: High-quality, full-life-cycle data is scarce: there is almost no data for large-scale engineering structures to completely fail, we can only obtain data in healthy and slightly damaged states, resulting in sample data scarcity and imbalance, and the cost of sample data labeling is high; Robustness and generalization ability are poor: deep learning technology is easily affected by noise when processing data, making it difficult to effectively diagnose and decide on the true damage condition.

[0009] Real-time and insufficient computing efficiency: although the deep learning technology has fast inference speed, it needs a large amount of computing resources and time when training a complex model, and its application in a real-time monitoring scene is a challenge. SUMMARY

[0010] In view of the deficiencies in the prior art, the present application provides a cable damage identification method, system and medium, which realizes high-precision identification of cable damage state.

[0011] In a first aspect, a cable damage identification method is provided, comprising the following steps: Obtaining historical piezoelectric signals of the cable under external load in various damage states collected by a flexible piezoelectric sensor, and preprocessing the piezoelectric signals; Converting the preprocessed piezoelectric signal data into Gram angle field image data using the Gram angle field algorithm; Using a conditional variational generative adversarial network to expand a damage sample data set, each sample including a combined input of piezoelectric signals and Gram angle field images and a corresponding damage class label; Building a CNN and Mamba network coupled model, CNN extracts features with Gram angle field images as input, then the output of CNN and piezoelectric signals are jointly input into Mamba network, and the classification head predicts the damage class based on the output of Mamba network; Training the coupled model using the expanded damage sample data set to obtain a cable damage identification model; Preprocessing and converting the real-time collected piezoelectric signals using the Gram angle field algorithm, and inputting the piezoelectric signals into the cable damage identification model to output the cable damage identification result.

[0012] Further, the preprocessing process includes denoising, segmenting and filtering, and normalizing the piezoelectric signals.

[0013] Further, the preprocessed piezoelectric signal data is converted into Gram angle field image data using the Gram angle field algorithm, specifically including: Re-normalizing the preprocessed one-dimensional piezoelectric signals to the interval [-1, 1]; Mapping the normalized signals to a polar coordinate system; Calculating the Gram angle field matrix based on the Gram angle and field or Gram angle difference field algorithm; Scaling the Gram angle field matrix to the pixel range of a gray image to generate a two-dimensional Gram angle field image.

[0014] Further, the damage signal sample data set is expanded using a conditional variational generative adversarial network, specifically including: An encoder, generator, and discriminator are constructed, and conditional variables representing the damage category are introduced into the input of all three. The encoder, generator, and discriminator are trained by a multi-iteration cyclic training method. After training, the generator is saved. By inputting conditional variables specifying the damage category and random noise sampled from a standard normal distribution into the generator, a new Gram angular field image conforming to the characteristics of the specified damage category is generated and combined with a piezoelectric signal under the same damage category to generate new samples, thereby expanding the damage sample dataset.

[0015] Furthermore, the CNN and Mamba network coupling model includes: CNN: A multi-layered stacked structure, each layer consisting of convolutional layers, non-linear activation layers, and pooling layers arranged sequentially; its input is a Gram angular field image; Mamba network: includes the Vision Mamba module and the cross-modal Mamba module; The Vision Mamba module takes piezoelectric signals and CNN outputs as inputs and performs global context modeling and feature refinement for each modality feature through a selective state-space model. The cross-modal Mamba module utilizes a cross-attention mechanism to fuse the two modal features output by the Vision Mamba module; Classification Head: Predicts the output damage category based on fused features from cross-modal Mamba module outputs.

[0016] Furthermore, when the cross-modal Mamba module uses the cross-attention mechanism to fuse the two modal features output by the Vision Mamba module, it uses the feature of one modality as the query vector and the feature of the other modality as the key vector and value vector.

[0017] Secondly, a cable-stayed bridge damage identification system is provided, including: Several flexible piezoelectric sensing acquisition devices are used to acquire piezoelectric signals of the stay cables under external loads; The data transmission module is used to transmit the data collected by the plurality of flexible piezoelectric sensing acquisition devices to the data processing module; The data processing module is equipped with a cable-stayed bridge damage identification model. This model preprocesses and converts the real-time acquired piezoelectric signals using the Gram angle field algorithm, then inputs the data into the cable-stayed bridge damage identification model and outputs the cable-stayed bridge damage identification results. The cable-stayed bridge damage identification model is obtained using the cable-stayed bridge damage identification method described above.

[0018] Furthermore, the flexible piezoelectric sensing acquisition device includes two sets of clamps, multiple cylindrical connecting rods, and a flexible piezoelectric sensor; the two ends of the multiple cylindrical connecting rods are respectively connected to the two sets of clamps to form an annular clamp; the flexible piezoelectric sensor is arranged on the surface of the multiple cylindrical connecting rods.

[0019] Thirdly, a cable-stayed bridge damage identification system is provided, including: A memory on which computer programs are stored; A processor is used to load and execute the computer program to implement the cable-stayed bridge damage identification method as described above.

[0020] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the cable-stayed bridge damage identification method as described above.

[0021] This invention proposes a method, system, and medium for identifying damage to stay cables, which has the following advantages compared with existing technologies: (1) Strong data generation capability, effectively solving the problems of sample scarcity and class imbalance: Conditional Variational Generative Adversarial Network (CVAE-GAN) can accurately generate target samples according to the specified damage category by introducing conditional labels, thereby effectively solving the problem of class imbalance caused by the scarcity of data on certain damage states. At the same time, its variational autoencoder structure can learn a continuous and smooth latent space distribution, which not only ensures the diversity and rationality of the generated samples, but also simulates the continuous evolution process of damage through latent space interpolation, thereby generating more physically meaningful transition state samples, significantly improving the generalization ability and early warning accuracy of subsequent damage identification models.

[0022] (2) High computational efficiency: The self-attention mechanism of Transformer requires calculating the relationship between each token in the sequence and all other tokens. Its computational cost and memory usage increase quadratically with the length of the sequence. Mamba network replaces the core self-attention mechanism of Transformer with Selective State Space Model (SSSM), which has a linear complexity, thus solving the computational bottleneck of Transformer in long sequence processing.

[0023] (3) High accuracy in damage signal recognition: The constructed CNN-MMN hybrid model fully leverages the advantages of convolutional neural networks (CNN) in local feature extraction and multimodal Mamba networks (MMN) in long-sequence global context modeling. Among them, the Vision Mamba module can efficiently encode image sequences and dynamically select key information related to damage; the cross-modal Mamba module deeply integrates the complementary information of piezoelectric signals and Gram angle field images. This architecture significantly improves the accuracy of damage recognition.

[0024] (4) Safety of full life cycle monitoring: Traditionally, piezoelectric sensors are directly attached to the surface of the steel strand cable to monitor the cable's health status. However, due to external factors such as ultraviolet rays and wind and rain, the adhesive will age and the sensor will fall off. At the same time, due to the gaps in the steel strand cable itself, the piezoelectric sensor does not provide full coverage monitoring of the steel strand cable's status, resulting in inaccurate monitoring signals. By attaching the piezoelectric sensor to the connecting rod of the clamp in full coverage, the health status of the steel strand cable can be indirectly monitored by monitoring the signal status of the connecting rod. This not only improves the accuracy of monitoring but also avoids the damage to the steel strand cable caused by directly grinding and cleaning the surface of the steel strand cable to attach the sensor. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are 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.

[0026] Figure 1 This is a flowchart of the cable-stayed bridge damage identification method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the ring clamp structure provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0028] In the description of this invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "top," "bottom," "inner," "outer," and "center," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. When an element is referred to as being "fixed to" another element, it can be directly on the other element or there may be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element present. The terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or order. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0029] To address the shortcomings of existing technologies, this invention provides a method, system, and medium for identifying damage to stay cables. Based on data acquired through flexible piezoelectric sensing and deep learning, it achieves high-precision identification of cable damage conditions. The technical solution of this invention is described in detail below with reference to specific embodiments.

[0030] like Figure 1 As shown, this embodiment of the invention provides a method for identifying damage to stay cables, including the following steps: S1: Acquire historical piezoelectric signals of the stay cable under external load in various damage states collected by the flexible piezoelectric sensor, and preprocess them.

[0031] Before using a flexible piezoelectric sensor to collect piezoelectric signals, it is necessary to determine the placement and relevant parameters of the flexible piezoelectric sensor. The specific process is as follows: S11: Establish a finite element simulation model with known physical parameters based on the cable-stayed components, and find the stress concentration area of ​​the cable-stayed steel strands under stress through finite element analysis; S12: Deploy flexible piezoelectric sensors in the stress concentration areas of the cable-stayed bridge; S13: Flexible piezoelectric sensor parameter settings: Center frequency range: 100kHz-250kHz (for waveguides with helical structures, such as cable-stayed steel strands, this frequency band can achieve a good balance between damage sensitivity and propagation distance). Sampling frequency: Follows the Nyquist sampling theorem (the sampling frequency must be at least twice the highest frequency of the signal); S14: Signal Acquisition Acquisition trigger mode: synchronous trigger (each data acquisition is strictly synchronized with the issuance of the excitation signal); Data collection duration: Sampling duration > 2 × cable length / group velocity, the group velocity needs to be estimated in advance through experiments or theoretical calculations; Flexible piezoelectric sensors (represented by PVDF and nanocomposite materials) utilize the flexibility and ductility of their materials to achieve large-area, conformal bonding, much like "electronic skin." This not only significantly improves the long-term consistency and reliability of monitoring signals but also enables them to be deployed in a distributed, arrayed manner to capture the global strain and local dynamic response of the cable-stayed bridge surface. This overcomes the shortcomings of traditional rigid sensors' incomplete "point-based" measurement coverage, laying a solid foundation for accurate damage identification and localization.

[0032] After signal acquisition is completed, the acquired piezoelectric signals need to be preprocessed. The preprocessing process includes noise reduction, segmentation and normalization of the acquired historical piezoelectric signals.

[0033] Specifically, the noise reduction uses the wavelet thresholding method: select an appropriate wavelet basis function to decompose the piezoelectric signal into different scales and positions, set a threshold, set the wavelet coefficients smaller than the threshold to zero or shrink them, identify them as noise, and then reconstruct the signal.

[0034] The segmented screening uses a fixed-length sliding window method: a fixed-length window slides across the piezoelectric signal with a certain overlap rate to extract a series of data segments of equal length, and selects data segments whose peak values ​​exceed a set peak threshold, thereby generating a large amount of basic sample data.

[0035] The normalization process uses Z-Score standardization, X norm = (X - μ) / σ, X and X norm Here, μ represents the signal before and after normalization, and σ represents the signal mean. The data is transformed into a distribution with a mean of 0 and a standard deviation of 1.

[0036] The acquired historical piezoelectric signals must include piezoelectric signals under healthy conditions and piezoelectric signals under various damage conditions of the stay cables, and be labeled with damage categories, including healthy, microcracks, and broken wires.

[0037] S2: The Gram angle field algorithm is used to convert the preprocessed piezoelectric signal data into Gram angle field image data. While preserving the integrity of the signal, it also preserves the signal's dependence on time, which facilitates the use of CNN's powerful image classification capabilities for learning in the later stage, making it more effective for cable-stayed bridge damage identification tasks.

[0038] Specifically, the conversion process includes: S21: Signal renormalization: The preprocessed one-dimensional piezoelectric signal is renormalized to the [-1,1] interval using maximum-min normalization; S22: Map the normalized signal to the polar coordinate system: Represent the one-dimensional piezoelectric signal data (time series data) as points in the polar coordinate system, thereby converting time information into angles and radii; S23: Calculate the Gram angle field matrix (GAF matrix) based on the Gram angle sum field or Gram angle difference field algorithm; where the Gram angle sum field is the cosine value of the sum of the angles of every two points, and the Gram angle difference field is the sine value of the difference of the angles of every two points. In specific implementation, only one of them needs to be selected. S24: The range of the Gram angle field matrix is ​​[-1,1]. Scale the Gram angle field matrix to the grayscale image pixel range of [0,255] to generate a grayscale image, which is a two-dimensional Gram angle field image.

[0039] S3: Use conditional variational generative adversarial networks to augment the damage sample dataset. Each sample includes a combined input consisting of a piezoelectric signal and a Gram angular field image, along with the corresponding damage category label.

[0040] Specifically, the dataset of augmented damage samples using conditional variational generative adversarial networks includes: An encoder, generator, and discriminator are constructed, and conditional variables representing damage categories are introduced into the input of all three. The network is trained by jointly optimizing the reconstruction loss, KL divergence loss, adversarial loss, and conditional classification loss. During training, a multi-iteration cyclic training method is used to train the encoder, generator, and discriminator. After training, the generator is saved. By inputting conditional variables specifying the damage category and random noise sampled from a standard normal distribution into the generator, a new Gram angular field image conforming to the characteristics of the specified damage category is generated and combined with a piezoelectric signal under the same damage category to generate new samples, thereby expanding the damage sample dataset.

[0041] More specifically, the process of using conditional variational generation to augment the dataset of damaged samples using adversarial networks includes: S31: Data Preparation and Preprocessing The Gram angle field image data processed in step S2 are used to construct a dataset; Size uniformity: Adjust all Gram field images to the same size; Pixel normalization: Normalize the pixel values ​​of the Gram angular field image from [0, 255] to the range [-1, 1].

[0042] S32: Model Building Define condition variables: Convert damage category labels (e.g., 0: healthy, 1: microcrack, 2: broken wire) into unique thermal coding form to serve as condition variable c; Build the encoder: Input: A real Gram angle field image and its corresponding conditional label; Structure: Consists of multiple convolutional layers, with progressive downsampling; Output: Two vectors of the same dimension, representing the mean and log-variance of the latent space, respectively; Build generator: Input: Latent variable z and condition variable c sampled from the distribution of encoder output; Network structure: It consists of multiple transposed convolutional layers for progressive upsampling, mapping the latent variable z and conditional variable c into a complete Gram angle field image; Output: The pseudo-Gram angle field image that satisfies the condition variable c; Constructing a discriminator: Input: An image (which can be a real Gram field image or a generated fake Gram field image) ) and its condition variable c (usually concatenated with the image feature map after depthwise convolution, or used as a spatial mapping); Network structure: It consists of multiple convolutional layers, used for progressive downsampling to extract features for judgment; Output: True / False Discrimination: A scalar representing the probability that the input image is real data; Category discrimination: A category probability distribution, representing the probability that the input image belongs to each damage category.

[0043] S33: Iterative training: Multiple iterations, in each iteration the discriminator is trained multiple times first, and then the generator and encoder are trained once; Discriminator training: Sample a small batch of real Gram angle field images x and their condition variable c; Generating pseudo-gram angle field images using an encoder and generator. ; The real Gram corner field image x (labeled 1) and the fake Gram corner field image (With label 0) Input each into the discriminator; Calculate the discriminator loss; the discriminator loss mainly consists of two parts: adversarial loss and gradient penalty loss. The adversarial loss is used to distinguish between real and generated images, and usually adopts the Wasserstein distance estimation form, that is, to maximize the difference between the real image discrimination score and minimize the generated image discrimination score. The gradient penalty loss, on the other hand, imposes a constraint on the gradient norm of the linear interpolation samples of real and generated images to ensure that the discriminator satisfies Lipschitz continuity. Backpropagation is used to update the discriminator parameters; Generator and encoder training: Generate another batch of fake Gram field images ; Calculate the total loss of the generator and encoder; the total loss mainly consists of four core losses: reconstruction loss (usually L1 or perceptual loss, to ensure that the input image can be accurately reconstructed), KL divergence loss (constraining the latent distribution of the encoder output to approximate a standard normal distribution), adversarial loss (making the generated image difficult to distinguish from the real image in the discriminator), and conditional classification loss (ensuring that the generated image can be correctly classified into the specified damage category). Backpropagation is performed, updating the parameters of both the generator and encoder simultaneously.

[0044] S34: Expanded Sample Generation: Save the generator model after training is complete; For a specific damage category that needs to be expanded, input its corresponding condition variable c, and sample random noise from a standard normal distribution. ; The generator outputs a new Gram angle field image that matches the characteristics of this damage category; The generated Gram angular field image is combined with the piezoelectric signal under the same damage category to generate new samples, thereby expanding the damage sample dataset.

[0045] The purpose of step S3 is to generate more data samples through Conditional Variational Generative Adversarial Network (CVAE-GAN) to improve the generalization ability of the model, and to use data augmentation to balance the number of samples of different classes in the monitoring data, reduce the impact of data skew on model training and performance, and improve the robustness and reliability of the model.

[0046] S4: Construct a coupled CNN and Mamba network model. The CNN extracts features from the Gram angle field image as input. Then, the output of the CNN and the piezoelectric signal are input into the Mamba network. The classification head predicts the damage category based on the output of the Mamba network.

[0047] In this embodiment, the CNN and Mamba network coupling model includes: CNN: Its input is a Gram-angle field image, and it adopts a multi-layer stacked structure. Each layer includes a convolutional layer, a non-linear activation layer and a pooling layer arranged in sequence. The convolutional layer slides the convolutional kernel on the input image to detect local patterns (such as edges and textures), and then introduces non-linearity through the ReLU activation function. The pooling layer then achieves feature dimensionality reduction and translation invariance enhancement. In practice, the number of stacked layers is preferably 4-6. Through layer-by-layer stacking, the network gradually abstracts the low-level pixel information into high-order features that represent the damage state, and finally outputs a feature vector for subsequent cross-modal Mamba network fusion and classification. Mamba network: includes the Vision Mamba module and the cross-modal Mamba module; The Vision Mamba module takes piezoelectric signals and CNN output as inputs. The CNN output needs to be converted into a sequence and injected with positional information. Then, a selective state space model is used to perform global context modeling and feature refinement on each modality feature, dynamically and selectively passing information down. The parameters of the selective state space model are not fixed, but dynamically generated according to the current token. Therefore, the model can selectively remember key information related to damage. Cross-modal Mamba module: The cross-attention mechanism is used to fuse the two modal features output by the Vision Mamba module to achieve deep feature interaction and information complementarity; specifically, the feature sequence of one modality is used as the query vector Q, and the feature sequence of the other modality is used as the key vector K and the value vector V to perform deep feature fusion. Classification head: Includes a global pooling layer and a fully connected layer, which predicts the output damage category based on the fused features of the cross-modal Mamba module output.

[0048] S5: The coupled model is trained using the expanded damage sample dataset to obtain the cable-stayed bridge damage identification model.

[0049] The expanded injury sample dataset is divided into training and validation sets, both of which must include samples in healthy and various injury states. The training set is then used to train a coupled CNN and Mamba network model, and the validation set is used to validate and evaluate the trained model. Through validation and evaluation, the model parameters are continuously optimized to improve the accuracy and reliability of the coupled model. During training, a weighted cross-entropy loss function is used to handle potential class imbalance. The optimizer is AdamW (initial learning rate set to 0.001, weight decay set to 0.01), and a cosine annealing hot-restart learning rate scheduler (initial period T0 set to 50 epochs) is used to dynamically adjust the training process.

[0050] S6: After preprocessing the real-time acquired piezoelectric signals and converting them using the Gram angle field algorithm, input the signals into the cable damage identification model and output the cable damage identification results.

[0051] Specifically, step S6 includes: S61: Signal Acquisition: Real-time piezoelectric signals in the stress concentration areas of the stay cable strands are acquired using flexible piezoelectric sensors; S62: Signal preprocessing: Signal denoising: The signal is decomposed into different scales and locations through wavelet transform. A threshold is set, and wavelet coefficients smaller than the threshold are set to zero or shrunk and identified as noise. Then the signal is reconstructed. Segmented filtering: The reconstructed signal is segmented using a fixed-length sliding window method, and the segment with the largest peak value is selected from it. Normalization: The selected piezoelectric signals are normalized by Z-Score to obtain the preprocessed real-time piezoelectric signals.

[0052] S63: Signal conversion: The preprocessed real-time piezoelectric signal data is converted into a real-time Gram angle field image using the Gram angle field algorithm; S64: Online model inference, which inputs the preprocessed real-time piezoelectric signal and the converted real-time Gram angle field image into the trained cable damage recognition model to predict and output the current damage state of the cable.

[0053] The cable-stayed bridge damage identification method provided in the above embodiments has the following advantages: With its strong data generation capabilities, Conditional Variational Generative Adversarial Networks (CVAE-GANs) effectively address the challenges of sample scarcity and class imbalance by introducing conditional labels. This allows for the precise generation of target samples based on specified damage categories, effectively resolving class imbalance caused by the scarcity of data for certain damage states. Furthermore, its variational autoencoder structure learns a continuous and smooth latent spatial distribution, ensuring not only the diversity and rationality of generated samples but also simulating the continuous evolution of damage through latent spatial interpolation. This generates more physically meaningful transitional state samples, significantly improving the generalization ability and early warning accuracy of subsequent damage identification models.

[0054] High computational efficiency: The Transformer's self-attention mechanism requires calculating the relationship between each token in the sequence and all other tokens. Its computational cost and memory usage increase quadratically with the sequence length. The Mamba network replaces the core self-attention mechanism of the Transformer with a Selective State Space Model (SSSM), which has a linear complexity, thus solving the computational bottleneck of the Transformer in processing long sequences.

[0055] High accuracy in damage signal recognition: The constructed CNN-MMN hybrid model fully leverages the advantages of Convolutional Neural Networks (CNNs) in local feature extraction and Multimodal Mamba Networks (MMNs) in long-sequence global context modeling. Specifically, the Vision Mamba module efficiently encodes image sequences and dynamically selects key information related to damage; the cross-modal Mamba module deeply integrates complementary information from piezoelectric signals and Gram angle field images. This architecture significantly improves damage recognition accuracy while ensuring efficient model operation thanks to the near-linear computational complexity of the Mamba model, meeting real-time monitoring requirements.

[0056] Achieving intelligent full-cycle diagnosis with outstanding early warning capabilities: This invention forms a complete automated processing flow from signal acquisition, preprocessing, image conversion, data enhancement to intelligent recognition. This system can perform online reasoning and accurate identification of multiple damage stages in cable stays, such as crack initiation, propagation, and wire breakage. It achieves a fundamental shift from periodic, offline inspections to continuous, online, and intelligent early warning, providing strong technical support for the safe operation and preventative maintenance of bridge structures.

[0057] This invention also provides a cable-stayed bridge damage identification system, comprising: Several flexible piezoelectric sensing acquisition devices are used to acquire piezoelectric signals of the stay cables under external loads; The data transmission module is used to transmit the data collected by the plurality of flexible piezoelectric sensing acquisition devices to the data processing module; The data processing module is equipped with a cable-stayed bridge damage identification model. This model preprocesses and converts the real-time acquired piezoelectric signals using the Gram angle field algorithm, then inputs the data into the cable-stayed bridge damage identification model and outputs the cable-stayed bridge damage identification results. The cable-stayed bridge damage identification model is obtained using the cable-stayed bridge damage identification method described above.

[0058] In this embodiment, the flexible piezoelectric sensing acquisition device includes two sets of clamps, multiple cylindrical connecting rods, and a flexible piezoelectric sensor; the two ends of the multiple cylindrical connecting rods are respectively connected to the two sets of clamps to form an annular clamp; the flexible piezoelectric sensor is attached to the surface of the multiple cylindrical connecting rods.

[0059] Figure 2An example of a ring clamp is given, in which each set of clamps consists of two semi-circular arc-shaped clamps 1. Threaded holes are machined on the ears of the two semi-circular arc-shaped clamps 1. The two semi-circular arc-shaped clamps 1 are arranged opposite each other and connected by bolts. The tightness of the clamps is adjusted by adjusting the bolts. In this embodiment, threaded holes are machined at the tri-division points of each semi-circular arc-shaped clamp for connection with cylindrical connecting rods 2. The two ends of the cylindrical connecting rods 2 are machined as threaded structures to connect threadedly with the threaded holes machined on the semi-circular arc-shaped clamps 1. In this example, there are four cylindrical connecting rods 2. In other embodiments, the number of cylindrical connecting rods 2 can be increased or decreased according to actual needs.

[0060] The purpose of the ring clamp design is to avoid damage to the steel strand cable caused by directly grinding and cleaning the surface for sensor adhesion, and to allow for repeated disassembly and assembly of the clamp, achieving "damage-free" monitoring throughout the cable's entire lifecycle, significantly reducing maintenance costs and the risks associated with personnel working at heights. In practice, to avoid fatigue and the adverse effects of the clamp's own weight, high-rigidity, lightweight materials are selected to manufacture the ring clamp. To ensure a reliable connection between the ring clamp and the cable, serrated anti-slip textures are machined on the inner surface of the clamp where it contacts the cable.

[0061] The cable damage identification system provided in the above embodiments, in addition to possessing the advantages of the aforementioned cable damage identification methods, also has the following advantages: Safety of full life-cycle monitoring: Traditional methods of directly attaching piezoelectric sensors to the surface of steel strand cables to monitor cable health can lead to sensor detachment due to adhesive aging caused by external factors such as ultraviolet radiation and wind and rain. Furthermore, due to the gaps inherent in the steel strand cables, the piezoelectric sensors do not provide complete coverage of the cable's condition, resulting in inaccurate monitoring signals. This is addressed by designing an indirect coupling scheme using a "ring clamp + connecting rod". This new method involves attaching piezoelectric sensors to the connecting rods of a clamp, completely covering the cable. This replaces the traditional method of directly attaching piezoelectric sensors to the surface of the cable strands. By monitoring the signal status of the connecting rods, the health of the cable strands is indirectly monitored. This "ring-like" installation eliminates the need for grinding or drilling into the cable sheath, completely avoiding damage to the cable itself during installation. It fundamentally solves the problem of sensor detachment caused by adhesive aging, improving monitoring accuracy and avoiding the damage caused by directly grinding and cleaning the cable strand surface to attach the sensor. The ring-shaped clamp can be repeatedly disassembled and reassembled, enabling safe and non-destructive monitoring of the cable throughout its entire lifecycle, significantly reducing the cost and risk of high-altitude maintenance.

[0062] Furthermore, embodiments of the present invention also provide a cable-stayed bridge damage identification system, comprising: A memory on which computer programs are stored; A processor is used to load and execute the computer program to implement the cable-stayed bridge damage identification method as described above.

[0063] In addition, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the cable-stayed bridge damage identification method as described above.

[0064] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0065] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0066] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0069] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for identifying damage to stay cables, characterized in that, Includes the following steps: The historical piezoelectric signals of the stay cable under external load under various damage states were acquired by a flexible piezoelectric sensor and preprocessed. The Gram angle field algorithm is used to convert the preprocessed piezoelectric signal data into Gram angle field image data; The damage sample dataset is augmented using a conditional variational generative adversarial network. Each sample includes a combined input consisting of a piezoelectric signal and a Gram angular field image, along with a corresponding damage category label. A coupled CNN and Mamba network model is constructed. The CNN extracts features from the Gram angle field image as input. Then, the output of the CNN and the piezoelectric signal are input into the Mamba network. The classification head predicts the damage category based on the output of the Mamba network. The coupled model was trained using the expanded damage sample dataset to obtain the cable-stayed bridge damage identification model; After preprocessing and transforming the real-time acquired piezoelectric signals using the Gram angle field algorithm, the signals are input into the cable-stayed bridge damage identification model, and the cable-stayed bridge damage identification results are output.

2. The cable-stayed bridge damage identification method according to claim 1, characterized in that, The preprocessing process includes: noise reduction, segmentation and normalization of the piezoelectric signal.

3. The cable-stayed bridge damage identification method according to claim 1, characterized in that, The Gram angle field algorithm is used to convert the preprocessed piezoelectric signal data into Gram angle field image data, specifically including: The preprocessed one-dimensional piezoelectric signal is renormalized to the [-1,1] interval; Map the normalized signal to the polar coordinate system; Calculate the Gram angle field matrix based on the Gram angle sum field or Gram angle difference field algorithm; The Gram angular field matrix is ​​scaled to the pixel range of the grayscale image to generate a two-dimensional Gram angular field image.

4. The cable-stayed bridge damage identification method according to claim 1, characterized in that, The conditional variational generative adversarial network is used to augment the dataset of damaged signal samples, specifically including: An encoder, generator, and discriminator are constructed, and conditional variables representing the damage category are introduced into the input of all three. The encoder, generator, and discriminator are trained by a multi-iteration cyclic training method. After training, the generator is saved. By inputting conditional variables specifying the damage category and random noise sampled from a standard normal distribution into the generator, a new Gram angular field image conforming to the characteristics of the specified damage category is generated and combined with a piezoelectric signal under the same damage category to generate new samples, thereby expanding the damage sample dataset.

5. The cable-stayed bridge damage identification method according to claim 1, characterized in that, The coupled CNN and Mamba network models include: CNN: A multi-layered stacked structure, each layer consisting of convolutional layers, non-linear activation layers, and pooling layers arranged sequentially; its input is a Gram angular field image; Mamba network: includes the Vision Mamba module and the cross-modal Mamba module; The Vision Mamba module takes piezoelectric signals and CNN outputs as inputs and performs global context modeling and feature refinement for each modality feature through a selective state-space model. The cross-modal Mamba module utilizes a cross-attention mechanism to fuse the two modal features output by the Vision Mamba module; Classification Head: Predicts the output damage category based on fused features from cross-modal Mamba module outputs.

6. The cable-stayed bridge damage identification method according to claim 5, characterized in that, When the cross-modal Mamba module uses the cross-attention mechanism to fuse the two modal features output by the Vision Mamba module, it uses the features of one modality as the query vector and the features of the other modality as the key vector and value vector.

7. A cable-stayed bridge damage identification system, characterized in that, include: Several flexible piezoelectric sensing acquisition devices are used to acquire piezoelectric signals of the stay cables under external loads; The data transmission module is used to transmit the data collected by the plurality of flexible piezoelectric sensing acquisition devices to the data processing module; The data processing module is equipped with a cable-stayed bridge damage identification model, which is used to preprocess the real-time acquired piezoelectric signals and convert them using the Gram angle field algorithm, input the data into the cable-stayed bridge damage identification model, and output the cable-stayed bridge damage identification results; wherein, the cable-stayed bridge damage identification model is obtained using the cable-stayed bridge damage identification method as described in any one of claims 1 to 6.

8. The cable-stayed bridge damage identification system according to claim 7, characterized in that, The flexible piezoelectric sensing acquisition device includes two sets of clamps, multiple cylindrical connecting rods, and a flexible piezoelectric sensor; the two ends of the multiple cylindrical connecting rods are respectively connected to the two sets of clamps to form a ring clamp; the flexible piezoelectric sensor is arranged on the surface of the multiple cylindrical connecting rods.

9. A cable-stayed bridge damage identification system, characterized in that, include: A memory on which computer programs are stored; A processor for loading and executing the computer program to implement the cable-stayed bridge damage identification method as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the cable-stayed bridge damage identification method as described in any one of claims 1 to 6.

Images