Bridge real-time monitoring method based on double-input self-encoding neural network

Abstract

The present application relates to a kind of bridge real-time monitoring methods based on double input self-encoding neural network, comprising the following steps: ①First, a small amount of dynamic vibration sensor is deployed in arbitrary position of bridge to collect vibration signal under the action of moving load;②Subsequently, signal is analyzed in spectrum, extract bridge structure fundamental frequency f1 as core parameter, based on the dynamic relationship of sampling frequency fs and fundamental frequency f1, according to the formula l≥2fs / f1 The rule is calculated to intercept moving window length;③Windowed signal is input into double-channel self-encoder neural network for training;④The network is used to realize structure state judgment after training, input the response data to be tested, output corresponding feature vector, and characteristic sensitive factor is obtained by calculation, whether the bridge structure has been damaged according to the change trend of characteristic sensitive factor is judged, and the real-time recognition of structure damage under the condition of unsupervised is realized.

Description

Technical Field

[0001] This invention patent relates to the field of bridge structural health monitoring, specifically an artificial intelligence technology that uses dual-input data to train an autoencoder neural network model to achieve real-time bridge monitoring. Background Technology

[0002] Current structural health monitoring methods based on artificial neural networks have promising prospects for practical engineering applications, but they mostly adopt single-input network structures, rely on a large amount of damage label data for supervised training, and require the prior establishment of structural finite element models or health status baseline databases as references.

[0003] Such methods face two major drawbacks in actual bridge monitoring: first, existing bridges often lack complete historical damage data, resulting in insufficient generalization ability of supervised learning models; second, traditional single-input networks are unable to capture the dynamic coupling characteristics of structural response under moving loads, and cannot achieve real-time determination of damage occurrence.

[0004] While recent studies have attempted to introduce unsupervised autoencoders, they still rely on structural model parameters or fixed-window signal extraction methods, leading to increased misjudgment rates under varying load speeds or environmental noise interference. Furthermore, existing neural network methods generally employ full-time data analysis, resulting in computational delays of tens of minutes, which is insufficient to meet the demands of minute-level real-time monitoring. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a real-time bridge monitoring method based on a dual-input architecture autoencoder neural network, solving the problems mentioned in the background section. The specific technical solution is as follows:

[0006] A real-time bridge monitoring method based on a dual-input autoencoder neural network includes the following steps:

[0007] ① First, a small number of dynamic vibration sensors are deployed at any location on the bridge. The dynamic vibration sensors collect vibration signals under the action of moving loads in real time.

[0008] ②Then, the signal is subjected to spectrum analysis, and the fundamental frequency f1 of the bridge structure is extracted as the core parameter. Based on the dynamic relationship between the sampling frequency fs and the fundamental frequency f1, the length of the truncated moving window is calculated according to the criterion l≥2fs / f1 (l represents the length of the truncated input moving window, fs represents the sampling frequency, and f1 represents the fundamental frequency).

[0009] ③ The windowed signal is input into the dual-channel autoencoder neural network for training. The nonlinear feature expression capability is introduced through the activation function to alleviate the gradient vanishing problem. The learning rate is dynamically adjusted by the optimization algorithm and combined with historical gradient information to smooth the parameter update process. During the training phase of the neural network, the neural network attempts to continuously reduce the difference between real data and generated data and attempts to learn the common features of the input data under the constraint of the loss function.

[0010] ④ The trained network is used to determine the structural state. The trained neural network no longer focuses on the generated data but only acts as a feature extractor. It takes the response data to be tested as input, outputs the corresponding feature vector, and calculates the feature sensitivity factor. The trend of the feature sensitivity factor is used to determine whether the bridge structure has been damaged.

[0011] In a preferred embodiment of the present invention, the dynamic vibration sensor in step ① is an acceleration sensor, which captures the vibration response x and y under the action of a moving load.

[0012] As a preferred embodiment of the present invention, in step ③, the neural network training stage, the measured data needs to be normalized and then augmented. The data augmentation method is to randomly extract segments along the time direction from the original time series data. These segments are added as new samples to the training data batch. In each iteration of the neural network, a certain number of data samples are selected to calculate the current gradient of the neural network. This sample set is called the data batch.

[0013] As a preferred embodiment of the present invention, if the data batch size is set to n, then n time series segments are extracted from m sets of original data, and these n data segments constitute the data batch of this iteration.

[0014] As a preferred embodiment of the present invention, in step ③, the dual-channel autoencoder neural network consists of an encoder and a decoder. The encoder is composed of several one-dimensional convolutional layers (Conv1D) and pooling layers. The encoder extracts features from the compressed input response data segments to obtain the corresponding feature vectors. The decoder is composed of deconvolutional layers and upsampling layers. Based on the feature vectors extracted by the encoder, the input response data segments are restored. In order to constrain the feature vectors output by the encoder, the neural network has two input response data segments at the same time. These two input response data segments are taken from two adjacent segments of the temporal response data. The input response data segments are synchronously input into the neural network and share neural network parameters. During the training process, certain constraints are applied to the two feature vectors output by the encoder to shorten the distance between the two feature vectors in the feature space.

[0015] As a preferred embodiment of the present invention, in the bridge real-time monitoring method based on a dual-input autoencoder neural network according to claim 1, in step ④, the windowed signal is input into the dual-channel autoencoder neural network for training, and the gradient vanishing problem is alleviated by introducing nonlinear feature representation capabilities through the SeLU (Scaled Exponential Linear Unit) automated standard activation function.

[0016] Its function expression is:

[0017] Where x is the input data, and λ and α are two constants. In this method, λ is taken as 1.05 and α as 1.67.

[0018] The RMSprop (Root Mean Square Propagation) optimization algorithm dynamically adjusts the learning rate and combines historical gradient information to smooth the parameter update process. The RMSprop algorithm only needs to set the initial learning rate at the beginning of training, and can automatically adjust the learning rate for different parameters during training. The specific algorithm is shown in Equation X:

[0019] v t =pv t-1 +(1-p)g t 2 , w t+1 =w t +Δw t

[0020] Where η is the initial learning rate; v t g is the exponential average value along the gradient direction; t Let p be the time gradient, p be the decay coefficient (taken as 0.9 in this method), and ε be 10. -8 w t For the parameters updated iteratively, t is the time step, t-1 is the previous time step, t+1 is the next time step, and V is the value of V. t-1 Equivalent to the exponential average of the previous step, Δw t It is the updated value of the parameter iteration, w t+1 The next step is the learning rate.

[0021] In a preferred embodiment of the present invention, in step ④, the trained network is used to determine the structural state. The response data to be tested is input, and the corresponding feature vector is output. A sequence is constructed based on the feature vector, assuming Z... i Z i+1 Let be the feature vectors of the i-th and i+1-th windows, respectively. The formula for the feature sensitivity factor is defined as: DI i =COS(Z) i Z i+1 ), (DI i(Representing structural similarity), the trend of changes in characteristic sensitive factors is used to determine whether the structure has been damaged.

[0022] This invention presents a bridge real-time monitoring method based on a dual-input autoencoder neural network. By fusing feature layer constraints of the dual-input autoencoder with a fundamental frequency dynamic window, it achieves real-time identification of structural damage under unsupervised conditions, and has the following advantages:

[0023] (1) Compared with traditional neural network methods, this method does not rely on structural finite element models or historical damage label data, but directly captures the moment of state change by monitoring vibration response signals in engineering projects. The trained network enables rapid judgment of the monitored vibration signals, significantly improving real-time monitoring efficiency.

[0024] (2) In response to the complex load speed changes and environmental noise interference in actual engineering, the structural sensitive features are effectively extracted through fundamental frequency adaptive window segmentation and autoencoder neural network, and high robustness is maintained under signal-to-noise ratio fluctuation conditions.

[0025] (3) This method can be flexibly adapted to different sensor configurations, providing a scalable technical path for long-term health monitoring and rapid diagnosis of existing bridges, and solving the engineering problem of real-time monitoring and identification under the condition of no baseline data. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of the present invention;

[0027] Figure 2 This is a flowchart of the data augmentation process of the present invention;

[0028] Figure 3 This is a diagram of the dual-channel autoencoder neural network architecture of the present invention;

[0029] Figure 4 This is a method identification diagram for the numerical experiment of this invention;

[0030] Figure 5 This invention identifies data from actual monitored bridges after they have been struck by ships. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] like Figure 1As shown, the real-time bridge monitoring method based on a dual-input autoencoder neural network includes the following steps:

[0033] ① First, a small number of dynamic vibration sensors are deployed at any location on the bridge. The dynamic vibration sensors collect vibration signals under the action of moving loads in real time.

[0034] ②Then, the signal is subjected to spectrum analysis, and the fundamental frequency f1 of the bridge structure is extracted as the core parameter. Based on the dynamic relationship between the sampling frequency fs and the fundamental frequency f1, the length of the moving window is calculated according to the criterion l≥2fs / f1 (l represents the length of the moving window for the input, fs represents the sampling frequency, and f1 represents the fundamental frequency) to ensure that the window length matches the vibration characteristics of the bridge.

[0035] ③ The windowed signal is input into the dual-channel autoencoder neural network for training. The nonlinear feature expression capability is introduced through the activation function to alleviate the gradient vanishing problem and improve the network's ability to capture low-frequency damage features. The learning rate is dynamically adjusted by the optimization algorithm and the parameter update process is smoothed by combining historical gradient information, which significantly accelerates the network convergence. During the neural network training stage, the neural network attempts to continuously reduce the difference between real data and generated data and attempts to learn the common features of the input data under the constraint of the loss function.

[0036] ④ The trained network is used to determine the structural state. The trained neural network no longer focuses on the generated data but only acts as a feature extractor. It takes the response data to be tested as input, outputs the corresponding feature vector, and calculates the feature sensitivity factor. The trend of the feature sensitivity factor is used to determine whether the bridge structure has been damaged.

[0037] By using a small number of vibration response sensors (accelerometers as an example) deployed on the bridge structure, structural acceleration data is acquired, capturing the vibration response x and y under moving load. Based on the bridge fundamental frequency f1 obtained from real-time spectrum analysis, the length of the moving window for truncating the dual input data is determined according to the formula l≥2fs / f1, so that the window length accurately matches the vibration period characteristics of the bridge and eliminates the signal truncation error caused by the traditional fixed window length.

[0038] In step ③, the neural network training phase, the measured data needs to be normalized before data augmentation. Data augmentation improves the model performance and generalization ability of the neural network. The data augmentation method involves randomly extracting segments along the time direction from the original time-series data. These segments are added as new samples to the training data batch. In each iteration of the neural network, a certain number of data samples are selected to calculate the current gradient of the neural network; this sample set is called the data batch. Figure 2As shown, if the data batch size is set to n, then n time series segments are extracted from m sets of original data, and these n data segments constitute the data batch of this iteration.

[0039] Specifically, in step ③, the dual-channel autoencoder neural network consists of an encoder and a decoder. The encoder comprises several one-dimensional convolutional layers (Conv1D) and pooling layers. The encoder compresses the input response data segment and extracts features to obtain the corresponding feature vector. The decoder consists of deconvolutional layers and upsampling layers. Based on the feature vector extracted by the encoder, it reconstructs the input response data segment. To constrain the feature vector output by the encoder, this technique proposes a novel network structure, such as... Figure 3 As shown, the neural network has two input response data segments. These two input response data segments are taken from two adjacent segments of the time-series response data. The input response data segments are synchronously input into the neural network and share neural network parameters. During the training process, certain constraints are applied to the two feature vectors output by the encoder to shorten the distance between the two feature vectors in the feature space.

[0040] like Figure 4 and 5 In step ④, the windowed signal is input into the dual-channel autoencoder neural network for training. The SeLU (Scaled Exponential Linear Unit) automated standard activation function is used to introduce nonlinear feature representation capabilities to alleviate the gradient vanishing problem.

[0041] Its function expression is:

[0042] Where x is the input data, and λ and α are two constants. In this method, λ is taken as 1.05 and α as 1.67.

[0043] The RMSprop (Root Mean Square Propagation) optimization algorithm is an adaptive learning rate optimization algorithm that dynamically adjusts the learning rate and combines historical gradient information to smooth the parameter update process. The RMSprop algorithm only needs to set an initial learning rate at the beginning of training, and can automatically adjust the learning rate for different parameters during training. Its specific algorithm is shown in Equation X:

[0044] v t =pv t-1 +(1-p)g t 2 , w t+1 =w t +Δw t

[0045] Where η is the initial learning rate; v tg is the exponential average value along the gradient direction; t Let p be the time gradient, p be the decay coefficient (taken as 0.9 in this method), and ε be 10. -8 w t For the parameters updated iteratively, t is the time step, t-1 is the previous time step, t+1 is the next time step, and V is the value of V. t-1 Equivalent to the exponential average of the previous step, Δw t It is the updated value of the parameter iteration, w t+1 The next step is the learning rate. Compared with the classic stochastic gradient descent (SGD) optimization algorithm, the RMSprop optimization algorithm can avoid the jagged gradient descent, which greatly speeds up the training of neural networks and significantly accelerates network convergence.

[0046] Figure 4 These represent different working conditions from 0 to 8000 seconds, from no damage to various damage conditions, and can identify the time of damage occurrence. The red arrows indicate the time of damage occurrence. In step ④, the trained network is used to determine the structural state. The input is the response data to be tested, and the output is the corresponding feature vector. A sequence is constructed based on the feature vectors, assuming Z... i Z i+1 Let be the feature vectors of the i-th and i+1-th windows, respectively. The formula for the feature sensitivity factor is defined as: DI i =COS(Z) i Z i+1 ), (DI i (Representing structural similarity), the trend of changes in characteristic sensitive factors is used to determine whether the structure has been damaged. Figure 5 The data is from actual monitoring of the bridge from 6 to 10 a.m., and the collision occurred at 8 a.m.

[0047] The above description is a further detailed explanation of the present invention in conjunction with specific preferred embodiments. It should not be considered that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, several simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the protection scope of the present invention.

Claims

1. A bridge real-time monitoring method based on a dual-input autoencoder neural network, characterized in that, Includes the following steps: ① First, a small number of dynamic vibration sensors are deployed at any location on the bridge. The dynamic vibration sensors collect vibration signals under the action of moving loads in real time. ②Then, spectral analysis was performed on the signal to extract the fundamental frequency of the bridge structure. As a core parameter, based on the sampling frequency With base frequency The dynamic relationship, according to the formula The criteria for calculating the length of the moving window are as follows: This represents the length of the moving window used to capture the input. Represents the sampling frequency. Represents the fundamental frequency; ③ A dual-channel autoencoder neural network consists of an encoder and a decoder. The encoder comprises several one-dimensional convolutional and pooling layers. The encoder compresses the input response data segments and extracts features to obtain corresponding feature vectors. The decoder consists of deconvolutional and upsampling layers. Based on the feature vectors extracted by the encoder, it reconstructs the input response data segments. To constrain the feature vectors output by the encoder, the neural network simultaneously has two input response data segments. These two segments are taken from two adjacent segments of the temporal response data. The input response data segments are synchronously input into the neural network, sharing neural network parameters. During training, certain constraints are applied to the two feature vectors output by the encoder to shorten the distance between the two feature vectors in the feature space. A windowed signal is input into a dual-channel autoencoder neural network for training. The SeLU automated standard activation function is used to introduce nonlinear feature representation capabilities to alleviate the gradient vanishing problem. Its function expression is: Where x is the input data, and λ and α are two constants. In this method, λ is taken as 1.05 and α as 1.

67. The RMSprop algorithm dynamically adjusts the learning rate and smooths the parameter update process by incorporating historical gradient information. The RMSprop algorithm only requires setting an initial learning rate at the start of training, and can automatically adjust the learning rate for different parameters during training. The specific algorithm is shown in the following formula: ， ， in The initial learning rate; It is the exponential average value along the gradient direction; Let p be the time gradient, p be the decay coefficient (taken as 0.9 in this method), and ε be 10. -8 , For the parameters updated iteratively, t is the time step, t-1 is the previous time step, t+1 is the next time step, and V is the value of V. t-1 Equivalent to the exponential average of the previous step, Δw t It is the updated value of the parameter iteration, w t+1 The next step is the learning rate. By dynamically adjusting the learning rate through optimization algorithms and combining historical gradient information to smooth the parameter update process, during the neural network training phase, the neural network attempts to continuously reduce the difference between real data and generated data, and tries to learn the common features of the input data under the constraint of the loss function. ④ The trained network is used to determine the structural state. The trained neural network no longer focuses on the generated data but only acts as a feature extractor. It takes the response data to be tested as input, outputs the corresponding feature vector, and calculates the feature sensitivity factor. The trend of the feature sensitivity factor is used to determine whether the bridge structure has been damaged.

2. The bridge real-time monitoring method based on a dual-input autoencoder neural network according to claim 1, characterized in that: In step ①, the dynamic vibration sensor is an acceleration sensor. The acceleration sensor captures the vibration response x and y under the action of moving load. x and y represent acceleration data or acceleration time series from two different sensors.

3. The bridge real-time monitoring method based on a dual-input autoencoder neural network according to claim 1, characterized in that: In step ③, during the neural network training phase, the measured data needs to be normalized before data augmentation. The data augmentation method involves randomly extracting segments along the time direction from the original time series data. These segments are added as new samples to the training data batch. In each iteration of the neural network, a certain number of data samples are selected to calculate the current gradient of the neural network. This sample set is called the data batch.

4. The bridge real-time monitoring method based on a dual-input autoencoder neural network according to claim 3, characterized in that: If the data batch size is set to n, then n time series segments are extracted from m sets of original data, and these n data segments constitute the data batch of this iteration.

5. The bridge real-time monitoring method based on a dual-input autoencoder neural network according to claim 1, characterized in that: In step ④, the trained network is used to determine the structural state. The input is the response data to be tested, and the output is the corresponding feature vector. A sequence is constructed based on these feature vectors. Assuming... , They are the first , The feature vector of each window, and the formula for the feature sensitivity factor are defined as follows: DI i It represents the similarity of structural characteristics, and determines whether the structure has been damaged based on the changing trend of the feature sensitivity factor.

Images