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

Abstract

The invention relates to a real-time bridge monitoring method based on a double-input self-encoding neural network. The method comprises the following steps: (1) deploying a small number of dynamic vibration sensors at any position of a bridge to collect vibration signals under the action of a moving load; (2) performing spectral analysis on the signal, extracting a fundamental frequency f1 of a bridge structure as a core parameter, and calculating an interception moving window length according to a criterion that a formula l is greater than or equal to 2fs / f1 on the basis of a dynamic relationship between a sampling frequency fs and the fundamental frequency f1; (3) inputting the windowed signal into a dual-channel auto-encoder neural network for training; and (4) realizing structural state judgment by utilizing the trained network, inputting response data to be tested, outputting a corresponding feature vector, calculating to obtain a feature sensitive factor, and judging whether the bridge structure is damaged or not according to a change trend of the feature sensitive factor, thereby realizing real-time structural damage identification under an unsupervised condition.

Description

Technical Field

[0001] The patent of this invention relates to the field of bridge structure health monitoring, specifically an artificial intelligence technology that uses dual-input data to train an autoencoding neural network model to achieve real-time monitoring of bridges. Background Art

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

[0003] This type of method faces two fatal flaws 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 have difficulty capturing the dynamic coupling characteristics of structural responses under moving loads, making it impossible to achieve real-time judgment of damage occurrence.

[0004] While recent research has attempted to introduce unsupervised autoencoders, these still rely on structural model parameters or fixed-window signal capture, leading to increased misjudgment rates when subjected to varying load velocity or environmental noise. Furthermore, existing neural network methods generally analyze full-time data, resulting in computational delays of tens of minutes, making them difficult to meet the demands of minute-level real-time monitoring. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a real-time bridge monitoring method based on an autoencoder neural network with a dual-input architecture to solve the problems raised in the above background technology. The specific technical solution is as follows:

[0006] The 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 random locations on the bridge. The dynamic vibration sensors collect vibration signals under the action of moving loads in real time.

[0008] ②The signal is then spectrally analyzed to extract the bridge structure's fundamental frequency f1 as the core parameter. Based on the dynamic relationship between the sampling frequency fs and the fundamental frequency f1, the interception moving window length is calculated according to the formula l ≥ 2fs / f1 (l represents the interception input moving window length, fs represents the sampling frequency, and f1 represents the fundamental frequency).

[0009] ③ The windowed signal is input into a dual-channel autoencoder neural network for training. The activation function introduces nonlinear feature expression capabilities to alleviate the gradient vanishing problem. The optimization algorithm dynamically adjusts the learning rate and combines 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 attempts to learn the common characteristics of the input data under the constraints of the loss function.

[0010] ④ Use the trained network to realize structural status judgment. The trained neural network no longer focuses on the situation of generating data but only acts as a feature extractor. It inputs the response data to be tested, outputs the corresponding feature vector, and calculates the feature sensitivity factor. According to the changing trend of the feature sensitivity factor, it is judged whether the bridge structure has been damaged.

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

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

[0013] As a preferred solution of the present invention, the data batch size is set to n, and n time series segments are intercepted from m groups of original data, and these n data segments constitute the data batch of this iteration.

[0014] As a preferred solution of the present invention, in step ③, the dual-channel autoencoder neural network is composed of an encoder and a decoder. The encoder is composed of several one-dimensional convolution layers (Conv1D) and pooling layers (Pooling). The encoder compresses the input response data segment and performs feature extraction to obtain the corresponding feature vector. The decoder is composed of a deconvolution layer and an upsampling layer. According to the feature vector extracted by the encoder, the input response data segment is restored. In order to constrain the feature vector 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 on the time series response data. The input response data segments are synchronously input into the neural network to share the neural network parameters. During the training process, certain constraints are imposed on 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 solution of the present invention, according to the real-time bridge monitoring method based on the dual-input autoencoder neural network described in claim 1, in step ④, the windowed signal is input into the dual-channel autoencoder neural network for training, and the nonlinear feature expression capability is introduced by the SeLU (Scaled Exponential Linear Unit) automated standard activation function to alleviate the gradient vanishing problem.

[0016] Its function expression is:

[0017] Where x is the input data, λ and α are two constants. In this method, λ is taken as 1.05 and α is taken 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 rates of different parameters during training. The specific algorithm is shown in Formula 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 is the exponential average of the gradient direction; g t is the time gradient, p is the attenuation coefficient, which is 0.9 in this method, and ε is 10 -8 , w t is the parameter for iterative update, t is the time step, t-1 is the previous step, t+1 is the next time step, V t-1 Equivalent to the exponential average of the previous step, Δw t is the updated value of parameter iteration, w t+1 is the learning rate for the next step.

[0021] As a preferred solution of the present invention, in step ④, the trained network is used to realize structural state judgment, the response data to be tested is input, the corresponding feature vector is output, and a sequence is formed according to the feature vector. Assume that Z i , Z i+1 They are the feature vectors of the i-th and i+1-th windows respectively. The formula of the feature sensitivity factor is defined as: DI i =COS(Z i ,Z i+1 ), (DI iRepresents the similarity of structural characteristics), and judge whether the structure is damaged based on the changing trend of the characteristic sensitivity factor.

[0022] The real-time bridge monitoring method based on the dual-input autoencoder neural network of the present invention achieves real-time identification of structural damage under unsupervised conditions by integrating the feature layer constraints of the dual-input autoencoder with the fundamental frequency dynamic window. It has the following advantages:

[0023] ⑴ Compared with traditional neural network methods, this method does not rely on structural finite element models or historical damage label data, and can automatically capture the moment of state mutation by directly monitoring the vibration response signal in the project. The trained network can quickly judge the monitored vibration signal, significantly improving the efficiency of real-time monitoring;

[0024] (2) Aiming at the complex load velocity changes and environmental noise interference in actual engineering, the system effectively extracts structural sensitive features through baseband adaptive window segmentation and autoencoder neural network, and maintains high robustness under conditions of signal-to-noise ratio fluctuations;

[0025] ⑶This 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 in the absence of baseline data. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the overall flow chart of the present invention;

[0027] Figure 2 is a data augmentation flow chart 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 It is a method identification diagram of the numerical experiment of the present invention;

[0030] Figure 5 It is a data graph of the present invention identifying the actual monitoring bridge after being hit by a ship; DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

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

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

[0034] ② The signal is then spectrally analyzed to extract the bridge structure's fundamental frequency f1 as a core parameter. Based on the dynamic relationship between the sampling frequency fs and the fundamental frequency f1, the interception moving window length is calculated according to the formula l ≥ 2fs / f1 (l represents the interception input moving window length, fs represents the sampling frequency, and f1 represents the fundamental frequency), ensuring that the window length matches the bridge's vibration characteristics.

[0035] ③ The windowed signal is input into a dual-channel autoencoder neural network for training. The activation function introduces nonlinear feature expression capabilities to alleviate the gradient vanishing problem and improve the network's ability to capture low-frequency damage features. The optimization algorithm dynamically adjusts the learning rate and combines historical gradient information to smooth the parameter update process, significantly accelerating network convergence. During the neural network training phase, the neural network attempts to continuously reduce the difference between real data and generated data, and attempts to learn the common characteristics of the input data under the constraints of the loss function.

[0036] ④ Use the trained network to realize structural status judgment. The trained neural network no longer focuses on the situation of generating data but only acts as a feature extractor. It inputs the response data to be tested, outputs the corresponding feature vector, and calculates the feature sensitivity factor. According to the changing trend of the feature sensitivity factor, it is judged whether the bridge structure has been damaged.

[0037] By arranging a small number of vibration response sensors (for example, accelerometers) on the bridge structure, the structural acceleration data is acquired to capture the vibration responses x and y under the action of moving loads. Based on the bridge fundamental frequency f1 obtained by real-time spectrum analysis, the length of the moving window for intercepting the dual input data is determined according to the formula l≥2fs / f1. This ensures 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] During the neural network training phase in step ③, the measured data needs to be normalized and then augmented. Data augmentation improves the model performance and generalization ability of the neural network. The data augmentation method is to randomly intercept fragments along the time direction in the original time series data. These fragments are added to the training data batch as new samples. In each iteration of the neural network, a certain number of data samples will be selected to calculate the current gradient of the neural network. This sample set is called a data batch. Figure 2As shown in the figure, if the data batch size is set to n, then n time series segments are intercepted from m groups 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 consists of several one-dimensional convolution 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 a deconvolution layer and an upsampling layer. According to the feature vector extracted by the encoder, the input response data segment is restored. In order to constrain the feature vector output by the encoder, a new network structure is proposed in this technology, such as Figure 3 As shown in the figure, 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 time series response data. The input response data segments are synchronously input into the neural network to share the neural network parameters. During the training process, certain constraints are imposed on 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, and the nonlinear feature expression capability is introduced through the SeLU (Scaled Exponential Linear Unit) automated standard activation function to alleviate the gradient vanishing problem.

[0041] Its function expression is:

[0042] Where x is the input data, λ and α are two constants. In this method, λ is taken as 1.05 and α is taken 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 the initial learning rate at the beginning of training, and can automatically adjust the learning rates of different parameters during training. The specific algorithm is shown in Formula 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 tis the exponential average of the gradient direction; g t is the time gradient, p is the attenuation coefficient, which is 0.9 in this method, and ε is 10 -8 , w t is the parameter for iterative update, t is the time step, t-1 is the previous time step, t+1 is the next time step, V t-1 Equivalent to the exponential average of the previous step, Δw t is the updated value of parameter iteration, w t+1 It is the learning rate for the next step. Compared with the classic stochastic gradient descent (SGD) optimization algorithm, the RMSprop optimization algorithm can avoid the sawtooth descent of the gradient, greatly speeding up the training of the neural network and significantly accelerating the network convergence.

[0046] Figure 4 The time from 0 to 8000s is different. From non-destructive to different damage conditions, the moment of damage occurrence can be identified. The red arrow indicates the moment of damage occurrence. In step ④, the trained network is used to judge the structural state. The response data to be tested is input and the corresponding feature vector is output. A sequence is formed based on the feature vector. Assume that Z i , Z i+1 They are the feature vectors of the i-th and i+1-th windows respectively. The formula of the feature sensitivity factor is defined as: DI i =COS(Z i ,Z i+1 ), (DI i Represents the similarity of structural characteristics), and judge whether the structure is damaged based on the changing trend of the characteristic sensitivity factor. Figure 5 This is the actual monitoring data of the bridge from 6 to 10 in the morning, and the scratch occurred at 8 o'clock.

[0047] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or replacements can be made without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A real-time bridge monitoring method based on a dual-input autoencoder neural network, characterized in that: The following steps are involved: ① First, a small number of dynamic vibration sensors are deployed at random locations on the bridge. The dynamic vibration sensors collect vibration signals under the action of moving loads in real time. ②The signal is then spectrally analyzed to extract the bridge structure's fundamental frequency f1 as the core parameter. Based on the dynamic relationship between the sampling frequency fs and the fundamental frequency f1, the interception moving window length is calculated according to the formula l ≥ 2fs / f1 (l represents the interception input moving window length, fs represents the sampling frequency, and f1 represents the fundamental frequency). ③ The windowed signal is input into a dual-channel autoencoder neural network for training. The activation function introduces nonlinear feature expression capabilities to alleviate the gradient vanishing problem. The optimization algorithm dynamically adjusts the learning rate and combines 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 attempts to learn the common characteristics of the input data under the constraints of the loss function. ④ Use the trained network to realize structural status judgment. The trained neural network no longer focuses on the situation of generating data but only acts as a feature extractor. It inputs the response data to be tested, outputs the corresponding feature vector, and calculates the feature sensitivity factor. According to the changing trend of the feature sensitivity factor, it is judged whether the bridge structure has been damaged.

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

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

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

5. The real-time bridge monitoring method based on a dual-input autoencoder neural network according to claim 1 is characterized in that: In step ③, the dual-channel autoencoder neural network consists of an encoder and a decoder. The encoder consists of several one-dimensional convolutional layers (Conv1D) and pooling layers. The encoder compresses the input response data segment to extract features and obtains the corresponding feature vector. The decoder consists of a deconvolution layer and an upsampling layer. According to the feature vector extracted by the encoder, the input response data segment is restored. In order to constrain the feature vector 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 time series response data. The input response data segments are synchronously input into the neural network to share the neural network parameters. During the training process, certain constraints are imposed on the two feature vectors output by the encoder to shorten the distance between the two feature vectors in the feature space.

6. The real-time bridge monitoring method based on a dual-input autoencoder neural network according to claim 1 is characterized in that: 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 expression capabilities to alleviate the gradient vanishing problem. Its function expression is: Where x is the input data, λ and α are two constants. In this method, λ is taken as 1.05 and α is taken as 1.

67. 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 rates of different parameters during training. The specific algorithm is shown in Formula X: v t =pv t-1 +(1-p)g t 2 , w t+1 =w t +Δw t Where η is the initial learning rate; v t is the exponential average of the gradient direction; g t is the time gradient, p is the attenuation coefficient, which is 0.9 in this method, and ε is 10 -8 , w t is the parameter for iterative update, t is the time step, t-1 is the previous time step, t+1 is the next time step, V t-1 Equivalent to the exponential average of the previous step, Δw t is the updated value of parameter iteration, w t+1 is the learning rate for the next step.

7. The real-time bridge monitoring method based on a dual-input autoencoder neural network according to claim 1 or 6, characterized in that: In step ④, the trained network is used to judge the structural state, input the response data to be tested, output the corresponding feature vector, and form a sequence based on the feature vector. Assume that Z i , Z i+1 They are the feature vectors of the i-th and i+1-th windows respectively. The formula of the feature sensitivity factor is defined as: DI i =COS(Z i ,Z i+1 ), (DI i Represents the similarity of structural characteristics), and judge whether the structure is damaged based on the changing trend of the characteristic sensitivity factor.

Images