Concrete dam damage identification method based on multi-modal parameter deep learning

Abstract

The invention discloses a concrete dam damage identification method based on multi-modal parameter deep learning, and the method comprises the steps: extracting the actual measurement modal parameters of a concrete dam through distributed optical fiber vibration sensing data, building a high-precision dam body dynamic model based on finite element simulation, randomly generating different material parameter combinations, obtaining the corresponding modal parameters, and carrying out the recognition of the damage of the concrete dam. And the simulation model is optimized and corrected, so that the simulation model better conforms to the actual dynamic characteristics of the dam body. In the corrected simulation model, randomly generating different damage positions, ranges and degrees, and constructing a training sample set covering multiple damage modes; and establishing a one-dimensional convolutional neural network, performing deep learning model training by using the training sample, and substituting the actually measured distributed vibration data to realize intelligent identification of the damage position, range and degree of the concrete dam. According to the method, the accuracy and adaptability of concrete dam damage identification are improved, the method can be widely applied to long-term health monitoring of hydraulic structures, and efficient and accurate technical support is provided for dam safety assessment.

Description

Technical Field

[0001] The present invention relates to the technical field of safe operation and maintenance of water conservancy projects, and in particular to a concrete dam damage identification method based on multi-modal parameter deep learning. Background Art

[0002] During long-term service, concrete dams may suffer from cracks, stiffness degradation and other damages due to environmental changes, loads and material aging, which may affect structural safety. Modal analysis based on vibration measurement is widely used in dam health monitoring. Its basic principle is to identify damage by using changes in modal parameters (such as modal frequency, damping ratio and vibration mode) of the structure.

[0003] However, the traditional modal analysis method relies on a limited number of point sensors. Due to the density of the measurement points, it is difficult to fully capture the large-scale vibration characteristics of the dam. In addition, the amplitude of the modal frequency change is usually within 10 -1 The level is greatly affected by environmental noise, resulting in insufficient stability in damage identification. In contrast, local changes in modal vibration shapes are more sensitive to damage, but traditional sensing technology is difficult to achieve high spatial resolution data acquisition, limiting its application in health monitoring of complex hydraulic structures.

[0004] In recent years, deep learning technology has shown great potential in the field of structural damage identification. Compared with traditional methods, deep learning can automatically learn complex data features, adapt to nonlinear damage patterns, and improve recognition accuracy and generalization ability. In concrete dam damage identification, neural networks can learn the mapping relationship between damage patterns and modal changes by training a large amount of finite element simulation data and measured modal parameters, so as to achieve more accurate damage location and degree assessment. In particular, convolutional neural networks (CNNs), due to their powerful feature extraction capabilities, can automatically extract deep structural information from vibration data and improve recognition accuracy. In addition, finite element analysis (FEA) can be used to simulate modal parameters under different damage conditions, providing a large number of training samples for deep learning. However, due to the uncertainty of concrete dam material parameters, there are differences between simulation models and actual measurements. Therefore, it is necessary to optimize and correct the simulation model in combination with measured data to improve the credibility and adaptability of deep learning models. In addition, the performance of deep learning is highly dependent on the quality of the data set. How to reasonably generate training data covering different damage patterns and ensure that it covers various possible damage conditions is a key challenge in current research. Summary of the invention

[0005] Purpose of the invention: In view of the problems pointed out in the background technology, the present invention proposes a concrete dam damage identification method based on deep learning of multi-modal parameters. By establishing a high-precision finite element dynamic model, different material parameter combinations are randomly generated, and the simulation results are optimized to match the measured modal parameters to ensure the authenticity and reliability of the model; a simulation data set with damage is constructed to cover different damage locations, ranges and degrees, and a convolutional neural network (CNN) is used to automatically learn the mapping relationship between modal parameter changes and damage characteristics, so as to achieve efficient and intelligent damage identification.

[0006] Technical solution: The present invention discloses a concrete dam damage identification method based on multimodal parameter deep learning, comprising the following steps: Step 1: Collect the distributed vibration measurement values ​​of the concrete dam, denoise the distributed vibration measurement values, identify the operating modes of the concrete dam based on the distributed vibration measurement values, and extract the measured multi-modal parameters of the concrete dam; Step 2: Establish a finite element model based on the concrete dam engineering data, randomly generate the overall finite element parameter combination of the model, and obtain the corresponding modal parameter indicators through dynamic simulation calculation; Step 3: Combining the modal parameter indicators simulated in step 2 with those measured in step 1, the overall parameter combination of the model is optimized to obtain a revised concrete dam simulation model; Step 4: Use the modified concrete dam simulation model to randomly generate the concrete dam damage degree, damage location and damage range, perform dynamic simulation calculations, and obtain a training sample set of the damaged concrete dam simulation model; Step 5: Establish the initial neural network model for concrete dam damage identification; Step 6: Based on the training sample set of the damaged concrete dam simulation model obtained in step 4, a neural network model is trained, and the measured distributed vibration sensor data is substituted to obtain the recognition results of multiple characteristic parameters of concrete dam damage.

[0007] Furthermore, the distributed vibration measurement denoising in step 1 is specifically as follows: Based on the spatiotemporal two-dimensional distribution characteristics of distributed vibration measurements, they are mapped to images, and data augmentation, logarithmic image processing, and histogram equalization methods are used to improve image quality. Image opening and closing operations and edge detection are used to extract typical image features. The noise-assisted two-dimensional EMD method is used to jointly denoise the distributed vibration measurements.

[0008] Furthermore, the concrete dam operation mode identification in step 1 is specifically as follows: Step 1.1: Extract the non-stationary characteristics of the amplitude and power spectrum of the structure's spatiotemporal distribution of vibration measurements, and select the response sequence set and reference point set representing the significant vibration behavior of the structure based on the non-stationary index; Step 1.2: Use the selected response sequence set and reference point set to construct a transmissibility function matrix, which describes the ratio of the structural response point to the reference response point within a range of 2 times the second-order frequency of the concrete dam, is organized into columns by measurement points, and is organized into rows by frequency to form a matrix, covering the frequency range and having a frequency resolution of less than or equal to 0.05 Hz; Step 1.3: Perform variation and singular value decomposition on the transmissibility function matrix of different response points to solve the dynamic model of the transmissibility-based operational modal analysis method. In the singular value decomposition process, the singular values ​​are sorted from large to small, and the first two singular values ​​and their corresponding singular vectors are retained, and the remaining singular values ​​are eliminated; Step 1.4: Extract key modal parameters based on singular value decomposition, analyze the singular values ​​and corresponding singular vectors in the singular value decomposition results, and identify the modal frequency, damping ratio and vibration shape corresponding to each singular value.

[0009] Furthermore, in step 2, a finite element parameter combination of the entire concrete dam is randomly generated, and the finite element parameter combination includes elastic modulus, density, and Poisson's ratio.

[0010] Furthermore, the optimization calculation of the overall parameter combination of the model in step 3 is as follows: The optimized objective function is as follows: ; in, x is the parameter combination to be corrected. Since no preset mass is required, the parameter combination includes the concrete dam density and material dynamic parameters. The parameter combination is within the range D x Inside, D x The estimated value of the material parameters fluctuates by 20% to 50%; T ( x )and F ( x ) are the objective functions of time domain index and frequency domain index respectively; f i is the measured seismic response sequence at the ith discrete time point; g j is the modal frequency identified at the order of the jth modal frequency; and are the seismic response sequence simulated at the ith discrete time point and the model natural frequency simulated at the jth modal frequency order; n 1 and n 2 are the number of discrete time points and the number of modal frequency orders respectively; The optimization process is: Step 3.1: Analyze the quality of the multimodal parameters obtained in step 1, eliminate false modes, and select modal parameters that meet the convergence requirements; Step 3.2: Estimate the parameter range of the finite element model and obtain the corresponding multi-modal parameters under each parameter combination of the concrete dam in step 2; Step 3.3: Calculate the fitness value based on optimizing the objective function in the time-frequency domain and construct an initial population; Step 3.4: Use the MOEA / D optimization algorithm to optimize, decompose the multi-objective problem into multiple weighted single-objective sub-problems, and perform crossover and mutation operations under neighborhood search to generate new individuals and update the Pareto frontier solution set; Step 3.5: Determine whether the termination condition is met. If not, continue iterating; otherwise, output the Pareto optimal parameter combination set.

[0011] Furthermore, in step 4, the damage degree, damage location and damage range of the concrete dam are randomly generated, and the specific implementation method is as follows: The damage position includes the horizontal and vertical coordinates of the damage center: first, the two-dimensional model of the dam body is divided into the smallest unit of triangle or trapezoid, and the probability that the damage center is located in the unit is determined by the ratio of the unit area to the total area; for the trapezoidal unit, two random numbers in the range of 0 to 1 are used to express the positions of two points randomly selected on the parallel sides, and then a new random number is used to express the random position on the line connecting the two points as the final damage center; The damage degree uses the maximum reduction ratio of the elastic modulus as a setting parameter, and randomly generates any value between the maximum reduction ratio and 1 to express the damage degree; The damage range uses the maximum damage width and the maximum damage height as setting parameters, randomly generates a value between 0 and the maximum setting parameters to define the damage range, and then reduces the unit elastic modulus of the generated finite element model to determine whether the centroid of each unit is within the damage range defined above. If so, the unit elastic modulus is reduced; For a two-dimensional finite element model, the damage characteristic parameter vector is as follows: ; in, They represent the horizontal and vertical coordinates of the damage center, the damage degree, and the width and height of the damage range respectively.

[0012] Furthermore, the initial neural network model for concrete dam damage identification in step 5 is a one-dimensional CNN architecture with multiple inputs and multiple outputs, including an input layer, an output layer, a convolution layer, a pooling layer, a hidden layer, and a fully connected layer. The input of the initial neural network model for concrete dam damage identification includes modal vibration shapes and modal frequencies. The modal vibration shapes include first-order vibration shapes and second-order vibration shapes. The modal frequencies include first-order frequencies and second-order frequencies. The input quantity group of the initial neural network model for concrete dam damage identification is [first-order vibration shape, second-order vibration shape, first-order frequency, second-order frequency], and the output quantity group is [damage proportion, damage location, damage range].

[0013] Furthermore, in the neural network model training in step 6, the convolution process calculation formula is as follows: ; Among them, bias represents the bias term, which is a trainable parameter of the output channel, input represents input, out represents output, and the size of the neural network model training input vector is ( N i , C in , L in ), the size of the neural network model training output vector is ( N i , C out , L out ), N i Indicates i Group training samples; C in represents the input vector channel, C out Represents the output vector channel; L in and L out Respectively represent the channel vector lengths of input and output, k Represents the current and convolution kernel w The dot product k input vector channels, Indicates j output vector channels; Represents the convolution operation, input( N i , k ) indicates the i The first k The input vector of channels, out( N i , ) indicates the iThe first j The output vector obtained by channels; Apply the maximum pooling method, and the calculation formula is as follows: ; in, s is the step length, p Represents the vector p Segment pooling area, m Indicates 0~ p -1, the pooling area number, C j Indicates j channels that perform pooling operations, out( N i , C j , p ) indicates the i Group training sample p Segment pooling area j The pooled output value obtained by the pooling channel, input( N i , C j , + m ) indicates the i Group training sample p Segment pooling area j The input vector of pooling channels.

[0014] Beneficial effects: 1. The present invention provides a method for concrete dam damage identification based on multi-modal parameter deep learning. The method adopts a deep learning method combined with finite element simulation and distributed optical fiber vibration data, and automatically extracts modal parameter features through a convolutional neural network (CNN). This method can achieve high-precision identification of the location, scope and degree of concrete dam damage, reduce manual intervention, and improve the intelligence level of identification.

[0015] 2. The present invention improves the authenticity of damage simulation by optimizing the finite element dynamics model and correcting the simulation parameters to match the measured modal parameters. In addition, the neural network can adapt to different types and degrees of damage, improve the adaptability to complex structures, and reduce the impact of environmental noise on the recognition results.

[0016] 3. The present invention adopts a randomized damage generation method to generate data sets of different damage conditions on the modified finite element model, so that the deep learning model can cover various possible damage modes, improve the generalization ability of the model, and enable it to identify unknown damage conditions.

[0017] 4. The present invention can be used for long-term monitoring of hydraulic structures such as dams, bridges, and tunnels, providing more accurate damage assessment results, providing a scientific basis for structural maintenance and safety management, and helping to improve the safety and economy of hydraulic structure operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Implement the process of multi-modal parameter deep learning method for concrete dam damage identification; Figure 2 This is a schematic diagram of typical measurement values ​​of distributed optical fiber; Figure 3 The schematic diagram of the finite element simulation model of a concrete dam with damage; Figure 4 Schematic diagram of model damage parameters and their correction method; Figure 5 This is a schematic diagram of the CNN architecture for dynamic identification of concrete dam damage; Figure 6 This is a diagram showing the convergence process of the CNN training effect evaluation indicators for dynamic identification of concrete dam damage. DETAILED DESCRIPTION

[0019] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0020] The present invention provides a concrete dam damage identification method based on multi-modal parameter deep learning, which realizes intelligent identification of the location, scope and degree of concrete dam damage by integrating distributed vibration sensor data, finite element simulation and deep learning technology. The method extracts multi-modal parameters based on measured vibration data, establishes a finite element model and optimizes parameters, and then constructs a diverse training sample set by randomly generating damage conditions. A one-dimensional convolutional neural network (CNN) is used to learn the mapping relationship between modal parameter changes and damage characteristics, and finally achieves efficient and accurate damage identification. This technology can overcome the shortcomings of limited measurement points and low modal parameter sensitivity in traditional methods, significantly improve the accuracy and reliability of concrete dam health monitoring, and provide technical support for the safety management of hydraulic structures.

[0021] See also Figure 1 , Figure 1 A schematic diagram of an implementation flow of a concrete dam damage identification method based on multi-modal parameter deep learning provided by an embodiment of the present invention, with reference to Figure 1 As shown, the technical implementation process includes the following steps: Step 1: De-noise the distributed vibration measurement matrix obtained (the distributed measurement matrix can provide the distribution information of the concrete dam vibration mode. When the measurement point spacing is less than 10m, the distributed optical fiber vibration sensing technology can be used. When the measurement point spacing is greater than 10m, the seismic array array can be dispersed to collect the vibration response data matrix. The measurement points should be arranged in the area with larger vibration response amplitude). Figure 2 To illustrate the spatiotemporal distribution of the measured distributed vibration values, five measuring points in a certain measurement range are selected from all the measuring points to make the diagram clear. The concrete dam structure operation modal identification based on the distributed vibration values ​​is performed, and the measured multi-modal parameters of the concrete dam are extracted. The measured multi-modal parameters include modal frequency, modal vibration shape, and modal amplitude.

[0022] The denoising of the distributed vibration measurement value matrix is ​​performed by mapping it to the image based on its spatiotemporal two-dimensional distribution characteristics. Data augmentation, logarithmic image processing, and histogram equalization methods are used to improve the image quality. Image opening, closing, and edge detection operations are used to extract typical image features. Then, the noise-assisted two-dimensional EMD method is used to jointly denoise the distributed vibration measurement value matrix.

[0023] The concrete dam structure operation mode identification based on distributed vibration measurement is as follows: Step 1.1: Extract the non-stationary characteristics of the amplitude and power spectrum of the structure's spatiotemporal distribution of vibration measurements, and select the response sequence set and reference point set representing the significant vibration behavior of the structure based on the non-stationary index.

[0024] Step 1.2: Use the selected response sequence and reference points to construct a transmissibility function matrix, which describes the ratio of the structural response point to the reference response point within a range of 2 times the second-order frequency before the concrete dam. Organize it into columns by measuring points and into rows by frequency to form a matrix. It must cover the above frequency range and have a frequency resolution of less than or equal to 0.05 Hz.

[0025] Step 1.3: Perform variation and singular value decomposition on the transmissibility function matrix at different response points to solve the dynamic model of the transmissibility-based operational modal analysis method as follows: ; in, M , N represents different excitation load conditions, i , j Indicates different response points, represents the transmissibility function, express r The eigenvalue of order, represents strain ( ) vibration mode; therefore for MUnder load conditions i Point and j The transfer rate function between points, for N Under load conditions i Point and j The transfer rate function between points, is the difference between the transmissibility functions under the above two loading conditions, for i Point r The value of the first-order strain mode.

[0026] During the singular value decomposition process, the singular values ​​are sorted from large to small, and the first two singular values ​​and their corresponding singular vectors are retained, which usually represent the most significant dynamic characteristics of the structure. The remaining singular values ​​are eliminated to help remove noise effects and unimportant dynamic components, thereby streamlining the model and improving the efficiency and accuracy of subsequent analysis.

[0027] Step 1.4: Extract key modal parameters based on singular value decomposition, analyze the singular values ​​and corresponding singular vectors in the singular value decomposition results, and identify the modal frequency, damping ratio and vibration shape corresponding to each singular value.

[0028] Step 2: Establish the finite element model of the concrete dam based on the actual engineering data, such as Figure 3 As shown in the model without damage parameters, the overall finite element parameter combination of the model is randomly generated, and the corresponding modal parameter indicators are obtained through dynamic simulation calculation.

[0029] Randomly generate the overall finite element parameter combination of the concrete dam. In this implementation, the elastic modulus, density, and Poisson's ratio are selected as parameters. The parameter preselection range is shown in Table 1 below, where: α and β is the Rayleigh damping coefficient, E Represents the elastic modulus, μ Denote Poisson's ratio: Table 1 Parameter preselection range

[0030] Step 3: Combining the simulated and measured modal parameter indicators, the overall parameter combination of the model is optimized to obtain the revised concrete dam simulation model.

[0031] The optimization calculation of the overall parameter combination of the model, the objective function of the optimization algorithm is as follows:

[0032] in, x is the parameter combination to be corrected. Since no preset mass is required, the parameter combination includes the concrete dam density and material dynamic parameters. The parameter combination is within the range Dx Inside, and D x The estimated value of the material parameters fluctuates by 20% to 50%; T ( x )and F ( x ) are the objective functions of time domain index and frequency domain index respectively; f i is the measured seismic response sequence at the ith discrete time point; g j is the modal frequency identified at the order of the jth modal frequency; and are the seismic response sequence simulated at the ith discrete time point and the model natural frequency simulated at the jth modal frequency order; n 1 and n 2 are the number of discrete time points and the number of modal frequency orders.

[0033] The optimization process is: Step 3.1: Analyze the quality of the multimodal parameters obtained in step 1, eliminate false modes, and select modal parameters that meet the convergence requirements.

[0034] In this embodiment, the modal frequency of the concrete dam is assisted in identifying by comparing the frequency spectrum density curve and the stability diagram of the free field vibration sequence.

[0035] When the preset orders are different, a suspected stable axis may be formed in the stability diagram; but when the same position in the frequency spectrum density curve is not prominent, the frequency is hardly identified, and it can be determined that the frequency value is a false modal frequency and can be eliminated.

[0036] If multiple modal frequencies are identified but fail to form a stable axis, this is also affected by the dominant frequency of earthquake excitation and can be eliminated.

[0037] They all form stable axes and there is no dominant frequency at the corresponding position of the seismic excitation frequency spectrum, which can be determined as the position of the natural frequency of the concrete dam.

[0038] As the frequency increases, the stability axis on the stability diagram becomes less obvious and the identified modes become more dense. In order to reduce the error, it is reasonable to use the first two modal frequency identification values ​​as frequency domain indicators to construct the objective function of model parameter correction.

[0039] Step 3.2: Estimate the parameter range of the finite element model and obtain the corresponding multi-modal parameters under each parameter combination of the concrete dam in step 2.

[0040] Step 3.3: Calculate the fitness value based on the time-frequency domain objective function and construct the initial population.

[0041] Step 3.4: Use MOEA / D for optimization to decompose the multi-objective problem into multiple weighted single-objective sub-problems, and perform crossover and mutation operations under neighborhood search to generate new individuals and update the Pareto frontier solution set.

[0042] Step 3.5: Determine whether the termination condition is met. If not, continue iterating; otherwise, output the Pareto optimal parameter combination set.

[0043] Step 4: Use the modified concrete dam simulation model to randomly generate damage degree, damage location and damage range, perform dynamic simulation calculations, and obtain a training sample set of the damaged concrete dam simulation model.

[0044] The generation of the damage location includes the horizontal and vertical coordinates of the damage center. First, the two-dimensional model of the dam body is divided into the smallest unit of triangle or trapezoid, and the probability that the damage center is located in the unit is determined by the ratio of the unit area to the total area. For the trapezoidal unit, two random numbers in the range of 0 to 1 are used to express the positions of two points randomly selected on the parallel sides, and then a new random number is used to express the random position on the line connecting the two points as the final damage center.

[0045] The damage degree uses the maximum reduction ratio of the elastic modulus as a setting parameter, and randomly generates any value between the maximum reduction ratio and 1 to express the damage degree.

[0046] The damage range uses the maximum damage width and the maximum damage height as setting parameters, and randomly generates values ​​between 0 and the maximum setting parameters to define the damage range. Then, the elastic modulus of the unit of the generated finite element model is reduced to determine whether the centroid of each unit is within the damage range defined above. If so, the elastic modulus of the unit is reduced.

[0047] For a two-dimensional finite element model, the damage characteristic parameter vector is as follows:

[0048] Among them, the five components represent the horizontal and vertical coordinates of the damage center, the damage degree, and the width and height of the damage range.

[0049] In this embodiment, the gravity dam simulation model damage simulation uses the rectangular frame unit, and the simulation measurement optical fiber is laid on the downstream side of the dam section. The center of the rectangle represents the damage position (the horizontal coordinate and vertical coordinate origin of the damage center point are the lower left corner of the gravity dam section, and the vertical coordinate changes along the vertical direction), and the damage range is represented by the length and width of the rectangular frame. The randomly generated damage parameters can be reflected in the simulation model ( Figure 3 ), where the red rectangle represents the preset damage range. However, in order to solve the problem that the output parameters may exceed the boundary of the simulation model, the above damage parameters are further modified ( Figure 4 ),like Figure 4The initially generated damage range (left side) exceeds the concrete dam model unit. At this time, the excess length brings the problem of multiple damage range parameters corresponding to the same damage situation to the damage identification model, which is easy to cause the damage identification to not converge. Therefore, the initially generated damage range needs to be reduced until it just includes the boundary of the concrete dam finite element model. At this time, the corrected damage range parameters (i.e., the length and width of the rectangle) are used as training values ​​for the model. Figure 4 Where (x1, y1) is the center coordinate of the initially generated damage range rectangle, and (x2, y1) is the center coordinate of the corrected damage range rectangle.

[0050] First, assume that the position of the damage center remains unchanged. According to the width of the damage range, calculate the left and right horizontal coordinates of the damage range to determine whether it exceeds the horizontal coordinate value on the left or right side of the model. If it exceeds, replace it with the horizontal coordinate value of the boundary. For the upper and lower vertical coordinates of the damage range, use the upper and lower vertical coordinate values ​​of the model boundary to make a judgment. Similarly, replace the vertical coordinate values ​​that exceed the model boundary. After the damage range is updated, update the position of the damage center. If the horizontal and vertical coordinates of the boundary are updated, the horizontal and vertical coordinates of the new damage center are calculated with the midpoint of the updated horizontal and vertical coordinates of the boundary. In addition, the degree of damage is included. A total of 1,000 damage combinations are generated, of which 800 groups are used as training sets and 200 groups are used as validation sets. The details are shown in Table 2 below: Table 2 Damage combination dataset

[0051] Step 5: Establish the initial neural network model for concrete dam damage identification.

[0052] The initial neural network model for concrete dam damage identification is a one-dimensional CNN architecture with multiple inputs and multiple outputs (see Figure 5 ), including input layer, output layer, convolution layer, pooling layer, hidden layer, and fully connected layer. In this problem, the input of the initial neural network model for concrete dam damage identification is the dynamic index of the concrete dam, including modal vibration shape and modal frequency. The modal vibration shape includes first-order vibration shape and second-order vibration shape. The modal frequency includes first-order frequency and second-order frequency. The input quantity group of the initial neural network model for concrete dam damage identification is [first-order vibration shape, second-order vibration shape, first-order frequency, second-order frequency], and the output quantity group is [damage proportion, damage location, damage range].

[0053] Step 6: Train the neural network model and substitute the measured distributed vibration sensor data to obtain the recognition results of multiple characteristic parameters of concrete dam damage.

[0054] Neural network model training, where the convolution process calculation formula is as follows: ; Among them, bias represents the bias term, which is a trainable parameter of the output channel, input represents input, out represents output, and the size of the neural network model training input vector is ( N i , C in , L in ), the size of the neural network model training output vector is ( N i , C out , L out ), N i Indicates i Group training samples; C in represents the input vector channel, C out Represents the output vector channel; L in and L out Respectively represent the channel vector lengths of input and output, k Represents the current and convolution kernel w The dot product k input vector channels, Indicates j output vector channels; Represents the convolution operation, input( N i , k ) indicates the i The first k The input vector of channels, out( N i , ) indicates the i The first j The output vector obtained by channels; Apply the maximum pooling method, and the calculation formula is as follows: ; in, s is the step length, p Represents the vector p Segment pooling area, m Indicates 0~ p -1, the pooling area number, C j Indicates j channels that perform pooling operations, out( N i , Cj , p ) indicates the i Group training sample p Segment pooling area j The pooled output value obtained by the pooling channel, input( N i , C j , + m ) indicates the i Group training sample p Segment pooling area j The input vector of pooling channels.

[0055] Figure 6 The convergence of the CNN model when training five feature quantities at the same time is demonstrated. Each feature quantity of damage can also be well predicted. The convergence process of the training set and the validation set is highly consistent, and the damage parameter prediction effect is better. This shows that the CNN model with added convolution and pooling operations has a stronger feature capture capability and is suitable for deep mining of modal vibration data information.

[0056] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A concrete dam damage identification method based on multimodal parameter deep learning, characterized in that: The following steps are involved: Step 1: Collect the distributed vibration measurement values ​​of the concrete dam, denoise the distributed vibration measurement values, identify the operating modes of the concrete dam based on the distributed vibration measurement values, and extract the measured multi-modal parameters of the concrete dam; Step 2: Establish a finite element model based on the concrete dam engineering data, randomly generate the overall finite element parameter combination of the model, and obtain the corresponding modal parameters through dynamic simulation calculation; Step 3: Combining the modal parameters simulated in step 2 with the multi-modal parameter indicators measured in step 1, the overall parameter combination of the model is optimized to obtain a revised concrete dam simulation model; Step 4: Use the modified concrete dam simulation model to randomly generate the concrete dam damage degree, damage location and damage range, perform dynamic simulation calculations, and obtain a training sample set of the damaged concrete dam simulation model; Step 5: Establish the initial neural network model for concrete dam damage identification; Step 6: Based on the training sample set of the damaged concrete dam simulation model obtained in step 4, a neural network model is trained, and the measured distributed vibration sensor data is substituted to obtain the recognition results of multiple characteristic parameters of concrete dam damage.

2. According to claim 1, a concrete dam damage identification method based on multi-modal parameter deep learning is characterized in that: The distributed vibration measurement denoising in step 1 is specifically as follows: Based on the spatiotemporal two-dimensional distribution characteristics of distributed vibration measurements, the distributed vibration measurements are mapped to images, and the image quality is improved using data augmentation, logarithmic image processing, and histogram equalization methods. Image opening and closing operations and edge detection are used to extract typical image features. The noise-assisted two-dimensional EMD method is used to jointly denoise the distributed vibration measurements.

3. The method for concrete dam damage identification based on multimodal parameter deep learning according to claim 1 is characterized in that: The concrete dam operation mode identification in step 1 is specifically as follows: Step 1.1: Extract the non-stationary characteristics of the amplitude and power spectrum of the structure's spatiotemporal distribution of vibration measurements, and select the response sequence set and reference point set representing the significant vibration behavior of the structure based on the non-stationary index; Step 1.2: Use the selected response sequence set and reference point set to construct a transmissibility function matrix, which describes the ratio of the structural response point to the reference response point within a range of 2 times the second-order frequency of the concrete dam, is organized into columns by measurement points, and is organized into rows by frequency to form a matrix, covering the frequency range and having a frequency resolution of less than or equal to 0.05 Hz; Step 1.3: Perform variation and singular value decomposition on the transmissibility function matrix of different response points to solve the dynamic model of the transmissibility-based operational modal analysis method. In the singular value decomposition process, the singular values ​​are sorted from large to small, and the first two singular values ​​and their corresponding singular vectors are retained, and the remaining singular values ​​are eliminated; Step 1.4: Extract key modal parameters based on singular value decomposition, analyze the singular values ​​and corresponding singular vectors in the singular value decomposition results, and identify the modal frequency, damping ratio and vibration shape corresponding to each singular value.

4. The method for concrete dam damage identification based on multimodal parameter deep learning according to claim 1 is characterized in that: In the step 2, a finite element parameter combination of the concrete dam as a whole is randomly generated, and the finite element parameter combination includes elastic modulus, density, and Poisson's ratio.

5. The method for concrete dam damage identification based on multimodal parameter deep learning according to claim 1 is characterized in that: The optimization calculation of the overall parameter combination of the model in step 3 is as follows: The optimized objective function is as follows: ； in, x is the parameter combination to be corrected. Since no preset mass is required, the parameter combination includes the concrete dam density and material dynamic parameters. The parameter combination is within the range D x Inside, and D x The estimated value of the material parameters fluctuates by 20% to 50%; T ( x )and F ( x ) are the objective functions of time domain index and frequency domain index respectively; f i is the measured seismic response sequence at the ith discrete time point; g j is the modal frequency identified at the order of the jth modal frequency; and are the seismic response sequence simulated at the ith discrete time point and the model natural frequency simulated at the jth modal frequency order; n 1 and n 2 are the number of discrete time points and the number of modal frequency orders respectively; The optimization process is: Step 3.1: Analyze the quality of the multimodal parameters obtained in step 1, eliminate false modes, and select modal parameters that meet the convergence requirements; Step 3.2: Estimate the parameter range of the finite element model and obtain the corresponding multi-modal parameters under each parameter combination of the concrete dam in step 2; Step 3.3: Calculate the fitness value based on optimizing the objective function in the time-frequency domain and construct an initial population; Step 3.4: Use the MOEA / D optimization algorithm to optimize, decompose the multi-objective problem into multiple weighted single-objective sub-problems, and perform crossover and mutation operations under neighborhood search to generate new individuals and update the Pareto frontier solution set; Step 3.5: Determine whether the termination condition is met. If not, continue iterating; otherwise, output the Pareto optimal parameter combination set.

6. The method for concrete dam damage identification based on multimodal parameter deep learning according to claim 1 is characterized in that: In step 4, the damage degree, damage location and damage range of the concrete dam are randomly generated. The specific execution method is as follows: The damage location includes the horizontal and vertical coordinates of the damage center: first, the two-dimensional model of the dam body is divided into the smallest unit of triangle or trapezoid, and the probability of the damage center being located in the unit is determined by the ratio of the unit area to the total area; For the trapezoidal element, two random numbers between 0 and 1 are used to express the positions of two points randomly selected on the parallel edges, and then a new random number is used to express the random position on the line connecting the two points as the final damage center; The damage degree uses the maximum reduction ratio of the elastic modulus as a setting parameter, and randomly generates any value between the maximum reduction ratio and 1 to express the damage degree; The damage range uses the maximum damage width and the maximum damage height as setting parameters, randomly generates a value between 0 and the maximum setting parameters to define the damage range, and then reduces the unit elastic modulus of the generated finite element model: determine whether the centroid of each unit is in the damage range defined above, and if so, reduce the unit elastic modulus; For a two-dimensional finite element model, the damage characteristic parameter vector is as follows: ； in, They represent the horizontal and vertical coordinates of the damage center, the damage degree, and the width and height of the damage range respectively.

7. The method for concrete dam damage identification based on multimodal parameter deep learning according to claim 1 is characterized in that: The initial neural network model for concrete dam damage identification in step 5 is a one-dimensional CNN architecture with multiple inputs and multiple outputs, including an input layer, an output layer, a convolution layer, a pooling layer, a hidden layer, and a fully connected layer. The input of the initial neural network model for concrete dam damage identification includes modal vibration shapes and modal frequencies. The modal vibration shapes include first-order vibration shapes and second-order vibration shapes. The modal frequencies include first-order frequencies and second-order frequencies. The input quantity group of the initial neural network model for concrete dam damage identification is [first-order vibration shape, second-order vibration shape, first-order frequency, second-order frequency], and the output quantity group is [damage proportion, damage location, damage range].

8. The method for concrete dam damage identification based on multimodal parameter deep learning according to claim 1 is characterized in that: In the step 6, the neural network model is trained, and the convolution process calculation formula is as follows: ； Among them, bias represents the bias term, which is a trainable parameter of the output channel, input represents input, out represents output, and the size of the neural network model training input vector is ( N i , C in , L in ), the size of the neural network model training output vector is ( N i , C out , L out ), N i Indicates i Group training samples; C in represents the input vector channel, C out Represents the output vector channel; L in and L out Respectively represent the channel vector lengths of input and output, k Represents the current and convolution kernel w The dot product k input vector channels, Indicates j output vector channels; Represents the convolution operation, input( N i , k ) indicates the i The first k The input vector of channels, out( N i , ) indicates the i The first j The output vector obtained by channels; Apply the maximum pooling method, and the calculation formula is as follows: ； in, s is the step length, p Represents the vector p Segment pooling area, m Indicates 0~ p -1, the pooling area number, C j Indicates j channels that perform pooling operations, out( N i , C j , p ) indicates the i Group training sample p Segment pooling area j The pooled output value obtained by the pooling channel, input( N i , C j , + m ) indicates the i Group training sample p Segment pooling area j The input vector of pooling channels.

Images