A structural damage detection method based on transmissibility function matrix and two-dimensional convolutional neural network

Abstract

The present invention proposes a structural damage detection method based on a transmissibility function matrix and a two-dimensional convolutional neural network. The transmissibility function matrix containing structural damage characteristics is constructed by using the transmissibility function between all measuring points of the structure, and the transmissibility function matrix is ​​used as the training data of the two-dimensional convolutional neural network for structural damage monitoring. The method utilizes the characteristic that structural damage will inevitably lead to corresponding changes in the transmissibility function between non-adjacent measuring points, converts the original vibration signals between all measuring points into transmissibility functions that characterize the structural damage state, and constructs a transmissibility function matrix with damage characteristics. The matrix is ​​used as the input of the two-dimensional convolutional neural network for training to obtain a network model of the structure. The transmissibility function matrix in the healthy state of any structure is imported into the trained structural network model to identify which damage state it belongs to, thereby achieving the purpose of monitoring structural damage.

Description

Technical Field

[0001] The present invention belongs to the technical field of structural health monitoring, and in particular relates to a structural damage detection method based on a transmissibility function matrix and a two-dimensional convolutional neural network. Background Art

[0002] Structural health monitoring (SHM) systems play an important role in the manufacturing, service and maintenance of structures. As an important research topic in SHM, vibration-based structural damage detection methods have attracted widespread attention from many scholars in recent years. The core idea is to extract damage-sensitive features of the structure, which are functions of the physical properties of the structure (mass, stiffness and damping coefficient). Changes in the physical properties of the structure caused by structural damage can lead to changes in damage-sensitive features, which can then be used to identify damage.

[0003] Structural damage detection methods based on deep learning neural networks provide an effective way for structural health monitoring. Generally, damage detection methods based on deep learning neural networks learn damage features from structural vibration responses and then identify damage. However, the effectiveness of existing methods is limited by key shortcomings, such as: a large amount of vibration response data is generated during the vibration test of the structure. These original vibration response data are generally high in dimension and have a lot of redundant information. Direct use in neural networks may cause problems such as training difficulties and low computational efficiency; in addition, the response of some structures is affected by the form of excitation, so that the vibration response of the tested structure may contain insufficient damage-sensitive features, making the stability of the neural network low.

[0004] Tongwei Liu et al. from Hohai University proposed a deep learning structural damage detection method based on transmissibility function in an article titled "A Data-Driven Damage Identification Framework Based on Transmissibility Function Datasets and One-Dimensional Convolutional Neural Networks: Verification on a Structural Health Monitoring Benchmark Structure" (DOI: 10.3390 / s20041059) in the journal Sensors. The study found that compared with traditional data based on time domain signals and fast Fourier transform frequency domain signals, transmissibility function data contains more important damage-related features, and the transmissibility function has the characteristic of being unaffected by excitation, which makes this method highly stable in damage identification. However, the article used the entire transmissibility function as the input of the one-dimensional convolutional neural network, but the entire transmissibility function has the characteristic of high dimensionality. Direct use will lead to problems such as long training time of the neural network model and low generalization ability. Sergio Cofre-Martel et al. from the University of Chile proposed a method for structural damage detection by converting the transmissibility function into a grayscale image as the input of a two-dimensional convolutional neural network in an article titled "Deep Convolutional Neural Network-Based Structural Damage Localization and Quantification Using Transmissibility Data" (DOI: 10.1155 / 2019 / 9859281) in the journal Shock and Vibration. Each row of the image is a transmissibility function between different measuring points. However, if the number of measuring points is reduced, the grayscale image will also become smaller, and the accuracy of damage identification will decrease accordingly. Summary of the invention

[0005] The neural network-based damage detection method learns damage-sensitive features from structural responses and then identifies damage. A large amount of vibration response data will be generated during the service or vibration test of the structure. These original vibration response data are generally high in dimension and have a lot of redundant information. Direct use may cause problems such as difficulty in network model training and low computational efficiency. In addition, although the entire segment transmissibility function is used as the input of the one-dimensional convolutional neural network, it has high stability in damage identification. However, the use of the entire segment transmissibility function for network training still has the characteristics of high dimension. Direct use will lead to problems such as long training time and low generalization ability of the neural network model. When converting the transmissibility function of different measuring points into grayscale images as the input of the two-dimensional convolutional neural network, a large number of measuring point responses are also required to ensure that enough transmissibility functions are converted into image data when detecting structural damage. If the number of measuring points is reduced, the grayscale image will also become smaller accordingly, and the accuracy of damage identification will decrease accordingly.

[0006] In order to avoid the shortcomings of the prior art, the present invention uses the transmissibility function between all the measuring points of the structure to construct a transmissibility function matrix (TFM) containing the structural damage characteristics, uses the transmissibility function matrix as the training data of the two-dimensional convolutional neural network to monitor the structural damage, and proposes a structural damage detection method based on the transmissibility function matrix and the two-dimensional convolutional neural network. The method uses the characteristic that structural damage will inevitably lead to the corresponding change of the transmissibility function between non-adjacent measuring points, converts the original vibration signals between all measuring points into the transmissibility function that characterizes the structural damage state, constructs a transmissibility function matrix with damage characteristics, and uses it as the input of the two-dimensional convolutional neural network for training to obtain a network model of the structure. The transmissibility function matrix in the healthy state of any structure is imported into the trained structural network model to identify which damage state it belongs to, thereby achieving the purpose of monitoring the structural damage.

[0007] The method of the present invention uses only one or a certain section of the function value of the transfer function between each measuring point to construct a transfer function matrix with the same dimension as the number of measuring points (a two-dimensional single-value transfer function matrix), which can greatly reduce the input dimension of the neural network while containing sufficient damage information, making its calculation more efficient. In addition, the transfer function matrix can be constructed by taking the function values ​​corresponding to multiple frequency points of the transfer function between each measuring point (a three-dimensional multi-value transfer function matrix). While the number of measuring points is reduced, another part of the damage information is supplemented in its third dimension, thereby improving the accuracy of damage identification. Since the two-dimensional convolutional neural network has multiple convolution kernels, its processing speed for the three-dimensional matrix is ​​not much different from that of the two-dimensional matrix, so while improving the accuracy of damage identification, it will not prolong the training time of the network. Under the premise of ensuring that the training time is not increased, the recognition accuracy of structural damage monitoring is improved.

[0008] The technical solution of the present invention is:

[0009] A structural damage detection method based on a transmissibility function matrix and a two-dimensional convolutional neural network comprises the following steps:

[0010] Step 1: Collect the structural response

[0011] The time domain response data of the structure under test in different health states are collected as the original data for measuring the structural damage characteristics.

[0012] Step 2: Response data slicing processing

[0013] The original time domain response data obtained in step 1 is sliced ​​to obtain response slice data of each response measurement point on the measured structure in its corresponding structural health state.

[0014] Step 3: Construct structural damage characteristic matrix

[0015] The response slice data of the measured structure in a certain healthy state obtained in step 2 are processed to obtain the transferability function between the response measurement points in the slice segment and assemble them into a transferability function matrix.

[0016] Since the transmissibility function is a frequency domain function, different frequencies correspond to different function values. Therefore, different numbers of frequencies can be selected to assemble different transmissibility function matrices according to needs. If the transmissibility function response value corresponding to only one frequency is used to assemble the transmissibility function matrix, a two-dimensional matrix form of the transmissibility function matrix can be obtained, which is called S-TFM (Single-valued TFM); if the transmissibility function response values ​​corresponding to multiple frequencies are used at the same time, and the transmissibility function matrices at these frequency points are stacked and assembled, a three-dimensional matrix form of the transmissibility function matrix can be obtained, which is called M-TFM (Multivalued TFM).

[0017] The transmissibility function matrix contains all transmissibility function combinations of the response data. When damage occurs at any position on the structure, there will always be multiple elements in the transmissibility function matrix that will change, that is, the structural damage state corresponds to the transmissibility function matrix one by one.

[0018] Step 4: Build a sample database

[0019] Repeat step 3 to process the response slice data of each slice segment of the structure under test in different health states, and finally obtain the transfer rate function matrix S-TFM or M-TFM of each slice segment in different health states, which can constitute the sample database of the deep learning neural network.

[0020] Step 5: Deep Learning Neural Network Training

[0021] The data of the sample database obtained in step 4 (S-TFM or M-TFM) is imported into the deep learning neural network for training, and finally a neural network model containing different health states is obtained. Depending on the damage information of the model, S-TFM or M-TFM can be selected as the training data according to the training time and accuracy. In theory, the more frequencies in M-TFM, the more information reflecting the structural damage characteristics, that is, the more accurate the network model obtained after training.

[0022] Step 6: Structural Damage Monitoring

[0023] The time domain response data of the structure in any state is collected, and after processing according to steps 2 and 3, the transmissibility function matrix in this state is obtained. The transmissibility function matrix is ​​input into the trained structural network model to identify which health state the state belongs to, and finally the damage monitoring of the structure is realized.

[0024] Furthermore, in step 1, the time domain response includes displacement response, velocity response, and acceleration response.

[0025] Furthermore, in step 2, for the structure under test in a certain healthy state, the original time domain responses obtained at n measuring points are a1(t), a2(t),…, a n (t), the time domain response of the overall n measurement points is sliced ​​in time, and a total of N c The length of the sampling points is N d The time domain response data segment of .

[0026] Furthermore, in step 3, the time domain response data in a slice segment under a certain healthy state of the measured structure is processed to obtain the transfer rate function T between the n measuring points in the segment. kl (ω), T kl (ω) represents the transmissibility function between the response at the measuring point k and the response at the measuring point l, and ω is the frequency value. The transfer functions between the n measuring points are assembled to obtain the transmissibility function matrix T(ω) about the frequency ω containing the structural damage characteristics:

[0027]

[0028] If only one frequency value is used to assemble the transmissibility function matrix, the transmissibility function matrix is ​​a single-valued transmissibility function matrix S-TFM. If multiple frequency values ​​are used at the same time, multiple two-dimensional transmissibility function matrices are stacked and assembled in the third dimension to obtain a multi-valued transmissibility function matrix M-TFM.

[0029] Furthermore, all transmissibility function values ​​in the single-valued transmissibility function matrix S-TFM are selected to correspond to the transmissibility values ​​at the natural frequency points when the structure is not damaged.

[0030] Furthermore, all transmissibility function values ​​in the single-valued transmissibility function matrix S-TFM are selected to correspond to the transmissibility value at the peak of the transmissibility function response.

[0031] Furthermore, the formation process of the multi-valued transfer function matrix M-TFM is:

[0032] Firstly, the transmissibility value corresponding to the natural frequency point when the structure is not damaged is used as the transmissibility function value to construct a two-dimensional transmissibility function matrix;

[0033] Then, the transmissibility values ​​corresponding to the adjacent frequency points on both sides of the natural frequency point when the structure is not damaged are used as transmissibility function values ​​to construct multiple two-dimensional transmissibility function matrices;

[0034] Finally, the constructed two-dimensional transmissibility function matrix is ​​superimposed in the three-dimensional direction to form the final three-dimensional multi-valued transmissibility function matrix.

[0035] Furthermore, the formation process of the multi-valued transfer function matrix M-TFM is:

[0036] Firstly, the transmissibility value corresponding to the peak value of the transmissibility function response is used as the transmissibility function value to construct a two-dimensional transmissibility function matrix;

[0037] Then, the transmissibility values ​​corresponding to the adjacent frequency points on both sides of the transmissibility function response peak are used as transmissibility function values ​​to construct multiple two-dimensional transmissibility function matrices;

[0038] Finally, the constructed two-dimensional transmissibility function matrix is ​​superimposed in the three-dimensional direction to form the final three-dimensional multi-valued transmissibility function matrix.

[0039] Furthermore, in step 6, the way of processing the time domain response data of the structure in any state is the same as the way of processing the sample database used for training the deep learning neural network.

[0040] A computer-readable storage medium stores computer-executable instructions, which are used to implement the method described above when executed.

[0041] A computer system comprises: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described above.

[0042] A computer program comprises computer executable instructions, which are used to implement the method described above when the instructions are executed.

[0043] Beneficial Effects

[0044] The beneficial effects of the present invention are as follows: the present invention proposes a structural damage detection method based on a transmissibility function matrix and a two-dimensional convolutional neural network, and uses the transmissibility function values ​​between all measuring points to construct a transmissibility function matrix as input, and each element in the matrix is ​​only a certain value or a certain range of values ​​of the transmissibility function, which greatly reduces the input dimension of the network. The processing speed of a two-dimensional convolutional neural network for a low-dimensional matrix is ​​much higher than that of a high-dimensional array, so the input of a convolutional neural network with a transmissibility function matrix as a dimension can accelerate the training process of the network. In addition, when there are fewer structural measuring points, the method of the present invention can construct a multi-valued transmissibility function matrix when the number of measuring points is reduced, that is, the transmissibility function matrix increases the dimension in the third dimension, thereby avoiding increasing the dimension in the first and second dimensions (the calculation speed of the convolutional neural network is greatly affected by the dimensions of the first and second dimensions, and the influence of the dimension of the third dimension on its calculation speed can be ignored), thereby increasing the damage information related to the structural damage and achieving the purpose of improving the accuracy of damage identification.

[0045] The invention has a simple principle and obvious effect, provides a new idea for the engineering application of structural health monitoring in the future, and has broad application prospects in structural health monitoring.

[0046] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0048] Figure 1 It is a schematic diagram of a method for selecting function values ​​in a single-valued transmissibility function matrix of a structure in a healthy state in the present invention;

[0049] Figure 2 It is a schematic diagram of a method for selecting function values ​​in a single-valued transmissibility function matrix of a structure in a healthy state in the present invention;

[0050] Figure 3 It is a schematic diagram of the construction process of the single-valued transfer rate function matrix and the multi-valued transfer rate function matrix in the present invention;

[0051] Figure 4 It is a schematic diagram of the process of deep learning neural network training in the present invention. DETAILED DESCRIPTION

[0052] Embodiments of the present invention are described in detail below. The embodiments are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0053] The present invention is a structural damage detection method based on a deep learning neural network. The method utilizes the characteristic that structural damage will inevitably lead to corresponding changes in the transmissibility function between non-adjacent measuring points, converts the original vibration signals between all measuring points into a transmissibility function that characterizes the structural damage state, and constructs a transmissibility function matrix with damage characteristics. The matrix is ​​used as the input of a two-dimensional convolutional neural network for training to obtain a network model of the structure. The transmissibility function matrix under any structural state is imported into the trained structural network model to identify which health state it belongs to, thereby achieving the purpose of monitoring structural damage. Finally, a structural damage detection method based on the transmissibility function matrix and a two-dimensional convolutional neural network is established.

[0054] The method of the present invention only uses one or a certain section of the function value of the transfer function between each measuring point to construct a transfer function matrix with the same dimension as the number of measuring points (a two-dimensional single-value transfer function matrix), which can greatly reduce the input dimension of the neural network while containing sufficient damage information, making its calculation more efficient. In addition, the transfer function matrix can be constructed by taking the function values ​​corresponding to multiple frequency points of the transfer function between each measuring point (a three-dimensional multi-value transfer function matrix), while reducing the number of measuring points, another part of the damage information is supplemented in its third dimension, thereby improving the accuracy of damage identification. Since the two-dimensional convolutional neural network has multiple convolution kernels, its processing speed for the three-dimensional matrix is ​​not much different from that of the two-dimensional matrix, so while improving the accuracy of damage identification, it will not prolong the training time of the network. Under the premise of ensuring that the training time is not increased, the recognition accuracy of structural damage monitoring is improved. The present invention provides a new idea for the engineering application of structural health monitoring in the future, and has broad application prospects in structural health monitoring.

[0055] The specific steps of this embodiment are:

[0056] Step 1: Collect the structural response

[0057] Assuming there are n measuring points on the structure, the time domain response of the structure under k different health states is collected as the original data for measuring the structural damage characteristics. The time domain response includes displacement response, velocity response, and acceleration response.

[0058] Step 2: Response data slicing processing

[0059] The original time domain response data obtained in step 1 is sliced ​​to obtain response slice data of each response measurement point on the measured structure in its corresponding structural health state.

[0060] For the structure under test in a certain healthy state, the original time domain responses obtained at n measuring points are a1(t), a2(t),…, a n (t), the time domain response of the overall n measurement points is sliced ​​in time, and a total of N c The length of the sampling points is N d The time domain response data segment of .

[0061] Step 3: Construct structural damage characteristic matrix

[0062] The response slice data of the measured structure in a certain healthy state obtained in step 2 are processed to obtain the transferability function between the response measurement points in the slice segment and assemble them into a transferability function matrix.

[0063] Specifically, the time domain response data of a slice segment in a certain healthy state of the measured structure is processed to obtain the transfer rate function T between the n measuring points in the segment. kl (ω), T kl (ω) represents the transmissibility function between the response at the measuring point k and the response at the measuring point l, and ω is the frequency value. The transfer functions between the n measuring points are assembled to obtain the transmissibility function matrix T(ω) about the frequency ω containing the structural damage characteristics:

[0064]

[0065] Since the transmissibility function is a frequency domain function, different frequencies correspond to different function values. Therefore, different numbers of frequencies can be selected to assemble different transmissibility function matrices according to needs. If the transmissibility function response value corresponding to only one frequency value is used to assemble the transmissibility function matrix, a two-dimensional matrix form of the transmissibility function matrix can be obtained, which is called the single-valued transmissibility function matrix S-TFM (Single-valued TFM). If the transmissibility function response values ​​corresponding to multiple frequencies are used at the same time, and the transmissibility function matrices at these frequency points are stacked and assembled in the third dimension, a three-dimensional matrix form of the transmissibility function matrix is ​​obtained, which is called the multi-valued TFM (M-TFM).

[0066] The transmissibility function matrix T(ω) contains all transmissibility function combinations of the time domain response data. When damage occurs at any position on the structure, there will always be multiple elements in the transmissibility function matrix that will change, that is, the structural health state corresponds to the transmissibility function matrix one by one. In addition, the transmissibility function is a frequency domain function, and different frequencies correspond to different function values. According to the knowledge of vibration dynamics, the amplitude and corresponding frequency of the response peak in the frequency response data are more sensitive to changes in structural characteristics. Therefore, the function values ​​at the positions sensitive to structural damage can be selected to construct the transmissibility function matrix (such as: the transmissibility function value corresponding to the natural frequency point when the structure is not damaged or the transmissibility function value corresponding to the peak value of the transmissibility function response) as the neural network input to achieve the purpose of damage detection.

[0067] The transmissibility function matrix is ​​a two-dimensional single-value transmissibility matrix S-TFM, that is, all transmissibility functions in the transmissibility function matrix only select the transmissibility function value corresponding to one frequency point to assemble into the transmissibility function matrix. The two-dimensional single-value transmissibility function matrix can be constructed in the following two ways:

[0068] 1. All transmissibility function values ​​in the transmissibility function matrix are selected from the transmissibility values ​​corresponding to the natural frequency points when the structure is not damaged, such as Figure 1 This is because local damage of the structure usually has a greater impact on the high-order modal parameters of the structure, so the transmissibility function values ​​at the frequency points corresponding to the high-order natural frequencies are selected to construct the transmissibility function matrix.

[0069] 2. All transmissibility function values ​​in the transmissibility function matrix are selected to correspond to the transmissibility value at the peak of the transmissibility function response, such as Figure 2 This is because damage can cause the peak value of the transmissibility function response to shift, and the peak value of the transmissibility function response mostly appears in the low-frequency region. Therefore, selecting the value corresponding to the peak value of the transmissibility function response to construct the transmissibility function matrix can cope with the situation where the high-frequency data cannot be used due to the limitation of the acceleration sensor range, such as the function value in the high-frequency region may be inaccurate.

[0070] For a three-dimensional multi-valued transmissibility function matrix, that is, all transmissibility functions in the transmissibility function matrix select a section of frequency points to assemble the transmissibility function matrix. On the basis of the two-dimensional transmissibility function matrix, it is superimposed in the three-dimensional direction according to different frequencies, and finally assembled into a multi-valued transmissibility function matrix containing transmissibility function values ​​at multiple frequency points. The three-dimensional multi-valued transmissibility function matrix can be constructed in the following two ways:

[0071] 1. First, the transmissibility value corresponding to the natural frequency point when the structure is not damaged is used as the transmissibility function value to construct a two-dimensional transmissibility function matrix. Then, the transmissibility values ​​corresponding to the adjacent frequency points on both sides of the natural frequency point when the structure is not damaged are used as the transmissibility function values ​​to construct multiple two-dimensional transmissibility function matrices. Finally, the two-dimensional transmissibility function matrices constructed above are superimposed in the three-dimensional direction to form the final three-dimensional multi-valued transmissibility function matrix.

[0072] 2. First, the transmissibility value corresponding to the peak value of the transmissibility function response is used as the transmissibility function value to construct a two-dimensional transmissibility function matrix. Then, the transmissibility values ​​corresponding to the adjacent frequency points on both sides of the peak value of the transmissibility function response are used as the transmissibility function values ​​to construct multiple two-dimensional transmissibility function matrices. Finally, the two-dimensional transmissibility function matrices constructed above are superimposed in the three-dimensional direction to form the final three-dimensional multi-valued transmissibility function matrix.

[0073] Each element in the transfer function matrix is ​​only a certain value or a certain range of values ​​of the transfer function, which greatly reduces the input dimension of the network. The processing speed of the two-dimensional convolutional neural network for low-dimensional matrices is much higher than that of high-dimensional arrays, so using the transfer function matrix as input can accelerate the training process of the network. In addition, the number of measuring points is a point of concern in the field of health monitoring engineering, and for some complex structures, it is often impossible to arrange enough measuring points. The method of the present invention can increase the number of selected function values ​​when the number of measuring points is reduced, that is, the dimension of the transfer function matrix is ​​increased in the third dimension, thereby avoiding the increase in the dimension in the first and second dimensions (the calculation speed of the two-dimensional convolutional neural network is greatly affected by the dimensions of the first and second dimensions, and the influence of the dimension of the third dimension on its calculation speed can be ignored), thereby increasing the damage information related to the structural damage and improving the accuracy of damage identification.

[0074] Therefore, for the structure in one of the structural health states, an S-TFM or M-TFM can be constructed in a time domain response segment of n measurement points after slicing.

[0075] Step 4: Build a sample database

[0076] Repeat step 3 for all N response data of n measuring points under k structural health states. c Repeat steps 2 and 3 to process the segment data, and finally get N c ×k S-TFMs or M-TFMs can constitute a sample database of a two-dimensional convolutional neural network.

[0077] Step 5: Deep Learning Neural Network Training

[0078] like Figure 4As shown in Figure 1, the data of the sample database obtained in step 4 (S-TFM or M-TFM) is imported into the deep learning neural network for training, and finally a neural network model containing different health states is obtained. Depending on the damage information of the model, S-TFM or M-TFM can be selected as the training data according to the training time and accuracy. In theory, the more frequencies in M-TFM, the more information reflecting the structural damage characteristics, that is, the more accurate the network model obtained after training.

[0079] In step 3, as the number of frequency value points selected in the constructed multi-valued transfer function matrix increases, the network model trained in step 5 becomes more accurate. Especially in the case of a small number of measurement points, the construction of a multi-valued transfer function matrix can make up for the situation of a small number of measurement points and insufficient structural damage characteristics. Moreover, the training speed of the two-dimensional convolutional neural network is greatly affected by the dimensions of the first and second dimensions. The effect of increasing the dimension of the third dimension of the transfer function matrix on the training speed of the two-dimensional convolutional neural network can be ignored.

[0080] Step 6: Structural Damage Monitoring

[0081] The time domain response data of the structure in any state is collected, and after processing according to steps 2 and 3, the transmissibility function matrix in the state is obtained, and the obtained transmissibility function matrix is ​​input into the trained structural network model to identify which health state the state belongs to, and finally the damage monitoring of the structure is realized. The way of processing the time domain response data of the structure in any state here is the same as the way of processing the sample database used for training the deep learning neural network.

[0082] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and intent of the present invention.

Claims

1. A structural damage detection method based on a transmissibility function matrix and a two-dimensional convolutional neural network, characterized in that: The following steps are involved: Step 1: Collect the structural response Collect the time domain response data of the structure under test in different health states as the original data for measuring the structural damage characteristics; Step 2: Response data slicing processing Slice the original time-domain response data obtained in step 1 to obtain response slice data of each response measurement point on the measured structure in its corresponding structural health state; Step 3: Construct structural damage characteristic matrix The response slice data of the measured structure in a certain healthy state obtained in step 2 are processed to obtain the transferability function between the response measurement points in this slice segment and assemble it into a transferability function matrix; specifically: The time domain response data of a slice segment in a certain healthy state of the measured structure is processed to obtain the transfer rate function T between the n measuring points in the segment. kl (ω), T kl (ω) represents the transmissibility function between the response at the measuring point k and the response at the measuring point l, and ω is the frequency value. The transfer functions between the n measuring points are assembled to obtain the transmissibility function matrix T(ω) about the frequency ω containing the structural damage characteristics: At the same time, multiple frequency values ​​are used to assemble the transmissibility function matrix, so multiple two-dimensional transmissibility function matrices are stacked and assembled in the third dimension to obtain a multi-valued transmissibility function matrix M-TFM; Step 4: Build a sample database Repeat step 3 to process the response slice data of each slice segment of the structure under test in different health states, and finally obtain the transfer rate function matrix of each slice segment in different health states to form a sample database of the deep learning neural network; Step 5: Deep Learning Neural Network Training Import the data of the sample database obtained in step 4 into the deep learning neural network for training, and finally obtain a neural network model containing different health states; Step 6: Structural Damage Monitoring The time domain response data of the structure in any state is collected, and after processing according to steps 2 and 3, the transmissibility function matrix in this state is obtained. The transmissibility function matrix is ​​input into the trained structural network model to identify which health state the state belongs to, and finally the damage monitoring of the structure is realized.

2. The structural damage detection method based on the transmissibility function matrix and the two-dimensional convolutional neural network according to claim 1, characterized in that: In step 1, the time domain response includes displacement response, velocity response, and acceleration response.

3. The structural damage detection method based on a transmissibility function matrix and a two-dimensional convolutional neural network according to claim 1, characterized in that: In step 2, for the structure under test in a certain healthy state, the original time domain responses obtained at n measuring points are a1(t), a2(t),…, a n (t), the time domain response of the overall n measurement points is sliced ​​in time, and a total of N c The length of the sampling points is N d The time domain response data segment of .

4. The structural damage detection method based on a transmissibility function matrix and a two-dimensional convolutional neural network according to claim 1, characterized in that: The formation process of the multi-valued transfer function matrix M-TFM is: Firstly, the transmissibility value corresponding to the natural frequency point when the structure is not damaged is used as the transmissibility function value to construct a two-dimensional transmissibility function matrix; Then, the transmissibility values ​​corresponding to the adjacent frequency points on both sides of the natural frequency point when the structure is not damaged are used as transmissibility function values ​​to construct multiple two-dimensional transmissibility function matrices; Finally, the constructed two-dimensional transmissibility function matrix is ​​superimposed in the three-dimensional direction to form the final three-dimensional multi-valued transmissibility function matrix.

5. The structural damage detection method based on transmissibility function matrix and two-dimensional convolutional neural network according to claim 1, characterized in that: The formation process of the multi-valued transfer function matrix M-TFM is: Firstly, the transmissibility value corresponding to the peak value of the transmissibility function response is used as the transmissibility function value to construct a two-dimensional transmissibility function matrix; Then, the transmissibility values ​​corresponding to the adjacent frequency points on both sides of the transmissibility function response peak are used as transmissibility function values ​​to construct multiple two-dimensional transmissibility function matrices; Finally, the constructed two-dimensional transmissibility function matrix is ​​superimposed in the three-dimensional direction to form the final three-dimensional multi-valued transmissibility function matrix.

6. The structural damage detection method based on transmissibility function matrix and two-dimensional convolutional neural network according to claim 1, characterized in that: In step 6, the time domain response data of the structure in any state is processed in the same manner as the sample database used to train the deep learning neural network.

7. A computer-readable storage medium storing computer-executable instructions, wherein the instructions are used to implement the method according to any one of claims 1 to 6 when executed.

Images