Structural damage identification method based on two-dimensional vibration data and multi-class augmentation

Abstract

The invention discloses a structural damage identification method based on two-dimensionalization and multi-category augmentation of vibration data, and relates to the technical field of damage identification algorithms. The identification method comprises the following steps: firstly, in order to make full use of the temporal and spatial correlation information between sensors, the signals of multiple sensors are constructed into a two-dimensional multi-channel time-frequency spectrum diagram as the input of a multi-scale feature extraction network GoogLeNet; then, in view of the problem that it is difficult to obtain a sufficient number of damage samples for training a deep learning network in actual engineering, an auxiliary discriminator generative adversarial network ADC-GAN is used to realize multi-category sample data augmentation; then, the proposed hybrid loss function is used to train the multi-scale feature extraction network GoogLeNet to improve the training effect of the backbone network; finally, GoogLeNet is used to classify the real-time input samples to obtain the evaluation result of the structural state.

Description

Technical Field

[0001] The invention relates to the technical field of damage identification algorithms, and in particular to a structural damage identification method based on two-dimensionalization of vibration data and multi-category augmentation. Background Art

[0002] Structural damage identification is defined as accurate qualitative and quantitative judgment of changes in static and dynamic characteristics that are closely related to the mechanical properties of the structure. During the service life of large equipment and structures, the coupling of factors such as load, impact, environmental corrosion and material aging will inevitably lead to the accumulation of structural damage. Once the structural damage accumulates to a certain extent, if it is not identified and handled in time, the damage will expand rapidly, leading to major accidents.

[0003] At present, the structural damage identification methods mainly include those based on strain signals, temperature signals and vibration signals. The method based on vibration signals has stronger global and local damage identification capabilities and is the most widely used. The methods based on vibration signals can be divided into two categories: signal processing-based methods and machine learning-based methods. Signal processing-based methods include random subspace identification, wavelet transform and blind modal identification. These methods identify structural damage by comparing the dynamic characteristic parameters of the structure in different states, which is the main method in the early days. The method based on machine learning first extracts damage-related features from the vibration signal, and then inputs these features into the classifier for classification to achieve damage identification. Common methods for feature extraction include prediction residual and coefficient analysis of autoregressive models, principal component analysis, wavelet transform analysis, etc. The classifiers used include Gaussian process, support vector machine, Bayesian learning, etc. The method based on machine learning has shown advantages in many applications.

[0004] In recent years, deep learning has been increasingly used in the field of structural damage identification. Unlike traditional machine learning, deep learning methods do not rely on manually selected damage features. Luo Yongpeng et al. and Sharma et al. used a one-dimensional convolutional neural network to extract damage features from acceleration signals and classify them. The recognition accuracy of damage samples in small steel frames and steel plate frames reached 100% and 88.9% respectively, showing good damage identification performance. In engineering applications, multiple sensors are usually used to collect vibration data to achieve comprehensive structural status monitoring. However, the features extracted by deep learning using a one-dimensional convolutional neural network cannot fully characterize the information association between different sensors at the same time, and the utilization of multi-signal information is not sufficient.

[0005] In order to utilize the information association between sensors, a method has emerged to construct one-dimensional vibration signals into images and use images for damage identification. Liang Tao et al. used Markov transition fields to create multiple vibration signals into multi-channel images, and used a two-dimensional convolutional neural network to extract their features and classify them. In the simulated damage identification task of the IASC-ASCE steel frame, the overall sample recognition accuracy of this method can reach 94.4%. However, the two-dimensional convolutional neural network structure and loss function used in the above study still have room for further improvement.

[0006] In engineering applications, due to the relative lack of structural damage sample data, convolutional neural networks often face the problem of insufficient training, which affects the final recognition accuracy to a certain extent. To address the above problems, Luleci et al. used the Wasserstein generative adversarial network (WGAN) to generate damage samples and expand the sample library. The method was tested on the QUGS vibration test dataset publicly available at Qatar University, and the research object was a small steel frame stand structure. By using the augmented sample library for training, the proposed convolutional neural network achieved a recognition accuracy of 97% for the overall sample. However, as an unsupervised deep learning model, WGAN needs to build a separate model when generating samples of different categories, which leads to its relatively low efficiency in training and generating samples. Summary of the invention

[0007] In view of the above technical problems to be solved, the present invention provides a structural damage identification method based on two-dimensionalization and multi-category augmentation of vibration data, which uses a lightweight MLP architecture containing only a shallow structure and a small number of parameters to achieve high accuracy and high timeliness.

[0008] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0009] A structural damage identification method based on two-dimensionalization and multi-category augmentation of vibration data. The identification method first constructs the signals of multiple sensors into a two-dimensional multi-channel time-frequency spectrum according to the time and space correlation information between the sensors, and uses it as the input of a multi-scale feature extraction network; then, an auxiliary discriminator is used to generate an adversarial network to augment the multi-category sample data; then, a hybrid loss function is used to train the multi-scale feature extraction network; finally, the multi-scale feature extraction network is used to classify the real-time input samples to obtain the evaluation result of the structural state.

[0010] As a further improvement of the above technical solution:

[0011] Preferably, the signals of the multiple sensors are constructed as a two-dimensional multi-channel time-frequency spectrum diagram, comprising the following steps:

[0012] S1-1, for multiple vibration signals collected by c sensors , first the vibration signal Perform continuous wavelet transform to obtain × The time-frequency spectrum coefficient matrix :

[0013] (1)

[0014] in, for The wavelet transform coefficients of is the frequency scale number, for The sequence length, is the modulo operation; The calculation formula is:

[0015] (2)

[0016] (3)

[0017] in, is the Gabor mother wavelet The conjugate of for The scaling factor, for The translation factor of

[0018] S1-2, The value range is mapped to the grayscale value interval of [0,255], and the size is scaled to 224x224 pixels to obtain the time-frequency spectrum grayscale image ; Then, the time-frequency spectrum grayscale image Gamma correction is performed to obtain , the calculation formula is:

[0019] (4)

[0020] in, and They are the maximum value operation and the minimum value operation respectively. for The gamma coefficient of The calculation formula is:

[0021] (5)

[0022] in, is the adjustable gamma factor, for The image entropy and the average value of grayscale gradient, ;

[0023] S1-3, spliced ​​gamma-corrected time-frequency spectrum grayscale image Obtain a multi-channel time-frequency spectrum; correlate the spatial information of multiple sensor signals by increasing the spatial dimension.

[0024] Preferably, the method of using an auxiliary discriminator to generate an adversarial network to perform multi-category sample data augmentation is specifically as follows:

[0025] Introducing Conditional Generators in Generative Adversarial Networks and discriminative classifier Forming ADC-GAN, optimization objective function of ADC-GAN for:

[0026] (8)

[0027] in, is the joint distribution of real data and label information, is the joint distribution of generated data and label information; λ is an adjustable hyperparameter; Determines whether the input sample z of the classifier belongs to the real sample y + The classification probability, Generate sample y for z attribute - The classification probability is calculated as follows:

[0028] (9)

[0029] (10)

[0030] in, It is a feature extraction layer shared by the classifier and the discriminator; and is a fully connected layer that maps the real sample labels and generated sample labels into learnable feature vectors.

[0031] Preferably, the hybrid loss function is composed of Softmax loss 、 Center loss and exclusive regularization loss It consists of three parts: Guide the Softmax layer of the backbone network to form a classification decision boundary; center loss The role of is to measure the distance of high-dimensional embedded features and reduce the distance of intra-class sample features extracted by the backbone network; exclusive regularization loss The role of is to measure the distance of high-dimensional embedded feature clusters and increase the distance between class sample features extracted by the backbone network.

[0032] Preferably, the calculation formula of the hybrid loss function L is:

[0033] (11)

[0034] in, and represents the input feature vector and category of the last fully connected layer of the backbone network of the i-th sample in the structural state recognition model, and d represents the feature dimension; Represents the weight matrix of the last fully connected layer of the backbone network of the structural state recognition model The jth column of The i-th column of the weight matrix W of the last fully connected layer of the backbone network of the structural state recognition model, n and m represent the number of categories and the number of samples of the mini-batch training samples respectively; For the learnable The class center eigenvector, Indicates taking L2 norm operation; and is an adjustable hyperparameter used to balance 、 and The loss between ; k is the training round, N is the adjustable annealing hyperparameter, Indicates that the L1 norm operation is performed.

[0035] Preferably, the multi-scale feature extraction network is trained using a hybrid loss function, firstly, the ADC-GAN is trained using real samples of the damage class, and high-quality pseudo samples are generated using the trained ADC-GAN and histogram normalization; the pseudo samples are mixed with the real samples to establish an augmented sample library; then, the structural state recognition model is trained using the augmented sample library and the hybrid loss function.

[0036] Compared with the prior art, the structural damage identification method based on two-dimensionalization and multi-category augmentation of vibration data provided by the present invention has the following advantages:

[0037] The structural damage identification method based on two-dimensional vibration data and multi-category augmentation of the present invention, first, in order to make full use of the temporal and spatial correlation information between sensors, the signals of multiple sensors are constructed into a two-dimensional multi-channel time-frequency spectrum as the input of the multi-scale feature extraction network GoogLeNet; then, in view of the problem that it is difficult to obtain a sufficient number of damage samples for training deep learning networks in actual engineering, the auxiliary discriminator generative adversarial network ADC-GAN is used to realize multi-category sample data augmentation; then, the proposed hybrid loss function is used to train GoogLeNet to improve the training effect of the backbone network; finally, GoogLeNet is used to classify the real-time input samples to obtain the evaluation results of the structural state. The various components of the method of the present invention: the data augmentation model ADC-GAN, the structural state identification model GoogLeNet and the hybrid loss function, all have the advantage of improving the accuracy of sample identification. Compared with the representative signal processing method and the new deep learning method, the method of the present invention is not only more accurate, but also has a lower demand for damage samples, and has good practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is the overall framework diagram of the present invention.

[0039] Figure 2 This is the overall structure diagram of the standard GAN.

[0040] Figure 3 It is the overall structure diagram of ADC-GAN in the present invention.

[0041] Figure 4 It is a flow chart of training and monitoring in the present invention.

[0042] Figure 5 It is a diagram of the sensor arrangement and structural damage location of the research object corresponding to the experimental data in the experimental verification of the present invention.

[0043] Figure 6 (a) shows the natural frequency of the truss tower from September 2020 to June 2021 in the experimental verification of the present invention.

[0044] Figure 6(b) shows the natural frequency of the truss tower from May to July 2021 in the experimental verification of the present invention. DETAILED DESCRIPTION

[0045] The specific embodiments of the present invention are described in detail below. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0046] The structural damage identification method based on two-dimensionalization and multi-category augmentation of vibration data of the present invention, firstly, in order to make full use of the temporal and spatial correlation information between sensors, the signals of multiple sensors are constructed into a two-dimensional multi-channel time-frequency spectrum diagram as the input of the multi-scale feature extraction network GoogLeNet; then, in view of the problem that it is difficult to obtain a sufficient number of damage samples for training deep learning networks in actual engineering, the auxiliary discriminator generative adversarial network ADC-GAN is used to realize multi-category sample data augmentation; then, the proposed hybrid loss function is used to train GoogLeNet to improve the training effect of the backbone network; finally, GoogLeNet is used to classify the real-time input samples to obtain the evaluation results of the structural state.

[0047] The structural damage identification method based on two-dimensionalization and multi-category augmentation of vibration data of the present invention includes two stages, online and offline. Figure 1 As shown in the figure. In the offline stage, the continuous wavelet transform is first used to transform the vibration signal of the sensor into a two-dimensional time-spectrogram. The gamma brightness correction is performed on the time-spectrogram to eliminate the brightness difference caused by the spatial position. The time-spectrogram is spliced ​​to construct a multi-channel time-spectrogram as the input of the structural state recognition model. Then, the data augmentation model is trained and pseudo samples of structural damage are generated. The pseudo samples are histogram-normalized to improve their image quality and realize multi-category sample data augmentation. Finally, the proposed hybrid loss function and augmented sample library are used to jointly train the structural state recognition model to improve its training effect.

[0048] In the online stage, a multi-channel time-frequency spectrum is constructed through online monitoring data, and the structural state recognition model trained in the offline stage is used to classify the real-time input to obtain the structural state evaluation results.

[0049] The structural damage identification method based on two-dimensionalization and multi-category augmentation of vibration data of the present invention comprises the following steps:

[0050] Step S1, construction of multi-channel time-frequency spectrum

[0051] In order to make full use of the temporal and spatial correlation information between sensors, the one-dimensional time series signals of multiple sensors are converted into multi-channel time-frequency spectrograms as the input of the subsequent structural state recognition model. Continuous wavelet transform can effectively capture the time-frequency characteristics of non-stationary and noisy vibration signals. In this embodiment, continuous wavelet transform is used to obtain the time-frequency grayscale image of a single vibration signal.

[0052] S1-1, for multiple vibration signals collected by c sensors , first the vibration signal Perform continuous wavelet transform to obtain × The time-frequency spectrum coefficient matrix :

[0053] (1)

[0054] in, for The wavelet transform coefficients of is the frequency scale number, for The sequence length, It is a modulo operation. The calculation formula is:

[0055] (2)

[0056] (3)

[0057] in, is the Gabor mother wavelet The conjugate of for The scaling factor, for The translation factor of .

[0058] S1-2, to adapt to the input format of the structural state recognition model, the time-frequency spectrum coefficient matrix Post-processing. The value range is mapped to the grayscale value interval of [0,255], and the size is scaled to 224x224 pixels to obtain the time-frequency spectrum grayscale image Then, in order to eliminate the brightness difference of the time-frequency spectrum caused by the spatial position of the sensor, the time-frequency spectrum grayscale image Gamma correction is performed to obtain , the calculation formula is:

[0059] (4)

[0060] in, and They are the maximum value operation and the minimum value operation respectively. for The gamma coefficient of The calculation formula is:

[0061] (5)

[0062] in, is the adjustable gamma factor, for The image entropy and the average value of grayscale gradient, .

[0063] In this embodiment, is 0.5, then The calculation formula is:

[0064] (6)

[0065] in, Represents the probability distribution of pixel gray levels in the image, H represents the height of the image, and W represents the width of the image. Represents the horizontal gradient of the image at pixel coordinates (n, m), Represents the horizontal gradient of the image at pixel coordinate (n, m).

[0066] S1-3, spliced ​​gamma-corrected time-frequency spectrum grayscale image Obtain a multi-channel time-frequency spectrum. By increasing the spatial dimension, the spatial information of multiple sensor signals can be effectively correlated.

[0067] Step S2, data augmentation

[0068] In actual engineering, the number of structural damage samples is usually limited and insufficient to fully train the structural state recognition model. To address this problem, the method proposed in this embodiment uses an improved generative adversarial network (GAN) to achieve multi-category sample data augmentation before training the structural damage model.

[0069] The standard GAN consists of two neural networks: the generator G and the discriminator D. The overall structure is as follows: Figure 2 As shown in Figure 2. The role of the generator is to generate realistic samples to deceive the discriminator, while the role of the discriminator is to accurately distinguish between real and fake samples. There is a zero-sum game between the generator and the discriminator. When the two reach a Nash equilibrium, the generator works best. The optimization objective function V(G,D) of GAN is:

[0070] (7)

[0071] Among them, z is the input sample of the discriminator D, and are the marginal distributions of real samples and generated samples respectively, Indicates that the real data distribution The expectation for z is Indicates that the data distribution is generated D(z) represents the expectation of z under the discriminator D, and D(z) represents the probability prediction value of z being a true sample.

[0072] The standard GAN is an unsupervised deep learning model. It has problems such as unstable training and mode collapse during application. In addition, due to the lack of supervision of category label information, the standard GAN can only generate pseudo samples of a single category. To solve these problems, this embodiment uses the ADC-GAN model. ADC-GAN introduces a conditional generator and discriminative classifier Improving the standard GAN. Conditional Generator By embedding conditional information into random noise input, conditional generation of samples of specified categories is achieved, which improves the controllability of generated samples. By introducing the classification loss of real and generated samples, the training process of the generator is supervised by the category labels, which improves the diversity of generated samples.

[0073] The overall structure of ADC-GAN is as follows Figure 3 Compared with the optimization objective function of the standard GAN, the optimization objective function of ADC-GAN is Optimization added The classification loss is:

[0074] (8)

[0075] in, is the joint distribution of real data and label information, is the joint distribution of generated data and label information; λ is an adjustable hyperparameter. Determines whether the input sample Z of the classifier belongs to the real sample y + The classification probability of Generate sample y for Z attribute - The classification probability is calculated as follows:

[0076] (9)

[0077] (10)

[0078] in, It is a feature extraction layer shared by the classifier and the discriminator; and is a fully connected layer that maps the real sample labels and generated sample labels into learnable feature vectors.

[0079] Step S3, structural state recognition model

[0080] In order to fully exploit the structural vibration information carried by the multi-channel time-frequency spectrum, this embodiment uses a two-dimensional deep convolutional neural network to extract time-frequency and spatial feature information to achieve structural state recognition.

[0081] The structural state recognition model consists of a backbone network and a Softmax classification layer. The backbone network extracts low-level and high-level features of the input data in a progressive manner, and this process directly determines the performance of the structural state recognition. Therefore, it is very important to select a suitable backbone network for the recognition model. Generally speaking, increasing the number of layers of the backbone network can improve the performance of the network; however, the increase in the number of layers will also bring some new problems, such as overfitting, gradient disappearance, and gradient explosion. GoogLeNet is a backbone network developed by Google. It uses the Inception architecture to introduce parallel processing to reduce the number of parameters in the network and improve the training speed and accuracy of the network. Compared with early backbone networks such as AlexNet and VGGNet, the network has fewer parameters and faster inference speed; compared with the subsequent ResNet, it can better extract multi-scale feature information of input samples. Therefore, this embodiment uses GoogLeNet as the backbone network of the structural state recognition model.

[0082] Softmax is a widely used loss function for deep learning model training. In the method proposed in this embodiment, the structural state recognition model is a multi-classification task model; and the difference between the damage class samples is not large. Under the above conditions, the model trained with Softmax is easily interfered by noise.

[0083] In order to improve the performance of the structural state recognition model, this embodiment proposes a hybrid loss function L based on the Softmax function as the loss function of the training process. L is composed of the Softmax loss 、 Center loss and exclusive regularization loss It consists of three parts. It is the traditional Softmax loss, which is responsible for guiding the Softmax layer of the backbone network to form a classification decision boundary. , this embodiment sets the bias of the last fully connected layer of the backbone network to 0, so The bias of the last fully connected layer is not taken into account. Center loss The role of is to measure the distance of high-dimensional embedded features and reduce the distance of intra-class sample features extracted by the backbone network; exclusive regularization loss The role of is to measure the distance of high-dimensional embedded feature clusters and increase the distance between class sample features extracted by the backbone network.

[0084] The calculation formula of L is:

[0085] (11)

[0086] in, and represents the input feature vector and category of the last fully connected layer of the backbone network of the i-th sample in the structural state recognition model, and d represents the feature dimension; Represents the weight matrix of the last fully connected layer of the backbone network of the structural state recognition model The jth column of The i-th column of the weight matrix W of the last fully connected layer of the backbone network of the structural state recognition model, n and m represent the number of categories and the number of samples of the mini-batch training samples respectively; For the learnable The class center eigenvector, Indicates taking L2 norm operation; and is an adjustable hyperparameter used to balance 、 and The loss between ; k is the training round, N is the adjustable annealing hyperparameter, Indicates that the L1 norm operation is performed.

[0087] Step S4, training and monitoring process

[0088] The overall training process of this embodiment is as follows Figure 4 As shown, the multi-channel time-spectrogram used is divided into 8:2 ratios, which are used for the offline stage and the online stage of the simulation respectively.

[0089] In the offline simulation stage, the data augmentation model is first trained with real samples of damage, and high-quality pseudo samples are generated using the trained model and histogram specification. The pseudo samples are mixed with real samples to establish an augmented sample library. Then, the structural state recognition model is trained using the augmented sample library and the mixed loss function. In the online stage, the input samples are classified with the help of the trained structural state recognition model to complete the structural state monitoring.

[0090] When training the data augmentation model, ADC-GAN is used as the data augmentation model, and formula (8) is used as its loss function. The parameter λ in formula (8) is set to 1. The optimization algorithms of the generator, discriminator, and discriminant optimizer of ADC-GAN all use the Adam optimizer, and the learning rates are set to 5×10 -5 , 2×10 -4 and 2×10 -4 The maximum training rounds of ADC-GAN are set to 50, and the mini-batch size of each round is 8.

[0091] When training the structural recognition model, GoogLeNet is used as the structural recognition model, and formula (11) is used as its loss function. The parameters in formula (11) are , and N are set to 1, 2, and 1 respectively. The optimization algorithm of GoogLeNet adopts Adam optimizer, and the learning rate is set to 2×10 -4 The maximum training round of GoogLeNet is set to 50, and the mini-batch size of each round is 52.

[0092] Experimental verification

[0093] (1) Data sources and generation of multi-channel time-frequency spectrograms

[0094] The experimental data is a publicly available structural vibration test data. The research object is a 9m high truss tower. The accelerometer and structural damage locations of the truss tower are as follows: Figure 5 As shown, ML1~ML9 correspond to the locations of accelerometers numbered from 01x to 09x, and the sampling frequency of the accelerometers is 1651Hz; DAM1-DAM6 correspond to the locations of structural damage.

[0095] The public data records the structural response vibration signals of the truss tower under natural excitation from August 1, 2020 to July 31, 2021. These signals include healthy state vibration signals and vibration signals of 6 different damage states. Some data collected by accelerometers numbered from 01x to 09x were selected for the experiment. The detailed information of these data is shown in Table 1.

[0096] Table 1 Experimental data details:

[0097]

[0098] The one-dimensional vibration signals collected by multiple sensors are processed jointly and converted into multi-channel time-frequency spectrograms as the input of the structural state recognition model. The specific process is as follows: (1) A sliding window with a length of 2048 is used to perform non-overlapping slicing on the selected experimental data (accelerometer numbers from 01x to 09x) to obtain signal samples. , where n=1,...,11592; (2) Convert the above signal samples into multi-channel time-spectrum diagrams. The detailed information of the multi-channel time-spectrum diagrams is shown in Table 2, where there are 6 structural damage categories, namely: DAM6(S), DAM4(S), DAM3(S), DAM6(M), DAM4(M) and DAM3(M).

[0099] Table 2 Multi-channel time-spectrum diagram details:

[0100]

[0101] (2) Experimental platform and model training details

[0102] The overall training process is as follows Figure 4 As shown in Figure 2. In each cross-validation, a total of 11,592 samples were used, of which 9,274 samples were used in the offline phase of the simulation and 2,318 samples were used in the online phase of the simulation. In the samples used in these two phases, the ratio of various types of damaged samples to healthy state samples was consistent, both 1:6.

[0103] In the offline stage, we first use all real samples (4626) of all damage classes to train the data augmentation model ADC-GAN and generate fake samples of damage classes. Pseudo samples and real samples are mixed to form an augmented sample library (13900), in which the ratio of damage class samples to healthy state samples is 2:6. The backbone network of the data augmentation model ADC-GAN is implemented using the backbone network of BigGAN, and the hyperparameter λ of its optimization objective function is set to 1. In ADC-GAN, the generator , discriminator D and discriminant classifier The optimization algorithms are all set to Adam optimizer, and their learning rates are set to 5×10 -5 , 2×10 -4 and 2×10 -4 The maximum training iteration number of ADC-GAN is set to 50, and the Mini-batch size of each round is set to 8.

[0104] Then, the augmented sample library (13900) is used to train the structural state recognition model. The mixed loss function is used for training, and its hyperparameters , and N are set to 1, 2, and 1, respectively. The optimization algorithm of the structural state recognition model is set to Adam optimizer, and the learning rate is set to 2×10 -4 ; The maximum number of training iterations is set to 50, and the Mini-batch per round is set to 52.

[0105] (3) Overall performance evaluation

[0106] The proposed method is compared with representative signal processing methods and deep learning methods by using Accuracy, Precision, Recall, F1-score, False Positive Rate (FPR) and False Negative Rate (FNR) evaluation indicators, so as to evaluate the overall performance of the method of the present invention.

[0107] The evaluation index of the online sample recognition result is obtained by using 5-fold cross validation, and the accuracy of the method of the present invention in identifying whether the structure has defects is shown in Table 3. As can be seen from Table 3, the Accuracy and F1-Score of the method of the present invention in identifying whether the structure has defects are 98.10% and 97.91% respectively.

[0108] Table 3 Recognition accuracy of the method of the present invention in the overall sample (%):

[0109]

[0110] The public data of this experiment uses a representative signal processing method of covariance-driven random subspace identification to analyze the difference in the natural frequency of the structure under different states. The results are shown in Figure 6. As shown in Figure 6 (a), the natural frequency (gray) of DAM6 (S) when severe damage occurs and the natural frequency (cyan and blue) when the structure is healthy are significantly different. The natural frequency can be used as a criterion to effectively distinguish the above two states; the natural frequencies of DAM3 (S) and DAM4 (S) when severe damage occurs are not significantly different from the natural frequencies (blue, red and green) when the structure is healthy. At this time, it is difficult to distinguish the above three states using the natural frequency as a criterion. As shown in Figure 6 (b), the calculation results of another time period data show that the natural frequencies (gray) of the three mild damages DAM6 (M), DAM4 (M) and DAM3 (M) are very close to the natural frequencies (blue, red and green) when the structure is healthy, and cannot be effectively distinguished. Therefore, the method of the present invention can effectively identify mild and severe structural damage.

[0111] The experiment proposes that 3DS-CNN network processes and fuses multiple vibration signal features to realize structural damage identification, which is a representative new method for structural damage identification based on deep learning. Among them, the ratio of various types of damaged samples to healthy samples is 1:1. The overall recognition accuracy of the experimental method in the validation set is 81.43%, which can achieve effective structural state identification. However, the method proposed in this invention significantly improves the overall recognition accuracy by 16.67%.

[0112] In summary, the method of the present invention shows superior performance compared to the signal processing method using covariance-driven random subspace recognition and the 3DS-CNN deep learning method. This method can not only effectively identify various structural states, but also has a high overall recognition accuracy. In addition, the experimental data sources of the method of the present invention and the comparative method are consistent, which are complex structural response vibration signals stimulated by natural excitation. Therefore, when processing vibration signals containing noise and interference generated by natural excitation, the proposed method has practical value in engineering applications.

[0113] The above implementation cases are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Although the present invention has been disclosed as above with preferred embodiments, they are not intended to limit the present invention. Therefore, any simple modification, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A structural damage identification method based on two-dimensional vibration data and multi-category augmentation, It is characterized in that The recognition method first constructs the signals of multiple sensors into a two-dimensional multi-channel time-frequency spectrum diagram according to the time and space correlation information between the sensors, and uses it as the input of the multi-scale feature extraction network; then, the auxiliary discriminator generates an adversarial network to perform multi-category sample data augmentation; and then, a mixed loss function is used to train the multi-scale feature extraction network; Finally, a multi-scale feature extraction network is used to classify the real-time input samples to obtain the evaluation results of the structural status; The signals of the multiple sensors are constructed into a two-dimensional multi-channel time-frequency spectrum diagram, comprising the following steps: S1-1, for Multiple vibration signals collected by sensors , first the vibration signal Perform continuous wavelet transform to obtain The time-frequency spectrum coefficient matrix : （1） in, for The wavelet transform coefficients of is the frequency scale number, for The sequence length, is the modulo operation; The calculation formula is: （2） （3） in, is the Gabor mother wavelet The conjugate of for The scaling factor, for The translation factor of S1-2, The value range is mapped to the grayscale value interval of [0,255], and the size is scaled to 224x224 pixels to obtain the time-frequency spectrum grayscale image ; Then, the time-frequency spectrum grayscale image Gamma correction is performed to obtain , the calculation formula is: （4） in, and They are the maximum value operation and the minimum value operation respectively. for The gamma coefficient of The calculation formula is: （5） in, is the adjustable gamma factor, for The image entropy and the average value of grayscale gradient, ; S1-3, spliced ​​gamma-corrected time-frequency spectrum grayscale image Obtain multi-channel time-frequency spectrograms; correlate the spatial information of multiple sensor signals by increasing the spatial dimension; The method of using the auxiliary discriminator to generate an adversarial network to perform multi-category sample data augmentation is specifically as follows: Introducing Conditional Generators in Generative Adversarial Networks and discriminative classifier Forming ADC-GAN, optimization objective function of ADC-GAN for: （8） in, is the joint distribution of real data and label information, To generate the joint distribution of data and label information; is an adjustable hyperparameter; Input samples for the discriminative classifier Attribution of real samples The classification probability, for Attribution generation sample The classification probability is calculated as follows: （9） （10） in, It is a feature extraction layer shared by the classifier and the discriminator; and is a fully connected layer that maps the real sample labels and generated sample labels into learnable feature vectors.

2. The structural damage identification method based on two-dimensionalization and multi-category augmentation of vibration data according to claim 1, It is characterized in that The hybrid loss function is composed of Softmax loss , center loss and exclusive regularization loss It consists of three parts: Guide the Softmax layer of the backbone network to form a classification decision boundary; center loss The role of is to measure the distance of high-dimensional embedded features and reduce the distance of intra-class sample features extracted by the backbone network; exclusive regularization loss The role of is to measure the distance of high-dimensional embedded feature clusters and increase the distance between class sample features extracted by the backbone network.

3. The structural damage identification method based on two-dimensionalization and multi-category augmentation of vibration data according to claim 2, It is characterized in that The hybrid loss function The calculation formula is: （11） in, and Representative The input feature vector and category of the last fully connected layer of the backbone network of the structural state recognition model. Represents feature dimension; Represents the weight matrix of the last fully connected layer of the backbone network of the structural state recognition model No. List, Represents the weight matrix of the last fully connected layer of the backbone network of the structural state recognition model No. List, and Respectively represent the number of categories and the number of samples in the mini-batch training sample; For the learnable The class center eigenvector, Indicates taking L2 norm operation; and is an adjustable hyperparameter used to balance , , Losses between is the training round, is an adjustable annealing hyperparameter, Indicates that the L1 norm operation is performed.

4. The structural damage identification method based on two-dimensionalization and multi-category augmentation of vibration data according to claim 3, It is characterized in that The method adopts a hybrid loss function to train a multi-scale feature extraction network, first uses real samples of damage type to train ADC-GAN, and uses the trained ADC-GAN and histogram normalization to generate high-quality pseudo samples; the pseudo samples are mixed with real samples to establish an augmented sample library; then, the augmented sample library and the hybrid loss function are used to train the structural state recognition model.

Images