Structural damage identification method and device based on parallel convolutional neural network
By using a parallel convolutional neural network-based method in structural damage recognition to process multi-channel vibration signals, the problem of insufficient recognition performance and fitting ability in the prior art is solved, and higher recognition accuracy and structural health monitoring capabilities are achieved.
Patent Information
- Application Number
- CN202111422491.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-26
- Publication Date
- 2025-03-14
- Estimated Expiration
- 2041-11-26
AI Technical Summary
The existing structural damage recognition model based on 1D-CNN cannot effectively process multi-channel vibration signals, resulting in insufficient recognition performance and fitting capabilities, which cannot meet the needs of modern civil engineering structure health monitoring.
A structural damage recognition method based on parallel convolutional neural networks (P-CNN) is proposed. Multiple parallel convolutional neural networks are combined into a network group, and the data of multiple acceleration sensors are processed separately, multi-domain features are extracted and damage type prediction is performed, damage degree is calculated and weighted averaged is performed, and the identification accuracy of structural damage is improved.
It improves the performance and fitting ability of structural damage recognition, can more accurately identify the degree and location of structural damage, and meets the needs of modern civil engineering structure health monitoring.
Smart Images
Figure CN114357855B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of bridge health monitoring, and in particular to a structural damage identification method and device based on a parallel convolutional neural network. Background Art
[0002] During the long-term service of civil engineering structures, due to external loads, material performance degradation, and the impact of emergencies, it is inevitable that structural damage of varying degrees will occur, which will lead to a reduction in the structural bearing capacity and affect the reliability and safety of the structure. The civil engineering structural health monitoring system (SHM) collects long-term monitoring data from sensors, monitors and evaluates the structural status, issues early warning information in a timely manner, and makes decisions on inspection, repair, and reinforcement. It is a cutting-edge technology for improving and upgrading the level of intelligent maintenance and management of civil engineering structures. In order to realize a structural health monitoring (SHM) system that provides early warning of structural damage, the existing technology has proposed a variety of structural damage detection methods.
[0003] When a structure is damaged, the modal parameters (frequency, vibration mode, damping ratio) of the structure will change, thus affecting the dynamic response of the entire structure. The vibration-based identification method uses the structural dynamic response to identify the physical parameters of the structure and their changes, thereby judging the existence, severity and location of structural damage, and thus evaluating the overall performance of the structure. Vibration-based global damage detection technology can be divided into parametric and non-parametric methods. Compared with the non-parametric method that uses statistical means to directly identify damage from the measured signal, the parametric method determines the modal parameters from the measured response.
[0004] In recent years, various machine learning algorithms have been applied to parametric and non-parametric global damage detection techniques. Machine learning-based damage detection mainly consists of two steps: feature extraction and classification. In the parametric-based machine learning methods, the modal parameters of the structure are calculated from the structural vibration response and regarded as the extracted features. Meanwhile, the non-parametric-based machine learning methods mainly include the following feature extraction techniques, such as simple statistical analysis methods, regression models, principal component analysis, wavelet transform, and other time-frequency analysis methods. For the classification step, a large number of classifiers have been applied to parametric and non-parametric machine learning, such as probabilistic neural networks, artificial neural networks, fuzzy neural networks, singular value decomposition, and support vector machines. As expected, the “hand-crafted” choice of features and classifiers leads to significant differences in their performance in different data modes.
[0005] With the rapid development of sensor technology and computing hardware, the concept of deep learning has shown considerable importance in structural damage detection in the face of high-dimensional data sets composed of massive structural dynamic responses. CNN, as a deep neural network method that combines feature extraction and classifiers, does not require manual setting of feature extractors. In addition, due to the characteristics of sparse connections and weight sharing, CNN has significant advantages in computational efficiency, generalization ability, and noise resistance.
[0006] At present, many structural damage identification methods based on deep learning have been developed, but there are still some limitations, which restrict the application of deep learning in structural health monitoring. Although 1D-CNN has powerful signal classification capabilities, it is only applicable to one-dimensional signals. Faced with vibration signals collected by multiple sensors in actual engineering, some scholars connect multiple channel data in series as a single data process. However, this will increase the dimension of the input signal and affect the performance of CNN. At the same time, the structural vibration signal is distributed in a wide frequency range and is affected by noise. It contains a lot of redundant information, and the damage characteristics in the signal are easily submerged; the damage identification model based on 1D-CNN in the existing technology cannot meet the needs of modern civil engineering structural health monitoring. Summary of the invention
[0007] In view of the above problems, the present invention aims to provide a structural damage identification method and device based on a parallel convolutional neural network.
[0008] The purpose of the present invention is achieved by adopting the following technical solutions:
[0009] First, a structural damage identification method based on parallel convolutional neural network is proposed, including:
[0010] S11 data sampling: collecting acceleration data recorded by an acceleration sensor, wherein the acceleration sensor is arranged at a set position of the target structure, and the acceleration data includes a vibration response recorded by the acceleration sensor after applying an excitation to the target structure;
[0011] S12 data preprocessing: preprocessing the acceleration data recorded by each acceleration sensor to obtain an acceleration data sequence and its time-frequency diagram;
[0012] S13 Damage Identification: The input data sets corresponding to each acceleration sensor are input into the trained damage identification model based on parallel convolutional neural network. The damage identification model extracts features and predicts damage categories of the input data sets to obtain the probability of meeting the data characteristics of each damage condition. The corresponding structural damage identification results are obtained based on the probability of meeting the data characteristics of each damage condition. The input data sets include acceleration time series and their time-frequency diagrams.
[0013] In one implementation, step S11 specifically includes:
[0014] applying an incentive to a target structure, wherein the target structure includes a framework structure;
[0015] Receive the vibration response recorded by an accelerometer set on the target structure:
[0016]
[0017] Where K p Represents the acceleration data recorded by the pth acceleration sensor, where p∈[1,P], and P represents the total number of acceleration sensors.
[0018] In one implementation, step S12 specifically includes:
[0019] The acceleration data recorded by each acceleration sensor is randomly intercepted, and a fixed-length data segment is intercepted as a data sample, and N segments of data are obtained to form a data sample:
[0020]
[0021] Where K p,n represents the nth segment of data randomly intercepted from the acceleration data recorded by the pth acceleration sensor, where n∈[1,N]; the length of each segment of data is set to L;
[0022] The data samples are subjected to zero-mean normalization and continuous wavelet transform to obtain the standardized time series and its wavelet time-frequency diagram.
[0023] In one embodiment, in step S13, the trained damage identification model based on parallel convolutional neural networks includes a network group composed of multiple parallel convolutional neural networks; wherein each parallel convolutional neural network corresponds to an acceleration sensor set at a set position of the target structure.
[0024] In one implementation, step S13 includes:
[0025] The input data sets corresponding to each acceleration sensor are respectively input into the corresponding parallel convolutional neural network, and each parallel convolutional neural network extracts the features of the input data sets and predicts the damage type, wherein the output result of the parallel convolutional neural network is the number of data segments that meet the characteristics of different damage conditions in the N data samples of the corresponding acceleration sensor;
[0026] For each parallel convolutional neural network, the probability of meeting the characteristics of each damage condition is calculated:
[0027]
[0028] Among them, POD jrepresents the probability that the parallel convolutional neural network determines that the input data set is damage condition j, and represents the similarity between the data features of the input data set and the data features of damage condition j, where j∈[1,m], m represents the total number of damage conditions set, and n j represents the output of the parallel convolutional neural network, which is the number of data segments that meet the data characteristics of damage condition j;
[0029] The weighted damage degree is calculated based on the probability of each damage condition characteristic:
[0030]
[0031] Where D twai is the weighted damage degree, which represents the damage assessment of the target structure by the ith parallel convolutional neural network, w j represents the damage degree set for damage condition j, where w j ∈[0,1];
[0032] The weighted damage degrees corresponding to each parallel convolutional neural network are weighted averaged to obtain the overall structural health assessment information of the target structure:
[0033]
[0034] Where D represents the damage degree of the target structure; D twai represents the weighted damage degree of the ith parallel convolutional neural network to the target structure as a whole, P′ represents the total number of parallel convolutional neural networks, c i represents the reduction coefficient corresponding to the i-th parallel convolutional neural network, which represents the ability of the data of the i-th acceleration sensor to reflect the damage of the target structure, where c i ∈[0,1];
[0035] Output the structural damage identification result corresponding to the damage degree D of the target structure.
[0036] In one embodiment, the method further comprises:
[0037] S0 training damage identification model, including:
[0038] S01 Training condition design: select the target structure, set m groups of damage conditions, and obtain the acceleration data corresponding to the m damage conditions as the original training data of the model;
[0039] S02 Data Sampling and Grouping: For target structures with different damage conditions, apply excitation to the target structure and obtain the acceleration response recorded by the acceleration sensor at the set position of the target structure:
[0040]
[0041] Among them, A m represents the acceleration response data set recorded by each acceleration sensor under the mth damage condition;
[0042] U m,p Indicates the acceleration data recorded by the pth sensor in the mth data group;
[0043] According to the acceleration response obtained above, m damage condition data sets recorded for each sensor are obtained:
[0044]
[0045] Among them, Group-p represents the acceleration response data recorded by the pth sensor under each damage condition;
[0046] S03 data preprocessing: constructing training data sets according to the data sets recorded by each acceleration sensor, including: randomly intercepting the data sets recorded by each acceleration sensor, performing zero-mean normalization on the randomly intercepted data samples, using the acceleration time series after zero-mean normalization as a one-dimensional training data set, performing continuous wavelet transform on the acceleration time series after zero-mean normalization, and using the time-frequency diagram obtained after continuous wavelet transform as a two-dimensional training data set;
[0047] S04 network training: using the one-dimensional training data set and the two-dimensional training data set obtained according to the data groups recorded by each acceleration sensor as the input of the corresponding parallel convolutional neural network, and completing the training of each parallel convolutional neural network;
[0048] S05 network group: repeat steps S03 and S04 to complete the training of each parallel convolutional neural network, combine the trained p parallel convolutional neural networks into a network group, and complete the training of the damage identification model.
[0049] In the second aspect, a structural damage identification device based on a parallel convolutional neural network is proposed, which is used to implement a structural damage identification method based on a parallel convolutional neural network as shown in any one of the embodiments of the first aspect above.
[0050] The beneficial effects of the present invention are:
[0051] The present invention proposes a structural damage identification method and device based on a parallel convolutional neural network. Based on the theory of multi-domain feature fusion, a parallel neural network (P-CNN) is proposed to be applied to structural damage identification. Compared with traditional CNN, it has higher recognition performance and fitting ability, which helps to improve the recognition effect of structural damage.
[0052] The proposed damage identification model based on parallel convolutional neural network is upgraded from binary classification to multi-classification, and the concept of damage degree is defined, which makes this method have higher damage identification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.
[0054] Figure 1 is an exemplary method flow chart of a structural damage identification method based on a parallel convolutional neural network of the present invention;
[0055] Figure 2 is an exemplary method flow chart of a neural network training module of the present invention;
[0056] Figure 3 is an exemplary method flow chart of the damage identification module of the present invention;
[0057] Figure 4 A schematic diagram of sensor arrangement for the ASCE Benchmark structure shown in an embodiment of the present invention;
[0058] Figure 5 A diagram showing a damage pattern of a reference structure shown in an embodiment of the present invention;
[0059] Figure 6 A schematic diagram of a standardized vibration signal according to an embodiment of the present invention;
[0060] Figure 7 This is a schematic diagram of a time-frequency diagram of a standard wavelet as shown in an embodiment of the present invention;
[0061] Figure 8 A schematic diagram of a parallel neural network structure shown in an embodiment of the present invention;
[0062] Fig. 9 This is a schematic diagram of damage identification results shown in an embodiment of the present invention. DETAILED DESCRIPTION
[0063] The present invention is further described in conjunction with the following application scenarios.
[0064] See also Figure 1 , which shows a structural damage identification method based on a parallel convolutional neural network, the method also includes:
[0065] S0 trains a damage recognition model based on parallel convolutional neural networks, see Figure 2 , which specifically includes the following steps:
[0066] S01 Training condition design: select the target structure, set m groups of damage conditions, and obtain the acceleration data corresponding to the m damage conditions as the original training data of the model.
[0067] In one scenario, a research object is selected, such as a frame structure, equipped with p accelerometers for measuring vibration response. The first step of this module is to design m groups of damage conditions to obtain sufficient vibration data to train p parallel convolutional neural networks (P-CNN) to form a network group. As the degree of structural damage increases, the accelerometer records the vibration response of each damage condition respectively. This case is set as follows: the first group of data recorded is the undamaged state of the structure (i.e., the intact state of the structure), and the degree of damage of the structure is artificially increased and defined in turn, such as the degree of damage w1 = 10%, w2 = 20%, w3 = 30%...., until the mth group of data corresponds to the complete damage state of the structure (i.e., the damage w m =100%), thereby obtaining the acceleration data of m damage conditions as the original training data of the model.
[0068] S02 Data sampling and grouping: For the target structure under different damage conditions, excitation is applied to the target structure, and the acceleration response of the structure is recorded by p accelerometers installed on the structure; for example, under the first damage condition, the acceleration response data group recorded by p sensors is expressed as A1=(U 1,1 ,U 1,2 , ...U 1,p ).
[0069] Therefore, the acceleration response recorded by each acceleration sensor set at the set position of the target structure is obtained as follows:
[0070]
[0071] Among them, A m represents the acceleration response data set recorded by each acceleration sensor under the mth damage condition;
[0072] U m,p Indicates the acceleration data recorded by the pth sensor in the mth data group;
[0073] In order to match the subsequent algorithm based on the damage identification model, the above recorded data groups are further recombined. The m damage condition data recorded by each sensor are combined into a group (Group), and each group of data is used to train a single P-CNN in the future.
[0074] Then, according to the acceleration response obtained above, the m damage condition data sets recorded by each sensor are obtained as follows:
[0075]
[0076] Group-p represents the acceleration response data recorded by the pth sensor under each damage condition.
[0077] S03 data preprocessing: construct training data sets according to the data groups recorded by each acceleration sensor, including: randomly intercepting the data groups recorded by each acceleration sensor, performing zero-mean normalization on the randomly intercepted data samples, and using the acceleration time series after zero-mean normalization as a one-dimensional training data set; performing continuous wavelet transform on the acceleration time series after zero-mean normalization, and using the time-frequency diagram obtained after continuous wavelet transform as a two-dimensional training data set.
[0078] In one scenario, the data of Group-1 is randomly intercepted (1024 data points as a segment), and the randomly intercepted data samples are zero-mean normalized. The acceleration time series after zero-mean normalization is used as a one-dimensional data set; the acceleration time series after zero-mean normalization is subjected to continuous wavelet transform, and the time-frequency diagram obtained after continuous wavelet transform is used as a two-dimensional data set.
[0079] S04 Network training: using the one-dimensional training data set and the two-dimensional training data set obtained according to the data groups recorded by each acceleration sensor as the input of the corresponding parallel convolutional neural network, and completing the training of each parallel convolutional neural network.
[0080] In one scenario, the data set obtained in step 3 is used as two (one-dimensional and two-dimensional) inputs of the parallel convolutional neural network (P-CNN1), and the training set, validation set, and test set are divided into a ratio of 7:2:1 for model training and testing.
[0081] S05 network group: repeat steps S03 and S04 to complete the training of each parallel convolutional neural network, combine the trained p parallel convolutional neural networks into a network group, and complete the training of the damage identification model.
[0082] In one scenario, steps three and four are repeated to process the data from Group-2 to Group-p, and a total of p parallel convolutional neural networks are trained to form a network group.
[0083] S1 identifies the damage degree of the target structure in an unknown state based on the trained damage identification model based on parallel convolutional neural network, see Figure 3 , which specifically includes the following steps:
[0084] S11 data sampling: collecting acceleration data recorded by an acceleration sensor, wherein the acceleration sensor is arranged at a set position of the target structure, and the acceleration data includes a vibration response recorded by the acceleration sensor after applying excitation to the target structure.
[0085] In one scenario, the position of the sensor needs to be recorded during model training; in subsequent actual measurements, the accelerometer needs to be set at the same position as during training.
[0086] In one implementation, step S11 specifically includes:
[0087] applying an incentive to a target structure, wherein the target structure includes a framework structure;
[0088] Receive the vibration response recorded by an accelerometer set on the target structure:
[0089]
[0090] Where K p Represents the acceleration data recorded by the pth acceleration sensor, where p∈[1,P], and P represents the total number of acceleration sensors.
[0091] S12 data preprocessing: preprocess the acceleration data recorded by each acceleration sensor to obtain an acceleration data sequence and its time-frequency diagram.
[0092] In one implementation, step S12 specifically includes:
[0093] The acceleration data recorded by each acceleration sensor is randomly intercepted, and the randomly intercepted data segments are used as data samples to obtain N data samples:
[0094]
[0095] Where K p,n represents the nth segment of data randomly intercepted from the acceleration data recorded by the pth acceleration sensor, where n∈[1,N]; the length of each segment of data is set to L;
[0096] The data samples are subjected to zero-mean normalization and continuous wavelet transform to obtain the standardized time series and its wavelet time-frequency diagram.
[0097] In one scenario, the randomly intercepted sample data contains N=100 data; the length of each data segment is L=1024.
[0098] S13 Damage Identification: The input data sets corresponding to each acceleration sensor are input into the trained damage identification model based on parallel convolutional neural network. The damage identification model extracts features and predicts damage categories based on the input data sets to obtain the probability of meeting the characteristics of each damage condition. The corresponding structural damage identification results are obtained based on the probability of meeting the characteristics of each damage condition. The input data sets include time series and time-frequency diagrams.
[0099] In one embodiment, in step S13, the trained damage identification model based on parallel convolutional neural networks includes a network group composed of multiple parallel convolutional neural networks; wherein each parallel convolutional neural network corresponds to an acceleration sensor set at a set position of the target structure.
[0100] In one implementation, step S13 includes:
[0101] The input data sets corresponding to each acceleration sensor are respectively input into the corresponding parallel convolutional neural network, and each parallel convolutional neural network extracts the features of the input data sets and predicts the damage type, wherein the output result of the parallel convolutional neural network is the number of data segments that meet the characteristics of different damage conditions in the N data samples of the corresponding acceleration sensor;
[0102] For each parallel convolutional neural network, the probability of meeting the characteristics of each damage condition is calculated:
[0103]
[0104] Among them, POD j represents the probability that the parallel convolutional neural network determines that the input data set is damage condition j, and represents the similarity between the data features of the input data set and the data features of damage condition j, where j∈[1,m], m represents the total number of damage conditions set, and n j represents the output of the parallel convolutional neural network, which is the number of data segments that meet the data characteristics of damage condition j; the weighted damage degree is calculated based on the obtained probability of each damage condition:
[0105]
[0106] Where D twai represents the weighted damage degree of the i-th parallel convolutional neural network to the target structure as a whole, w j represents the damage degree set for damage condition j, where w j ∈[0,1];
[0107] The weighted damage degrees corresponding to each parallel convolutional neural network are weighted averaged to obtain the overall structural health assessment information of the target structure:
[0108]
[0109] Where D represents the damage degree of the target structure; D twai represents the weighted damage degree of the ith parallel convolutional neural network to the target structure as a whole, P′ represents the total number of parallel convolutional neural networks, c irepresents the reduction coefficient corresponding to the i-th parallel convolutional neural network, which represents the ability of the data of the i-th acceleration sensor to reflect the damage of the target structure, where c i ∈[0,1];
[0110] Output the structural damage identification result corresponding to the damage degree D of the target structure.
[0111] In addition, the present invention also proposes a structural damage identification device based on a parallel convolutional neural network, including a network training module and a damage identification module; wherein,
[0112] The network training module is used to achieve the above Figure 1 The method steps included in any one of the implementations corresponding to step S0 in the structural damage identification method based on parallel convolutional neural network;
[0113] The damage identification module is used to implement the above Figure 1 The method steps included in any one of the implementations corresponding to step S1 in the structural damage identification method based on parallel convolutional neural network;
[0114] The present invention will not be described again here.
[0115] The structural damage identification method and device based on parallel convolutional neural network proposed in the present invention have the following beneficial effects:
[0116] Based on the theory of multi-domain feature fusion, a parallel neural network (P-CNN) is proposed and applied to structural damage identification.
[0117] The network model is upgraded from binary classification to multi-classification, and the damage is quantified at the same time.
[0118] Faced with the transition from binary classification to multi-classification, the backbone network of this article uses P-CNN instead of 1D-CNN, which is an improvement made to adapt to the increased difficulty of classification tasks. The high-performance network model is the guarantee of scientific and accurate classification results.
[0119] The task is transformed from a binary classification task to a multi-classification task, and the intact-completely damaged damage degree range is refined. The damage characteristics of the unknown structural state have more references in the sample labeling space, and the output classification results are more accurate. At the same time, the concept of damage degree is introduced to quantify the degree of structural damage. On the one hand, corresponding to the multi-classification output results, the damage degree is used as the damage weight of the training condition to participate in the calculation of the damage probability. On the other hand, the damage degree will be used as a criterion for the damage identification results of the test condition to guide the training of the network.
[0120] According to the characteristics of vibration acceleration data of bridge health monitoring, an improved damage identification method based on damage probability is proposed based on damage identification method and multi-domain feature extraction concept.
[0121] The backbone network of this method uses a parallel convolutional neural network with dual feature extraction channels. The two branches extract the features of the time series signal and the wavelet time-frequency graph respectively, which are stretched and concatenated into fusion features in the aggregation layer as the input of the final classifier. Compared with the traditional CNN, it has higher recognition performance and fitting ability. The network model is multi-classification, and the concept of damage degree is defined at the same time, which makes this method have higher damage recognition accuracy.
[0122] As an exemplary description, the present invention shows that the process of modeling and structural damage identification of the ASCE-Benchmark structure based on the structural damage identification method of the parallel convolutional neural network is as follows:
[0123] Network training module
[0124] 1. Working condition design
[0125] This case uses the second phase experimental data of the ASCE health monitoring benchmark study, with a sampling frequency of 200 Hz. The experimental model is a 4-story 2×2 span steel frame structure with a plane size of 2.5m×2.5m and a floor height of 0.9m. Each floor contains 9 column components, 12 floor beam components and 8 diagonal bracing components, such as Figure 4 The model beams and columns are consolidated, and the diagonal braces are flexibly connected to the beams and columns. The diagonal braces can be freely disassembled to simulate different structural damages. 12 acceleration sensors are set between each layer, and the sensor layout is as follows: Figure 5 The random excitation is applied by an exciter placed on the top floor of the structure. The random excitation is a band-limited white noise in the southwest direction with a component between 5 and 50 Hz.
[0126] Five types of damage conditions are set for the benchmark structure (the intact state is regarded as a characteristic damage condition), as shown in Table 1. Figure 6 Five types of damage modes of the structure are shown.
[0127] Table 1 Five types of damage conditions
[0128]
[0129] 2. Data sampling, grouping and preprocessing
[0130] The exciter applies random excitation, and the accelerometer collects vibration signals under various working conditions of the structure. Data from any one of working conditions 2, 3, and 4 is selected for the damage identification module to verify the damage identification accuracy of the model, and the remaining four working condition data are used in the network training module to train the network model.
[0131] A “moving window” with a window length of 1024 is used to randomly intercept the original vibration signal. A single sensor signal window randomly intercepts 100 data samples. There are 12 sensors in total, so each working condition contains 12×100 data samples.
[0132] In the network training module, the vibration signals recorded by the same acceleration sensor of the four damage conditions are combined into one group, totaling 12 groups. The 12 groups of data will train 12 parallel neural networks to form a network group. The data used in the network training module and the damage identification module are shown in Table 2.
[0133] Table 2 Network training dataset and damage detection target dataset
[0134]
[0135]
[0136] The MATLAB code is used to generate the data set, and the four dimensions of the data set represent the number of sub-datasets, signal length, number of sensors, and number of data samples.
[0137] The working condition number is used as the data label, and the acquired data samples are subjected to zero-mean normalization and continuous wavelet transform to obtain two types of data: standardized time series and wavelet time-frequency diagram, which are used as one-dimensional data sets and two-dimensional data sets respectively, and are divided into training set, validation set and test set in the ratio of 7:2:1. Figure 6 , 7 The standardized vibration signals and wavelet time-frequency diagrams of sensor signals under five types of working conditions are shown.
[0138] 3. Network training
[0139] (1) Network Model
[0140] The parallel neural network contains two feature extraction branches: one-dimensional feature extraction branch and two-dimensional feature extraction branch. Each branch includes two convolutional layers and two pooling layers. The features extracted by the two branches are stretched and connected in series in the pooling layer as the input of the fully connected layer. Finally, the raw output data of the fully connected layer is input to the classifier. The parallel neural network structure is as follows: Figure 8 As shown, Table 3 gives the proposed network parameters.
[0141] 1. The first and second convolutional layers contain 6 and 16 convolution kernels respectively, and the convolution kernel sizes are 1×5 and 5×5 respectively;
[0142] 2. The pooling layer is used to reduce network parameters, and the downsampling factors are 1×2 and 2×2 respectively;
[0143] 3. The linear rectification function (ReLU) is used as the activation function after each pooling operation.
[0144] 4. The first layer of the fully connected layer contains a total of 120 neurons, and the second layer of the fully connected layer contains a total of 84 neurons;
[0145] 5. The final raw output data is input into the Softmax classifier.
[0146] Table 3 Network parameters
[0147]
[0148]
[0149] In each iterative training, the data batch size is 64, and the Adam optimizer is used for training optimization. The learning rate of the Adam algorithm is 0.001, and the momentum weight and decay parameters of the learning rate are selected as β1 = 0.9 and β2 = 0.99. In the process of model training, the overfitting problem is very easy to occur. We use L2 regularization and 50% random dropout neurons (Dropout) to enhance the generalization ability and robustness of the model, and use the cross entropy loss objective function to measure the quality of the model.
[0150] (2) Evaluation indicators
[0151] In the statistical analysis of multiple classifications of deep learning, accuracy, precision, recall and F1 score are often used to comprehensively analyze the performance of the model, where the F1 score is the harmonic mean of accuracy and recall. Accuracy represents the proportion of correct classifications in all damage data. Generally speaking, accuracy is often used as an overall evaluation of multi-classification algorithms; the F1 score indicator combines the output results of precision and recall, and the value range is [0, 1]. The higher the F1 score, the better the trained model is. It is often used to evaluate the recognition performance of a multi-classification algorithm for a certain category. In this case, accuracy and F1 score are used as performance indicators for overall damage recognition of the model. The specific calculation formula is as follows:
[0152]
[0153] in: Indicates accuracy;
[0154] and y represent the predicted label and the true label
[0155] TP stands for “true positive” (the number of samples whose object class is correctly classified into the object class);
[0156] FP stands for “false positive” (the number of samples of other classes that are misclassified into the object class);
[0157] TN stands for “true negative” (the number of samples of other classes that are correctly classified into other classes);
[0158] FN stands for “false negatives” (the number of samples where an object class is misclassified into other classes);
[0159] (3) Training Results
[0160] In order to describe the performance of the classification model on the test set, Table 4 shows the accuracy of the network groups trained in each group of experiments.
[0161] Table 4 Network group accuracy of each group of experiments
[0162]
[0163] Damage Identification Module
[0164] 1. Data sampling and preprocessing
[0165] The damage identification module data is shown in Table 2.
[0166] 2. Damage identification
[0167] In order to verify the performance of the damage identification model, one or more damage states of the structure are used as test conditions. The acceleration data samples of the test conditions of the three groups of experiments are respectively input into the trained network group to identify the damage degree of the test conditions. Table 5 gives the output results of the experiment.
[0168] For each parallel convolutional neural network, the probability of meeting the characteristics of each damage condition is calculated:
[0169]
[0170] Among them, POD j represents the probability that the parallel convolutional neural network determines that the test condition is the damage condition j, and represents the similarity between the test condition and the damage condition j, where j∈[1,m], m represents the total number of damage conditions set, and n j It represents the output result of the parallel convolutional neural network, which is the number of data segments that meet the data characteristics of damage condition j; the test condition is an unknown damage state of the target structure, which is used to verify the performance of the damage identification model.
[0171] The weighted damage degree is calculated based on the obtained probability of each damage condition:
[0172]
[0173] Where D twai represents the weighted damage degree of the i-th parallel convolutional neural network to the target object as a whole, w jrepresents the damage degree set for damage condition j, where w j ∈[0,1];
[0174] The weighted damage degrees corresponding to each parallel convolutional neural network are weighted averaged to obtain the overall structural health assessment information of the target object:
[0175]
[0176] Where D represents the damage degree of the target object; D twai represents the weighted damage degree of the ith parallel convolutional neural network to the target object as a whole, P′ represents the total number of parallel convolutional neural networks, c i represents the reduction coefficient corresponding to the i-th parallel convolutional neural network, which represents the ability of the data of the i-th acceleration sensor to reflect the damage of the target object, where c i ∈[0,1];
[0177] Output the structural damage identification result corresponding to the damage degree D of the target structure.
[0178] Table 5 Damage identification output results
[0179]
[0180] Fig. 9 The final results and reference values of the proposed method are shown. It can be seen that when identifying various damage conditions, the output damage degree of the proposed method is extremely close to the defined real damage of the structure, and it has broad application prospects in the field of structural damage identification.
[0181] It should be noted that each functional unit / module in each embodiment of the present invention may be integrated into one processing unit / module, or each unit / module may exist physically separately, or two or more units / modules may be integrated into one unit / module. The above-mentioned integrated unit / module may be implemented in the form of hardware or in the form of software functional unit / module.
[0182] Through the description of the above implementation mode, it can be clearly understood by those skilled in the art that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code or any appropriate combination thereof. For hardware implementation, the processor can be implemented in one or more of the following units: application specific integrated circuit (ASIC), digital signal processor (DSP), digital signal processing device (DSPD), programmable logic device (PLD), field programmable gate array (FPGA), processor, controller, microcontroller, microprocessor, other electronic units designed to implement the functions described herein or their combination. For software implementation, part or all of the process of the embodiment can be completed by instructing the relevant hardware through a computer program. When implemented, the above program can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein the communication medium includes any medium that facilitates the transmission of a computer program from one place to another. The storage medium can be any available medium that can be accessed by a computer. Computer-readable media can include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer.
[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should analyze that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.
Claims
1. A structural damage identification method based on parallel convolutional neural network, characterized in that: include: S11 data sampling: collecting acceleration data recorded by an acceleration sensor, wherein the acceleration sensor is arranged at a set position of the target structure, and the acceleration data includes a vibration response recorded by the acceleration sensor after applying an excitation to the target structure; S12 data preprocessing: preprocess the acceleration data recorded by each acceleration sensor to obtain an acceleration data sequence and its time-frequency diagram; S13 Damage Identification: Input the input data sets corresponding to each acceleration sensor into the trained damage identification model based on parallel convolutional neural network. The damage identification model extracts features and predicts damage categories of the input data sets to obtain the probability of meeting the data characteristics of each damage condition. The corresponding structural damage identification results are obtained based on the probability of meeting the data characteristics of each damage condition. The input data sets include acceleration time series and its time-frequency diagram. The trained damage identification model based on parallel convolutional neural network includes a network group composed of multiple parallel convolutional neural networks; each parallel convolutional neural network corresponds to an acceleration sensor set at a set position of the target structure; the parallel convolutional neural network includes two feature extraction branches: a one-dimensional feature extraction branch and a two-dimensional feature extraction branch, the two branches extract the features of the time series signal and the wavelet time-frequency graph respectively, and each branch includes two convolution layers and two pooling layers; the features extracted by the two branches are stretched and connected in series in the pooling layer as the input of the fully connected layer.
2. The structural damage identification method based on parallel convolutional neural network according to claim 1 is characterized in that: Step S11 specifically includes: applying an incentive to a target structure, wherein the target structure includes a framework structure; Receive the vibration response recorded by an accelerometer set on the target structure: K: Where K p Represents the acceleration data recorded by the pth acceleration sensor, where p∈[1,P], and P represents the total number of acceleration sensors.
3. The structural damage identification method based on parallel convolutional neural network according to claim 2 is characterized in that: Step S12 specifically includes: The acceleration data recorded by each acceleration sensor is randomly intercepted, and a fixed-length data segment is intercepted as a data sample, and N segments of data are obtained to form a data sample: Where K p,n represents the nth segment of data randomly intercepted from the acceleration data recorded by the pth acceleration sensor, where n∈[1,N]; the length of each segment of data is set to L; The data samples are subjected to zero-mean normalization and continuous wavelet transform to obtain the standardized time series and its wavelet time-frequency diagram.
4. The structural damage identification method based on parallel convolutional neural network according to claim 1 is characterized in that: Step S13 includes: The input data sets corresponding to each acceleration sensor are respectively input into the corresponding parallel convolutional neural network, and each parallel convolutional neural network extracts the features of the input data sets and predicts the damage type, wherein the output result of the parallel convolutional neural network is the number of data segments that meet the characteristics of different damage conditions in the N data samples of the corresponding acceleration sensor; For each parallel convolutional neural network, the probability of meeting the characteristics of each damage condition is calculated: Among them, POD j represents the probability that the parallel convolutional neural network determines that the input data set is damage condition j, and represents the similarity between the data features of the input data set and the data features of damage condition j, where j∈[1,m], m represents the total number of damage conditions set, and n j represents the output of the parallel convolutional neural network, which is the number of data segments that meet the data characteristics of damage condition j; the weighted damage degree is calculated based on the probability of each damage condition characteristic obtained: Where D twai is the weighted damage degree, which represents the damage assessment of the target structure by the ith parallel convolutional neural network, w j represents the damage degree set for damage condition j, where w j ∈[0,1]; The weighted damage degrees corresponding to each parallel convolutional neural network are weighted averaged to obtain the overall structural health assessment information of the target structure: Where D represents the damage degree of the target structure; D twai represents the weighted damage degree of the i-th parallel convolutional neural network to the target structure as a whole, P ′ represents the total number of parallel convolutional neural networks, c i represents the reduction coefficient corresponding to the i-th parallel convolutional neural network, which represents the ability of the data of the i-th acceleration sensor to reflect the damage of the target structure, where c i ∈[0,1]; Output the structural damage identification result corresponding to the damage degree D of the target structure.
5. The structural damage identification method based on parallel convolutional neural network according to claim 1 is characterized in that: The method further includes: S0 training damage identification model, including: S01 training condition design: selecting the target structure, setting m groups of damage conditions, and obtaining acceleration data corresponding to the m damage conditions as original training data of the model; S02 Data Sampling and Grouping: For target structures with different damage conditions, apply excitation to the target structure and obtain the acceleration response recorded by the acceleration sensor at the set position of the target structure: Among them, A m represents the acceleration response data set recorded by each acceleration sensor under the mth damage condition; U m,p Indicates the acceleration data recorded by the pth sensor in the mth data group; According to the acceleration response obtained above, m damage condition data sets recorded for each sensor are obtained: Wherein, Group-p represents the acceleration response data recorded by the pth sensor under each damage condition; S03 data preprocessing: constructing training data sets according to the data sets recorded by each acceleration sensor, including: randomly intercepting the data sets recorded by each acceleration sensor, performing zero-mean normalization on the randomly intercepted data samples, using the acceleration time series after zero-mean normalization as a one-dimensional training data set, performing continuous wavelet transform on the acceleration time series after zero-mean normalization, and using the time-frequency diagram obtained after continuous wavelet transform as a two-dimensional training data set; S04 network training: using the one-dimensional training data set and the two-dimensional training data set obtained according to the data groups recorded by each acceleration sensor as the input of the corresponding parallel convolutional neural network, and completing the training of each parallel convolutional neural network; S05 network group: repeat steps S03 and S04 to complete the training of each parallel convolutional neural network, combine the trained p parallel convolutional neural networks into a network group, and complete the training of the damage identification model.
6. A structural damage identification device based on parallel convolutional neural network, characterized in that: The device is used to implement the structural damage identification method based on a parallel convolutional neural network as described in any one of claims 1 to 5.