Strong motion acceleration record vertical component identification method based on twin neural network

Abstract

The application discloses a strong motion acceleration record vertical component identification method based on a twin neural network. Firstly, the cross-correlation coefficient of the acceleration time history and the cross-correlation coefficient of the Fourier spectrum between the three components of each record in the database are calculated and analyzed. Then, the twin neural network is used to learn the acceleration characteristics of each combination, the residual network is used to learn the Fourier spectrum characteristics of each combination, and artificial judgment is combined. The method is realized by comparing the similarity of the acceleration waveform and the FAS between the three components of a group of seismic motion records, and quite high accuracy is achieved. Normal and marked error records in the test set can be accurately identified. Compared with the identification by using the traditional artificial method alone, the efficiency is significantly improved while the identification accuracy is ensured.

Description

TECHNICAL FIELD

[0001] The application relates to a strong motion acceleration record vertical component identification method based on a twin neural network. BACKGROUND

[0002] High-quality strong motion records are the data basis for seismic engineering and engineering earthquake research, which is of great significance for site effect analysis, ground motion attenuation relationship research, structural seismic response input, and earthquake warning research. However, due to the involvement of many links in the operation, maintenance and management of the strong motion network, especially the relatively insufficient management experience in the early stage of network construction, some human errors are inevitable. Therefore, for the increasingly dense strong motion network, how to ensure the quality of strong motion data is an urgent and challenging task, and the quality monitoring of massive strong motion data sets is particularly important.

[0003] Data quality problems include S-wave triggering, spikes, multiple events, etc. Some strong motion data processing programs include quality control processes and identify specific quality problems. The direction problem is part of the metadata problem, which is usually manifested as direction error or missing. Previous studies have developed various direction checking methods. For example, the sensor direction can be determined by polarization analysis of seismic signals, or the direction can be estimated by analyzing the polarization of Rayleigh waves retrieved from environmental noise cross-correlation. In addition, some researchers have corrected the direction of the seismometer by comparing the waveform consistency between different types of co-located sensors. These methods are mainly used to determine the northward offset of the seismometer. In fact, during the quality evaluation of strong motion records, it was found that some records had the phenomenon of interchanging three-component channel data, i.e., mistakenly recognizing the vertical component of a set of records as the horizontal component, which may be caused by inconsistent instrument output standards or mistakes during post-processing. This is an abnormal data phenomenon because the vertical component of most records has a correct channel label.

[0004] The vertical component of the strong motion record is crucial in the process of structural design and the application of the H / V spectral ratio method. Therefore, when conducting seismic analysis and seismic design method research, the direction of the strong motion record must be accurate to avoid errors caused by the direction problem of the strong motion record. In most cases, the acceleration waveform and Fourier amplitude spectrum (FAS) of the east-west (EW) and north-south (NS) two horizontal components are more similar to the vertical component. Traditionally, the vertical component is identified in a set of strong motion records by manually comparing the acceleration waveform and Fourier spectrum shape of the three components, but this method is low in efficiency and impractical when facing a large number of records. Therefore, it is necessary to find a fast and accurate method to identify the vertical strong motion. SUMMARY

[0005] In order to solve the above technical defects, the application provides a strong earthquake acceleration record vertical component identification method based on a twin neural network, which is fast and has high accuracy.

[0006] The application is realized by the following measures:

[0007] A strong earthquake acceleration record vertical component identification method based on a twin neural network comprises the following steps:

[0008] Step 1: In the strong earthquake record database, select several groups of strong earthquake data with high accuracy as samples, each group of strong earthquake data contains a horizontal component H1, a horizontal component H2 and a vertical component UD, combine and mark two components in each group of strong earthquake data as combined H1-H2, combined H1-UD and combined H2-UD, and then use the cross-correlation coefficient to calculate the acceleration waveform similarity and the Fourier spectrum similarity of the two components in each combination, and conclude that the similarity between the two components in the combined H1-H2 is higher than that between the two components in the other two combinations;

[0009] Step 2: input the acceleration characteristic data of each combination in each group of strong earthquake data in step 1 into the twin neural network for deep learning and training to obtain a network model based on the twin neural network;

[0010] Step 3: draw the acceleration time history curve of the channel where the three components in each group of strong earthquake data to be judged are located, then combine the three components two by two according to the mode in step 1, and input them one by one into the network model obtained in step 2 for comparison to obtain the time domain similarity T sim of each combination;

[0011] Step 4: perform Fourier transform on the acceleration data of the channel where the three components in each group of strong earthquake data in step 3 are located and perform smoothing processing to obtain the smoothed fast Fourier transform data FAS, then combine the three components two by two according to the mode in step 1, and input them one by one into the network model obtained in step 2 for comparison to obtain the frequency domain similarity F sim of each combination;

[0012] Step 5: according to the time domain similarity T sim and the frequency domain similarity F sim obtained in steps 3 and 4, compare whether the similarity between the two components in the combined H1-H2 of a group of strong earthquake data is higher than that between the two components in the other two combinations, if yes, judge that the group of strong earthquake data is correct, if not, judge that the channel where the vertical component of the group of strong earthquake data is located has label error, and proceed to step 6;

[0013] Step 6: Combined with the detailed parameters of ground motion, further artificial discrimination is carried out, and finally the ground motion data with label error of vertical component is obtained.

[0014] The twin neural network will generate a multi-dimensional feature vector after feature extraction of the input content. In order to input the multi-dimensional feature vector into the subsequent layer, the Flatten layer is used to perform one-dimensional processing on the vector, and then the Manhattan distance PDeltaXP1 of the two one-dimensional vectors x1 and x2 is calculated in the comparison layer. The formula is as follows, and the abs in the formula is the absolute value:

[0015] PDeltaXP1=abs(x1-x2) (1)

[0016] Then, the distance is input into the full connection layer for processing, finally, the output is connected to a neuron, and the Sigmoid activation function is applied to the neuron as the output layer, and the result is converted to the range of [0, 1] to represent the similarity between the two input elements. If they are similar, the output of this function will be close to 1; if they are not similar, the output value will be close to 0.

[0017] In step 2, the training environment of the network model is composed of Python 3.9, and the deep learning framework TensorFlow 2.10 and Keras 2.10 are installed. The model is trained by using the Adam optimizer through back propagation, the bitch_size size is 64, the learning rate is 0.01, dropout and early_stopping are used in the training process, and the best model weight is stored according to the validation loss to prevent overfitting. The dropout rate is 0.5, and the binary cross entropy is used as the loss function, and the formula is as follows:

[0018]

[0019] In the formula, m is the total number of samples, y i is a binary label 0 or 1, and p(y i ) is the probability of the output label y i .

[0020] The beneficial effects of the present application are: the method is realized by comparing the similarity of acceleration waveforms and FAS between three components of a group of seismic records, and a relatively high accuracy is achieved. By using two methods for common identification, normal and labeled error records in the test set can be accurately identified. In addition to artificial identification, it is proved that the method can effectively identify the vertical component in a group of seismic records. Compared with the identification method using traditional artificial method alone, the efficiency is significantly improved while ensuring the identification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 Cumulative distribution of cross-correlation coefficients of acceleration waveforms;

[0022] Figure 2 Distribution of cross-correlation coefficients of acceleration waveforms of different combinations;

[0023] Figure 3 Cumulative distribution of cross-correlation coefficients of acceleration Fourier spectra;

[0024] Figure 4 Distribution of cross-correlation coefficients of acceleration Fourier spectra;

[0025] Figure 5 In which a represents the acceleration time history graph, and b represents the Fourier spectrum graph;

[0026] Figure 6 In which a represents the acceleration time history graph recorded by RSN1189, and b represents the FAS graph;

[0027] Figure 7 In which a represents the acceleration time history graph, and b represents the Fourier spectrum graph; DETAILED DESCRIPTION

[0028] The application will be described in further detail below with reference to the accompanying drawings:

[0029] The application discloses a seismic record vertical component channel discrimination method based on deep learning to realize efficient and rapid discrimination. Firstly, the cross-correlation coefficients of acceleration time histories and the cross-correlation coefficients of Fourier spectra between three components of each record in the NGA-West2 database are calculated and analyzed. Then, a twin neural network is used to learn the acceleration features of each combination, a residual network is used to learn the Fourier spectrum features of each combination, and a method process combined with artificial further judgment is introduced in the third section. Then, the model training and verification, and the method results. Finally, the seismic record vertical component discrimination method is applied to Chinese strong motion records to prove the effectiveness of the method in identifying vertical component channel errors. The specific steps are as follows:

[0030] Step 1, select a number of strong motion data with high accuracy in the strong motion record database as samples, each of which contains horizontal component H1, horizontal component H2 and vertical component UD, combine two components in each strong motion data and mark them as combination H1-H2, combination H1-UD and combination H2-UD, then use cross-correlation coefficients to calculate the acceleration waveform similarity and Fourier spectrum similarity of two components in each combination, and conclude that the similarity between two components in combination H1-H2 is higher than that between two components in the other two combinations;

[0031] Step 2, input the acceleration characteristic data of each group of strong motion data in each group of strong motion data in step 1 into the twin neural network for deep learning and training to obtain a network model based on the twin neural network;

[0032] Step 3, draw the acceleration time history curve of the channel where the three components of each group of strong motion data to be judged are located, then combine the three components in pairs according to the mode in step 1, and input them into the network model obtained in step 2 one by one for comparison to obtain the time domain similarity T of each combination sim ;

[0033] Step 4, Fourier transform and smooth the acceleration data of the channel where the three components of each group of strong motion data in step 3 are located to obtain the smoothed fast Fourier transform data FAS, then combine the three components in pairs according to the mode in step 1, and input them into the network model obtained in step 2 one by one for comparison to obtain the frequency domain similarity F of each combination sim ;

[0034] Step 5, according to the time domain similarity T sim and the frequency domain similarity F sim obtained in steps 3 and 4, compare whether the similarity between the two components in the combination H1-H2 of a group of strong motion data is higher than that between the two components in the remaining two combinations, if yes, it is judged that the group of strong motion data is correct, if not, it is judged that the vertical component channel of the group of strong motion data exists label error, and step 6 is performed;

[0035] Step 6, combined with the detailed parameters of ground motion, further artificial discrimination is carried out, and finally the ground motion data with vertical component label error is obtained.

[0036] The object of the application is to distinguish the vertical component from the three components of a group of ground motion records. Since the correlation between the two horizontal components is usually stronger, we first analyze the cross-correlation between all three components. This is achieved by combining the components in each group of records in pairs, where "H1-H2" represents the combination of two horizontal components, "H1-UD" and "H2-UD" represent the combination of two horizontal components and the vertical component, respectively.

[0037] The NGA-West2 database was developed by the Pacific Earthquake Engineering Research Center (PEER) to extend the NGA-West1 database and improve the quality and accuracy of ground motion prediction equations. The database has been carefully checked and the accuracy of the ground motion records in the database is well recognized. The records in the NGA-West2 database with different number of data points in three components and the records with missing vertical component were removed. Finally, 18447 sets of strong motion records were selected and used. The cross-correlation coefficient was used to calculate the correlation of acceleration waveforms and Fourier spectra of each set of combinations. Since the components of a set of records are not exactly the same, the cross-correlation coefficient p xy (τ) will not reach 1.

[0038] First, the cross-correlation coefficients of acceleration waveforms of all combinations were calculated and statistically analyzed. Figure 1 The cumulative distribution of cross-correlation coefficients is given. It can be seen that the cross-correlation coefficients between two horizontal components (H1-H2) are generally larger than those between horizontal and vertical components. The expected values of cross-correlation coefficients of three combinations H1-H2, H1-UD and H2-UD are 0.2254, 0.1961 and 0.1968, respectively, indicating that the correlation between two horizontal components is higher. The similarity of the cumulative distribution curves of H1-UD and H2-UD combinations further indicates that the correlation between horizontal components is higher.

[0039] Figure 2 The scatter plots of cross-correlation coefficients between different combinations are given, in which the blue dashed line is the fitting straight line of H1-UD combination and the black dashed line is the fitting straight line of H2-UD combination. It can be seen from the figure that the slope of the fitting straight line is obviously smaller than 1, indicating that the cross-correlation coefficients of H1-UD and H2-UD combinations are mostly lower than that of H1-H2 combination. However, among the 18447 records, only 51.5% of the records have the largest cross-correlation coefficient in H1-H2 combination, which is a relatively low proportion.

[0040] In addition, the cross-correlation coefficients between FAS of all combinations were calculated, Figure 3 The cumulative distribution of cross-correlation coefficients is given. It can be seen that in the overall trend, the correlation between two horizontal components (H1-H2) is always higher than that between horizontal and vertical components. In addition, the correlation between these Fourier spectra (expected values are 0.8062 (H1-H2), 0.7744 (H1-UD) and 0.7761 (H2-UD)) is significantly larger than that obtained from acceleration waveforms. Figure 4The scatter plot of the cross-correlation coefficients of different combinations shows that the correlation between the two horizontal components (H1-H2) is higher than that of the other two combinations. Among the 18447 records, 71.1% of the records have the largest cross-correlation coefficient in the H1-H2 combination. This result indirectly indicates that the correlation between FAS of each combination is stronger than that between the acceleration waveforms.

[0041] Nevertheless, there are still a large number of samples above the y=x line, indicating this trend. When considering both the acceleration waveforms and the FAS, 78.92% of the records have the largest cross-correlation coefficient in the H1-H2 combination. This indicates that it is difficult to identify the vertical component using only the cross-correlation coefficient. In Figure 2 In the upper left corner, the points represent records with relatively high correlation between H1 and the vertical component. Preliminary analysis of these records shows that the vertical component of these records may be mislabeled channels, Figure 5 A set of example records (RSN9325) is shown in

[0042] For the identification of the vertical component of ground motion records, the traditional method generally compares the similarity between the waveforms and the three components of the Fourier spectrum by manual comparison. The traditional method is feasible when identifying a small number of records. However, with the rapid accumulation of ground motion records, when faced with a large number of records, the traditional method requires a large amount of time and effort to accurately identify the channel of the vertical component. Therefore, this paper proposes a method that uses deep learning to preliminarily judge a large number of ground motion records, and then combines manual identification of the vertical component, aiming to improve the accuracy and efficiency of identification, while avoiding the subjectivity of manual identification and the standardization problem when multiple people cooperate. This method includes two main parts, respectively a Siamese neural network for comparing the similarity of acceleration waveforms and another Siamese neural network for comparing the similarity of FAS.

[0043] Siamese neural network (Siamese network) is a "twin network" mainly used for tasks that require comparing two similar inputs. The "twin" of the network refers to two identical subnetworks that share the same parameters and are used to process two independent inputs. After the input content is processed by the feature extraction network, a multi-dimensional feature vector is generated. In order to input this multi-dimensional feature vector into the subsequent layer, the Flatten layer is used to process the vector into a one-dimensional vector, and then the Manhattan distance PΔxP1 between the two one-dimensional vectors x1 and x2 is calculated in the comparison layer:

[0044] PΔxP1 = abs(x1-x2) (3)

[0045] The distance is then input into a fully connected layer for processing, and finally, the output is connected to a neuron to which a sigmoid activation function is applied as an output layer to convert the result to the range [0, 1] to represent the similarity between the two input elements. If they are more similar, the output of this function will be close to 1; if they are not similar, the output value will be close to 0. Residual neural networks (ResNet) have a structure that can quickly accelerate the training of neural networks and significantly improve the accuracy of models. Given the high performance of ResNet18 in processing images with simple features and the simple features input in this study, the main part of ResNet18 is used for image feature extraction in this paper.

[0046] The seismic record vertical component identification method provided by the application is developed on the basis of the traditional identification method, and the method uses deep learning technology to replace part of the artificial work in the acceleration waveform comparison and Fourier spectrum comparison in the traditional identification method. The specific steps of the method are as follows:

[0047] 1. Draw the acceleration time history curve of the three channels according to the original data of the seismic record, then combine them in pairs, and input them into the twin neural network for training and comparison, so as to obtain the similarity T of each combination sim (time domain similarity).

[0048] 2. Fourier transform and smooth the acceleration data of the three channels of the seismic record, then input the smoothed FAS into the twin neural network for training and comparison, and obtain the F of each combination sim (frequency domain similarity).

[0049] 3. According to the size of T sim and F sim of each combination, the correctness of the vertical component channel of a group of records is judged. For a group of normal records, the waveform correlation coefficient and the Fourier spectrum correlation coefficient should be H1-H2 combination compared with other combinations, if not, it is preliminarily judged that the vertical component channel of the record may have label error.

[0050] 4. For the record preliminarily judged as the suspected vertical component label error, combined with the detailed information of the seismic motion (epicentral distance, fault distance, PGA and other information), further artificial discrimination is carried out, and finally the possible vertical component label error seismic record is obtained

[0051] The 18447 sets of records in the NGA-West2 database were combined pairwise, a total of 55341 combinations, and each combination was labeled to obtain a sample data set, where H1-H2 label is 1, H1-UD and H2-UD label is 0, that is, the combination of two horizontal components is a positive sample, and other combinations are negative samples. In addition, 14983 sets of records with significant differences between the vertical component and the horizontal component of the waveform and FAS were carefully screened by manual visual inspection, and then randomly divided into 60% as the training set, 20% as the validation set and 20% as the test set.

[0052] The training environment consists of Python 3.9, which installs the deep learning framework TensorFlow 2.10

[38] and Keras2.10. The model is trained using the Adam optimizer through backpropagation, with a batch_size size of 64 and a learning rate of 0.01. Dropout and early stopping are used during training, and the best model weights are stored according to the validation loss to prevent overfitting (and reduce training time), with a dropout rate of 0.5. Binary cross-entropy is used as the loss function, as shown in Equation 4.

[0053]

[0054] where m is the total number of samples, y i is a binary label of 0 or 1, p(y i ) is the probability of the output label y i .

[0055] The established sample data set is used as input for the training and validation of the Siamese neural network. The training model that achieves the lowest loss in the validation set is then applied to evaluate the test set. For the three combinations in a set of seismic records, the three similarity values output by the model are compared. The combination with the highest similarity is designated as the H1-H2 combination. To evaluate the model's ability to identify channel label error records, 50% (1500 sets) of records in the test set are randomly selected, and their vertical component labels are exchanged with their horizontal component labels to intentionally create label error records. This label exchange operation is repeated 100 times, and the average of the individual and combined identification of acceleration waveform and FAS similarity is used respectively.

[0056] For normal samples, the accuracy rates of acceleration waveform, FAS and common identification are 98.7%, 95.6% and 94.3% respectively, while for the created label error samples, the accuracy rates of the three cases are 99.3%, 97.9% and 100% respectively. The trained Siamese neural network achieves high accuracy on the test set, and the intentionally created label error records are also accurately identified.

[0057] In fact, through the identification of the training model and further meticulous manual identification of the 18447 sets of records in the NGA-West2 database, it was found that 14 sets of seismic records showed significantly higher similarity between the vertical component and one of the two horizontal components (excluding records with PGA less than 1 gal). This indicates possible labeling errors in the three components of these records, and a set of example records (RSN11341) is shown as Figure 6

[0058] Since the operation of the China Digital Strong Motion Seismic Network, a large number of strong earthquake records with engineering application value have been obtained. Over the past decade, the strong earthquake network has gradually improved. With the construction of the "National Intensity Rapid Reporting and Early Warning Project", the number and coverage of China's strong earthquake observation stations have increased year by year, and a large amount of valuable data has been accumulated. The proposed vertical component identification method for seismic records is used to analyze and discuss the channel correctness of the collected Chinese strong earthquake records.

[0059] High-quality records (each signal window is clear and distinguishable) were selected from Chinese strong earthquake records for vertical component identification, thus avoiding errors caused by poor records. Through careful manual verification, a total of 3381 sets of strong ground motion records were used.

[0060] The trained model was used to evaluate the accuracy of the vertical component labels of the 3381 sets of strong earthquake records. Table 1 shows the results of separate and joint judgments using acceleration waveforms and FAS. Further manual identification was performed on the preliminarily selected seismic records, and finally 308 sets of suspected channel error seismic records were obtained. This is consistent with the manual identification results of the 3381 sets of records. This is consistent in all 3381 sets of records. The vertical component has been confused from the acceleration waveform and FAS. The 308 sets of suspected vertical component label error records in Chinese strong earthquake records are distributed in 170 stations in 12 provinces. The channel direction of the record appears to be incorrect, and the main reason may be that the station is problematic. The stations of these records were analyzed.

[0061] Table 1 Identification results of suspected channel label errors in Chinese strong earthquake records

[0062]

[0063] There are 28 stations with more than 3 channel error records at the same station and the total error rate is greater than 50%. Table 2 provides the codes of these stations.

[0064] Table 2 Number of channel label error records in 28 stations

[0065]

[0066] ​

[0067] It can be seen that the vertical component channel error rate of 16 stations reaches 100%, especially 7 records in 053LDS station are all judged as channel error, which may be due to the problems in channel connection, output setting, etc. Therefore, the station related staff can use the method in this paper to screen the stations that may have problems.

[0068] The method is realized by comparing the similarity of acceleration waveforms and FAS between three components of a group of ground motion records. First, using the NGA-West2 database, the three components of each group of ground motion records are combined two by two, respectively marked as H1-H2, H1-UD and H2-UD. Then, the correlation analysis of the acceleration time history curves and Fourier spectrum of each combination of each group of records is carried out using the cross-correlation coefficient. Subsequently, the acceleration time history curves and the image of the smoothed FAS of each combination are input into the twin neural network for training respectively. According to the discrimination results of acceleration waveforms and FAS, combined with artificial identification, the channel of the vertical component is determined. Finally, this method is applied to the Chinese strong motion records (2007-2018), and the main conclusions are as follows:

[0069] 1. The correlation analysis of the records in the NGA-West2 database shows that the correlation of the H1-H2 combination of the two horizontal components is generally high, but the proportion of the maximum cross-correlation coefficient of the H1-H2 combination is only 51.5% (for acceleration waveforms) and 71.1% (for Fourier spectrum), respectively. This shows that the two horizontal components are generally correlated, but it is difficult to judge whether the combination is H1-H2 combination by this standard.

[0070] 2. The vertical component identification method proposed in this paper achieves a relatively high accuracy. Although some records are incorrectly identified as channel label errors when only using acceleration waveforms or FAS for identification, the intentionally created vertical component label error records are accurately identified in the test set. By using both methods for co-identification, normal and labeled error records in the test set can be accurately identified. These findings, combined with artificial identification, confirm that the method can effectively identify the vertical component in a group of ground motion records. Using the method proposed in this paper, it is found that there may be vertical component label errors in 14 groups of ground motion records in NGA-West2.

[0071] The vertical component identification method of seismic records proposed in the present application identifies the Chinese strong motion records, and finds that the labels of 308 records of 170 stations are suspected to be wrong in the vertical component channel, which is basically consistent with the result obtained by simply using the traditional manual method, thereby verifying the effectiveness of the proposed method. The method proposed in this paper significantly improves the efficiency while ensuring the accuracy of identification. In addition, among the records with suspected vertical component label errors, there are 28 stations with more than 3 suspected channel error records and an error rate of more than 50%. Station staff can use the method proposed in this paper to quickly find the stations that may have problems and check these stations.

[0072] The above only describes the preferred embodiments of the patent, and it should be noted that for ordinary technical personnel in the technical field, several improvements and replacements can be made without departing from the technical principles of the patent, and these improvements and replacements should also be considered as the protection scope of the patent.

Claims

1. A strong motion acceleration record vertical component identification method based on twin neural network, characterized in that, Comprise the following steps: Step 1, select several groups of strong motion data with high accuracy from the strong motion record database as samples, each group of strong motion data contains horizontal component H1, horizontal component H2 and vertical component UD, combine two components in each group of strong motion data and mark as combination H1-H2, combination H1-UD and combination H2-UD, then use cross-correlation coefficient to calculate the similarity of acceleration waveform and the similarity of Fourier spectrum of two components in each combination, and conclude that the similarity between two components in combination H1-H2 is higher than that between two components in the other two combinations; Step 2, input the acceleration characteristic data of each combination in each group of strong motion data in step 1 into the Siamese neural network for deep learning and training to obtain a network model based on Siamese neural network; Step 3, draw the acceleration time history curve of the channel where the three components of each group of strong motion data to be judged are located, then combine the three components in pairs according to the manner in step 1, and input them into the network model obtained in step 2 for comparison one group after another to obtain the time domain similarity T of each combination sim ; Step 4, Fourier transform and smooth the acceleration data of the channels where the three components in each group of strong motion data in step 3 are located to obtain the smoothed fast Fourier transform data FAS, then combine the three components two by two in the manner in step 1, and input them into the network model obtained in step 2 one group after another for comparison to obtain the frequency domain similarity F of each combination sim ; Step 5, comparing the similarity between two components in the combination H1-H2 of a certain group of strong motion data with the similarity between two components in the remaining two combinations, if the former is higher than the latter, it is judged that the group of strong motion data is correct, if not, it is judged that the vertical component of the group of strong motion data is in a channel with label error, and step 6 is performed. sim and the frequency domain similarity F sim , comparing the similarity between two components in the combination H1-H2 of a certain group of strong motion data with the similarity between two components in the remaining two combinations, if the former is higher than the latter, it is judged that the group of strong motion data is correct, if not, it is judged that the vertical component of the group of strong motion data is in a channel with label error, and step 6 is performed. Step 6, combine the ground motion detailed parameters to further artificially distinguish, and finally obtain the ground motion data with label error of vertical component.

2. The method according to claim 1, wherein the method is characterized by: The Siamese neural network will produce a multi-dimensional feature vector after feature extraction of the input content. In order to input this multi-dimensional feature vector into the subsequent layer, the Flatten layer is used to process the vector into one-dimensional, and then the Manhattan distance PDeltaXP1 of the two one-dimensional vectors x1 and x2 is calculated in the comparison layer, as follows: abs in the formula is the absolute value: PDeltaXP1 = abs(x1-x2) (1) Then, input the distance into the fully connected layer for processing, finally, connect the output to a neuron, apply the Sigmoid activation function to the neuron as the output layer, and convert the result to the [0,1] range to represent the similarity between the two input elements, if they are similar, the output of this function will be close to 1; if they are not similar, the output value will be close to 0.

3. The method according to claim 1, wherein the method is characterized by: In step 2, the training environment of the network model is composed of Python 3.9, which is installed with deep learning framework TensorFlow2.10 and Keras 2.10, the model uses Adam optimizer to train through back propagation, the bitch_size size is 64, the learning rate is 0.01, dropout and early stopping are used in the training process, and the best model weight is stored according to the validation loss to prevent overfitting, the dropout rate is 0.5, and binary cross entropy is used as the loss function, as follows: where m is the total number of samples, y i is a binary label 0 or 1, p(y i ) is the probability of output label y i .

Images