Strong vibration acceleration record vertical component identification method based on twinning neural network

Abstract

The invention discloses a strong vibration acceleration record vertical component identification method based on a twin neural network. Firstly, a cross correlation coefficient of an acceleration time history and a cross correlation coefficient of a Fourier spectrum among three components recorded in each group of a database are calculated and analyzed; and then a twin neural network is adopted to learn acceleration characteristics of each combination, a residual network is adopted to learn Fourier spectrum characteristics of each combination, and further judgment is carried out in combination with manual work. The method is realized by comparing the similarity of acceleration waveforms and FAS among three components of a group of seismic oscillation records, and high accuracy is achieved. Normal and mismarked records in a test set can be accurately identified. Compared with a traditional manual method for recognition, the method has the advantages that the recognition accuracy is guaranteed, and meanwhile, the efficiency is remarkably improved.

Description

Technical Field

[0001] The invention relates to a method for identifying vertical components of strong vibration acceleration records based on twin neural networks. Background Art

[0002] High-quality strong motion records are the data basis for earthquake engineering and engineering earthquake research, and are of great significance for site effect analysis, seismic attenuation relationship research, structural seismic response input, and earthquake early warning. However, since the operation, maintenance, and management of the strong motion network involve many links, 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 imminent and challenging task, and quality monitoring of massive strong motion data sets is particularly important.

[0003] Data quality issues include S-wave triggering, spikes, multiple events, etc. Some strong motion data processing programs include quality control processes and identify specific quality issues. Direction issues are part of metadata issues, usually manifested as wrong or missing directions. 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 the cross-correlation of ambient noise. 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 north offset of the seismometer. In fact, during the quality assessment of seismic records, it was found that some records had the phenomenon of three-component channel data interchange, that is, the vertical component of a set of records was mistakenly recognized as the horizontal component, which may be caused by inconsistent instrument output standards or errors in post-processing. This is a data anomaly because most of the vertical components of the records have correct channel labels.

[0004] The vertical component of seismic records is crucial in structural design, application of H / V spectrum ratio method, etc. Therefore, when conducting seismic analysis and seismic design method research, the direction of seismic records must be accurate to avoid errors caused by seismic recording direction problems. In most cases, the acceleration waveforms and Fourier amplitude spectra (FAS) of the two horizontal components, east-west (EW) and north-south (NS), are more similar to the vertical component. Traditionally, the vertical component is generally identified in a set of seismic records by manually comparing the acceleration waveforms of the three components with the shapes of the Fourier spectra, but this method is inefficient and is unrealistic when faced with a large number of records. Therefore, it is very necessary to find a fast and accurate method for identifying vertical seismic records. Summary of the invention

[0005] In order to solve the above technical deficiencies, the present invention provides a method for identifying the vertical component of strong vibration acceleration records based on a twin neural network, which has fast identification and high accuracy.

[0006] The present invention is achieved by the following measures:

[0007] The vertical component identification method of strong vibration acceleration records based on twin neural network includes the following steps:

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

[0009] Step 2, inputting the acceleration characteristic data of each combination in each group of strong vibration data in step 1 into the twin neural network for deep learning and training, so as 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 vibration data to be judged are located, and then combine the three components in pairs according to the method in step 1, and input them into the network model obtained in step 2 in groups for comparison to obtain the time domain similarity T of each combination. sim ;

[0011] Step 4: Perform Fourier transform and smoothing on the acceleration data of the channels where the three components of each group of strong vibration data in step 3 are located to obtain the smoothed fast Fourier transform data FAS. Then, the three components are combined in pairs according to the method in step 1, and input into the network model obtained in step 2 in groups for comparison to obtain the frequency domain similarity F of each combination. sim ;

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

[0013] Step 6: Combined with the detailed parameters of the earthquake motion, further manual discrimination is performed to finally obtain earthquake motion data with incorrect vertical component labels.

[0014] The mutual correlation coefficient in step 1 is used to calculate the similarity of the acceleration waveform and the similarity of the Fourier spectrum of each combination. The specific calculation method is as follows:

[0015]

[0016] Where x(t) and y(t) are the input signals, representing the two components of each combination; τ is the delay time, N is 1 / 2 of the sampling point, and R xy Each point of (τ) has the average of N products, and there must be 2N sampling points; the mutual correlation coefficient ρ xy (τ) is a value between -1 and 1. The larger its absolute value is, the higher the similarity between the two input signals is.

[0017] After the twin neural network extracts the input content, it will generate a multi-dimensional feature vector. In order to input this multi-dimensional feature vector into the subsequent layers, it is necessary to use the Flatten layer to process the vector into one dimension, and then calculate the Manhattan distance PΔxP1 between the two one-dimensional vectors x1 and x2 in the comparison layer. The formula is as follows:

[0018] PΔxP1=abs(x1-x2) (3)

[0019] Then, the distance is input into the fully connected 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. 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.

[0020] In step 2, the training environment of the network model consists of Python 3.9, and the deep learning frameworks TensorFlow 2.10 and Keras 2.10 are installed. The model is trained by backpropagation using the Adam optimizer, the bitch_size is 64, the learning rate is 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. The dropout rate is 0.5, and binary cross entropy is used as the loss function. The formula is as follows:

[0021]

[0022] Where N is the total number of samples, y i is a binary label 0 or 1, p(y i) is the output label y i probability.

[0023] The beneficial effects of the present invention are as follows: the method is implemented by comparing the similarity of the acceleration waveform and FAS between the three components of a set of seismic records, and a fairly high accuracy rate is achieved. By using the two methods for joint recognition, the normal and mislabeled records in the test set can be accurately identified. Combined with manual recognition, it is confirmed that the method can effectively identify the vertical component in a set of seismic records. Compared with the recognition using traditional manual methods alone, the efficiency is significantly improved while ensuring the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is the cumulative distribution diagram of the acceleration waveform cross-correlation coefficient;

[0025] Figure 2 It is the distribution diagram of the mutual correlation coefficient of different combinations of acceleration waveforms;

[0026] Figure 3 is the cumulative distribution diagram of the acceleration Fourier spectrum cross-correlation coefficient;

[0027] Figure 4 is the distribution diagram of the mutual correlation coefficient of acceleration Fourier spectrum;

[0028] Figure 5 a represents the acceleration time course graph, and b represents the Fourier spectrum;

[0029] Figure 6 a in the middle shows the acceleration time course curve recorded by RSN1189, and b shows the FAS graph;

[0030] Figure 7 Indicates the error in the label of the vertical component channel. In the example record, a represents the acceleration time course graph, and b represents the Fourier spectrum. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0032] The present invention discloses a method for distinguishing the vertical component channel of seismic records based on deep learning to achieve efficient and rapid distinction. First, the mutual correlation coefficients of the acceleration time histories and the mutual correlation coefficients of the Fourier spectra between the three components of each group of records in the NGA-West2 database are calculated and analyzed. Then, a twin neural network is used to learn the acceleration features of each combination, and a residual network is used to learn the Fourier spectrum features of each combination, and the method flow combined with further manual judgment is introduced in the third section. Then, the model is trained and verified, as well as the results of the method. Finally, the vertical component identification method of seismic records is applied to China's strong earthquake records to prove the effectiveness of this method in identifying errors in the vertical component channel. The specific steps are as follows:

[0033] Step 1, select several groups of strong vibration data with high accuracy as samples in the strong vibration record database, each group of strong vibration data contains a horizontal component H1, a horizontal component H2 and a vertical component UD, and combine the three components in each group of strong vibration data in pairs and mark them as combination H1-H2, combination H1-UD and combination H2-UD, and then use the mutual correlation coefficient to calculate the acceleration waveform similarity and Fourier spectrum similarity of the two components of each combination, and conclude that the similarity between the two components in the combination H1-H2 is higher than the similarity between the two components in the other two groups of combinations;

[0034] Step 2, inputting the acceleration characteristic data of each combination in each group of strong vibration data in step 1 into the twin neural network for deep learning and training, so as to obtain a network model based on the twin neural network;

[0035] Step 3: Draw the acceleration time history curve of the channel where the three components in each group of strong vibration data to be judged are located, and then combine the three components in pairs according to the method in step 1, and input them into the network model obtained in step 2 in groups for comparison to obtain the time domain similarity T of each combination. sim ;

[0036] Step 4: Perform Fourier transform and smoothing on the acceleration data of the channels where the three components of each group of strong vibration data in step 3 are located to obtain the smoothed fast Fourier transform data FAS. Then, the three components are combined in pairs according to the method in step 1, and input into the network model obtained in step 2 in groups for comparison to obtain the frequency domain similarity F of each combination. sim ;

[0037] Step 5: Based on the time domain similarity T obtained in steps 3 and 4 sim and frequency domain similarity F sim, compare whether the similarity between the two components in the combination H1-H2 of a certain group of strong vibration data is higher than the similarity between the two components in the other two groups of combinations, if yes, then judge that the group of strong vibration data is correct, if not, then judge that there is a label error in the channel where the vertical component of the group of strong vibration data is located, and proceed to step 6;

[0038] Step 6: Combined with the detailed parameters of the earthquake motion, further manual discrimination is performed to finally obtain earthquake motion data with incorrect vertical component labels.

[0039] The goal of the present invention is to distinguish the vertical component from the three components of a set of ground motion records. Given that 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 set of records in pairs, where "H1-H2" represents the combination of two horizontal components, and "H1-UD" and "H2-UD" represent the combination of two horizontal components with a vertical component, respectively.

[0040] The NGA-West2 database was proposed by the Pacific Earthquake Engineering Research Center (PEER) to expand the NGA-West1 database and improve its quality and the accuracy of the earthquake motion prediction equation. The database has been carefully checked, so the accuracy of the earthquake motion records in the database is recognized. The records with different numbers of data points of the three components in the same group of records in the NGA-West2 database and the records with missing vertical components were eliminated, and finally a total of 18447 groups of strong motion records were selected and used. The correlation coefficient is used to calculate the correlation between the acceleration waveform and the Fourier spectrum of each combination. The specific calculation method is as follows:

[0041]

[0042] Where x(t) and y(t) are input signals, representing the two components of the acceleration waveform or Fourier spectrum. τ is the delay time, and N is 1 / 2 of the sampling point, because R xy Each point of (τ) has the average of N products, so there must be 2N sampling points. It should be noted that the correlation coefficient ρ xy (τ) is a value between -1 and 1. The larger its absolute value, the higher the correlation between the two input signals. Since the components of a set of records are not exactly the same, their mutual correlation coefficient ρ xy (τ) will never reach 1.

[0043] Firstly, the cross-correlation coefficients of all combinations of acceleration waveforms were calculated and statistically analyzed. Figure 1The cumulative distribution of the cross-correlation coefficients is given. It can be seen that the cross-correlation coefficient between the two horizontal components (H1-H2) is generally larger than the cross-correlation coefficient between the horizontal and vertical components. The expected values ​​of the cross-correlation coefficients for the three combinations H1-H2, H1-UD, and H2-UD are 0.2254, 0.1961, and 0.1968, respectively, indicating a high correlation between the two horizontal components. The similarity of the cumulative distribution curves for the H1-UD and H2-UD combinations further indicates that the correlation between the horizontal components is higher.

[0044] Figure 2 The scatter plot of the mutual correlation coefficients between different combinations is given, where the blue dotted line is the fitting straight line of the H1-UD combination, and the black dotted line is the fitting straight line of the H2-UD combination. It can be seen from the figure that the slope of the fitting straight line is significantly less than 1, indicating that the mutual correlation coefficients of the H1-UD and H2-UD combinations are mostly lower than those of the H1-H2 combination. However, among the 18,447 sets of records, only 51.5% of the records have the largest mutual correlation coefficient in the H1-H2 combination, which is a relatively low proportion.

[0045] In addition, the cross-correlation coefficients between all combined FAS were calculated. Figure 3 Their cumulative distributions are shown. It can be seen that in the general trend, the correlation between the two horizontal components (H1-H2) is always higher than the correlation between the horizontal and vertical components. In addition, the correlations between these Fourier spectra (expected values ​​of 0.8062 (H1-H2), 0.7744 (H1-UD), and 0.7761 (H2-UD)) are significantly larger than those derived from the acceleration waveforms. Figure 4 The scatter plot of the cross-correlation coefficients of different combinations is shown. The straight line fitted by the scatter plots shows that the correlation between the two horizontal components (H1-H2) is higher than that of the other two combinations. Among the 18447 sets of records, 71.1% of the records have the largest cross-correlation coefficient in the H1-H2 combination. This result indirectly shows that the correlation between the FAS of each combination is stronger than the correlation between the acceleration waveforms.

[0046] Nevertheless, there are still a large number of samples located above the y=x line, indicating this trend. When considering both the acceleration waveform and FAS, 78.92% of the records have the largest cross-correlation coefficient for the H1-H2 combination. This shows that identifying the vertical component using only the cross-correlation coefficient is difficult to achieve. Figure 2 In the figure, the points in the upper left corner represent records with relatively high correlation between H1 and the vertical component. After preliminary analysis, the vertical components of these records may be channel mislabeling. Figure 5 A set of example records is shown in (RSN9325).

[0047] For the identification of the vertical component of seismic records, the traditional method generally judges by manually comparing the similarity between the waveform and the three components of the Fourier spectrum. The traditional method is feasible when identifying a small number of records. However, with the rapid accumulation of seismic records, the traditional method requires a lot of time and effort to accurately identify the channels where the vertical components of massive records are located. Therefore, this paper proposes a method that uses deep learning to make preliminary judgments on a large number of seismic records, and then combines manual identification of vertical components, aiming to improve the accuracy and efficiency of identification, while avoiding the subjectivity in the manual identification process and the problem of standard unification when multiple people collaborate. The method consists of two main parts, one twin neural network for comparing the similarity of acceleration waveforms and the other twin neural network for comparing the similarity of FAS.

[0048] Siamese network is a "Siamese network" that is mainly used for tasks that require comparing two similar inputs. The "Siamese" of a network refers to two identical sub-networks that share the same parameters and are used to process two independent inputs. After the input content passes through the feature extraction network, a multi-dimensional feature vector will be generated. In order to input this multi-dimensional feature vector into subsequent layers, it is necessary to use the Flatten layer to process the vector into one dimension, and then calculate the Manhattan distance PΔxP1 between the two one-dimensional vectors x1 and x2 in the comparison layer:

[0049] PΔxP1=abs(x1-x2) (3)

[0050] Then, the distance is input into the fully connected layer for processing, and 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. Residual Neural Network (ResNet) has a structure that can quickly accelerate neural network training and significantly improve the accuracy of the model. In view of the high performance of ResNet18 in processing images with simple features and the simple features of the input in this study, this paper uses the main part of ResNet18 for image feature extraction.

[0051] The vertical component identification method of earthquake records proposed in this invention is developed based on the traditional identification method. This method uses deep learning technology to replace some manual work in the acceleration waveform comparison and Fourier spectrum comparison in the traditional identification method. The specific steps of this method are as follows:

[0052] 1. Draw the acceleration time history curves of the three channels based on the original data of the earthquake 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).

[0053] 2. Perform Fourier transform and smoothing on the acceleration data of the three channels of earthquake records. Then, the smoothed FAS are input into the twin neural network in pairs for training and comparison to obtain the FAS of each combination. sim (Frequency Domain Similarity).

[0054] 3. According to the T of each combination sim and F sim The size of determines the correctness of the channel where the vertical component of a set of records is located. For a set of records with normal channels, the waveform correlation coefficient and the Fourier spectrum correlation coefficient should be the largest for the H1-H2 combination compared to other combinations. If not, it is preliminarily determined that the vertical component channel of the set of records may have label errors.

[0055] 4. For the records that are initially judged to be suspected of having wrong vertical component labels, further manual judgment is carried out in combination with the detailed information of the earthquake motion (epicenter distance, fault distance, PGA, etc.), and finally the earthquake motion records that may have wrong vertical component labels are obtained.

[0056] The 18447 sets of records in the NGA-West2 database were combined in pairs, totaling 55341 combinations. Each combination was labeled to obtain a sample data set, where the H1-H2 label was 1, the H1-UD and H2-UD labels were 0, that is, the combination of the two horizontal components was a positive sample, and the other combinations were negative samples. In addition, 14983 sets of records with significant differences between the vertical and horizontal components of the waveform and FAS were carefully screened by manual visual inspection, and then 60% were randomly divided as training sets, 20% as validation sets, and 20% as test sets.

[0057] The training environment consists of Python 3.9, and the deep learning framework TensorFlow 2.10 is installed

[38] and Keras 2.10. The model was trained using the Adam optimizer via backpropagation, with a bitch_size of 64 and a learning rate of 0.01. Dropout and early_stopping were used during training, and the best model weights were stored based on the validation loss to prevent overfitting (and reduce training time), with a dropout rate of 0.5. Binary cross entropy was used as the loss function, as shown in Equation 4.

[0058]

[0059] Where N is the total number of samples, y i is a binary label 0 or 1, p(y i ) is the output label y i probability.

[0060] The established sample dataset was used as input to the twin neural network for training and validation. The trained model that achieved the lowest loss in the validation set was then applied to evaluate the test set. For the three combinations in a set of ground motion records, the three similarity values ​​output by the model were compared. The combination with the highest similarity was designated as the H1-H2 combination. In order to evaluate the model's ability to identify channel mislabeled records, 50% (1500 groups) of the records in the test set were randomly selected and their vertical component labels were swapped with horizontal component labels to intentionally create mislabeled records. This label swapping operation was repeated 100 times, using the average of the acceleration waveform and FAS similarities identified individually and in combination.

[0061] For normal samples, the accuracy rates of using acceleration waveform, FAS and joint recognition were 98.7%, 95.6% and 94.3% respectively, while for the created mislabeled samples, the accuracy rates of the three cases were 99.3%, 97.9% and 100% respectively. The trained twin neural network achieved a high accuracy rate on the test set, and the deliberately created mislabeled records were also accurately identified.

[0062] In fact, through the recognition of the training model and further careful manual recognition of 18,447 sets of records in the NGA-West2 database, it was found that 14 sets of ground motion 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 that there may be label errors in the three components of these records. A set of example records (RSN11341) such as Figure 6 shown.

[0063] Since its operation, China Digital Strong Motion Network has obtained a large number of strong motion records with engineering application value. Over the past decade, the construction of the strong motion network has been 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 of seismic records was used to analyze and discuss the channel correctness of the collected strong motion records in China.

[0064] High-quality records (each signal window is clearly discernible) were selected from the strong earthquake records in China 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.

[0065] The trained model was used to evaluate the accuracy of the vertical component labels of 3381 sets of strong earthquake records. Table 1 shows the results of individual and joint judgments using acceleration waveforms and FAS. The preliminary screened seismic records were further manually identified, and 308 sets of seismic records with suspected channel errors were finally obtained. This is consistent with the manual identification results of the 3381 sets of records. This is consistent in all 3381 sets of records. It can be determined from the acceleration waveform and FAS that the vertical component has been confused. The 308 sets of suspected vertical component label errors in China's strong earthquake records are distributed in 170 stations in 12 provinces. The main reason for the error in the recorded channel direction may be problems with the station, and the stations of these records were analyzed.

[0066] Table 1. Identification results of suspected channel labels errors in China's strong earthquake records

[0067]

[0068] There are 28 stations with more than three channel errors and a total error rate greater than 50%. Table 2 provides the codes of these stations.

[0069] Table 2 Number of channel label errors recorded at 28 stations.

[0070]

[0071] It can be seen that the vertical component channel error rate of the 16 stations reached 100%, especially the 7 records of the 053LDS station were all judged as channel errors. It may be that these stations have problems in channel connection, output settings, etc. Therefore, the relevant staff of the stations can use the method in this article to screen the stations that may have problems.

[0072] This method is achieved by comparing the similarity of acceleration waveforms and FAS between the three components of a set of seismic records. First, the three components of each set of seismic records are combined in pairs using the NGA-West2 database, marked as H1-H2, H1-UD and H2-UD respectively. Then, the acceleration time history curves and Fourier spectra of each combination of each set of records are correlated using the mutual correlation coefficient. Subsequently, the acceleration time history curves and smoothed FAS images of each combination are input into the twin neural network for training. According to the discrimination results of the acceleration waveform and FAS, the channel where the vertical component is located is determined in combination with manual identification. Finally, this method was applied to the strong earthquake records in China (2007-2018), and the main conclusions are as follows:

[0073] 1. Correlation analysis of the records in the NGA-West2 database shows that the correlation of the two horizontal component combinations H1-H2 is generally high, but the H1-H2 combination has the largest correlation coefficient, accounting for only 51.5% (for acceleration waveform) and 71.1% (for Fourier spectrum). This shows that the two horizontal components are generally correlated, but it is difficult to use this standard to determine whether the combination is an H1-H2 combination.

[0074] 2. The vertical component identification method proposed in this paper achieved a fairly high accuracy rate. Although some records were mistakenly identified as channel label errors when only acceleration waveforms or FAS were used for identification, deliberately created vertical component label error records were accurately identified in the test set. By using both methods for joint identification, normal and mislabeled records in the test set can be accurately identified. These findings, coupled with manual identification, confirm that the method can effectively identify the vertical component in a set of ground motion records. Using the method proposed in this paper, it was found that there may be vertical component label errors in 14 sets of ground motion records in NGA-West2.

[0075] The vertical component identification method of earthquake records proposed in this invention was used to identify China's strong earthquake records, and it was found that 308 groups of records from 170 stations had suspected vertical component channel label errors, which was basically consistent with the results obtained by simply using traditional manual methods to identify, thereby verifying the effectiveness of the proposed method. The method proposed in this article significantly improves the efficiency while ensuring the recognition accuracy. In addition, among the records with suspected vertical component label errors, 28 stations had more than 3 records with suspected channel errors and the error rate was greater than 50%. Station staff can use the method proposed in this article to quickly find stations that may have problems and verify these stations.

[0076] The above is only a preferred implementation of this patent. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of this patent. These improvements and substitutions should also be regarded as the scope of protection of this patent.

Claims

1. A method for identifying the vertical component of strong vibration acceleration records based on twin neural networks, characterized in that: The following steps are involved: Step 1, select several groups of strong vibration data with high accuracy as samples in the strong vibration record database, each group of strong vibration data contains a horizontal component H1, a horizontal component H2 and a vertical component UD, and combine the three components in each group of strong vibration data in pairs and mark them as combination H1-H2, combination H1-UD and combination H2-UD, and then use the mutual correlation coefficient to calculate the acceleration waveform similarity and Fourier spectrum similarity of the two components of each combination, and conclude that the similarity between the two components in the combination H1-H2 is higher than the similarity between the two components in the other two groups of combinations; Step 2, inputting the acceleration characteristic data of each combination in each group of strong vibration data in step 1 into the twin neural network for deep learning and training, so as to obtain a network model based on the twin neural network; Step 3: Draw the acceleration time history curve of the channel where the three components in each group of strong vibration data to be judged are located, and then combine the three components in pairs according to the method in step 1, and input them into the network model obtained in step 2 in groups for comparison to obtain the time domain similarity T of each combination. sim ; Step 4: Perform Fourier transform and smoothing on the acceleration data of the channels where the three components of each group of strong vibration data in step 3 are located to obtain the smoothed fast Fourier transform data FAS. Then, the three components are combined in pairs according to the method in step 1, and input into the network model obtained in step 2 in groups for comparison to obtain the frequency domain similarity F of each combination. sim ; Step 5: Based on the time domain similarity T obtained in steps 3 and 4 sim and frequency domain similarity F sim , compare whether the similarity between the two components in the combination H1-H2 of a certain group of strong vibration data is higher than the similarity between the two components in the other two groups of combinations, if yes, then judge that the group of strong vibration data is correct, if not, then judge that there is a label error in the channel where the vertical component of the group of strong vibration data is located, and proceed to step 6; Step 6: Combined with the detailed parameters of the earthquake motion, further manual discrimination is performed to finally obtain earthquake motion data with incorrect vertical component labels.

2. The method for identifying the vertical component of strong vibration acceleration records based on twin neural networks according to claim 1 is characterized in that: The mutual correlation coefficient in step 1 is used to calculate the similarity of the acceleration waveform and the similarity of the Fourier spectrum of each combination. The specific calculation method is as follows: Where x(t) and y(t) are the input signals, representing the two components of each combination; τ is the delay time, N is 1 / 2 of the sampling point, and R xy Each point of (τ) has the average of N products, and there must be 2N sampling points; the mutual correlation coefficient ρ xy (τ) is a value between -1 and 1. The larger its absolute value is, the higher the similarity between the two input signals is.

3. The method for identifying the vertical component of strong vibration acceleration records based on twin neural networks according to claim 1 is characterized in that: After the twin neural network extracts the input content, it will generate a multi-dimensional feature vector. In order to input this multi-dimensional feature vector into the subsequent layers, it is necessary to use the Flatten layer to process the vector into one dimension, and then calculate the Manhattan distance PΔxP1 between the two one-dimensional vectors x1 and x2 in the comparison layer. The formula is as follows: PΔxP1=abs(x1-x2) (3) Then, the distance is input into the fully connected 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. 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.

4. The method for identifying the vertical component of strong vibration acceleration records based on twin neural networks according to claim 1 is characterized in that: In step 2, the training environment of the network model consists of Python 3.9, and the deep learning frameworks TensorFlow 2.10 and Keras 2.10 are installed. The model is trained by back propagation using the Adam optimizer, the bitch_size is 64, the learning rate is 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. The dropout rate is 0.5, and binary cross entropy is used as the loss function. The formula is as follows: Where N is the total number of samples, y i is a binary label 0 or 1, p(y i ) is the output label y i probability.

Images