Training method, prediction method and device of aftershock time-frequency spectrum prediction model

Abstract

A training method, prediction method and device for an aftershock time-frequency spectrum prediction model, the method comprising: constructing an initial prediction model, the initial prediction model comprising an encoder and a decoder connected in sequence; acquiring a training data set and a test data set, wherein each piece of data comprises a time-frequency spectrum corresponding to a main earthquake acceleration time history, a main earthquake seismological parameter characteristic matrix and an aftershock seismological parameter characteristic matrix; training the decoder to obtain a trained decoder; training the encoder by using the training data set, the test data set and the trained decoder to obtain a trained prediction model; the prediction model is used for obtaining an aftershock time-frequency spectrum corresponding to the main shock based on the main and aftershock seismological parameters and the time-frequency spectrum corresponding to the main shock acceleration time history, and the aftershock time-frequency spectrum is used for obtaining an aftershock acceleration time history after phase recovery. According to the method, the decoder is trained first and then the decoder is trained, so that the model training difficulty can be reduced, the training efficiency can be improved, machine jamming can be avoided, and the computer performance can be improved.

Description

Technical Field

[0001] The invention relates to the technical fields of earthquake engineering, neural network and structural engineering, and in particular to a training method, a prediction method and a device for an aftershock time-frequency spectrum prediction model. Background Art

[0002] Aftershocks come from the uneven release of concentrated stress on faults caused by the previous mainshock. One or more aftershocks after a strong mainshock may aggravate structural damage, causing more building collapses and casualties.

[0003] The earthquake acceleration time history curve refers to the curve of the acceleration at a certain point changing with time during an earthquake. The earthquake acceleration time history curve usually takes time as the horizontal axis and acceleration as the vertical axis, reflecting the characteristics of short-period and long-period fluctuations during an earthquake. The aftershock acceleration time history refers to the acceleration time history of an earthquake that occurs after the main shock in an earthquake sequence. The aftershock acceleration time history is crucial for evaluating the performance of building structures and infrastructure in multiple earthquake events. Studying the characteristics of the aftershock acceleration time history and analyzing its impact on the structure can help better understand the cumulative damage caused by earthquakes to buildings.

[0004] Aftershock acceleration time history can be obtained by observing measured data or by numerical simulation. Among them, the measured data method is to use instruments such as accelerometers to record and collect the vibration at a certain point on the ground in real time after the earthquake occurs, and obtain the actual observed acceleration-time data. These data can be used to draw acceleration time history curves and are used for earthquake engineering design and structural seismic performance evaluation. Due to the limitations of the layout of observation equipment, usually only a limited number of observation point data can be obtained.

[0005] The numerical simulation method is to establish an empirical model or earthquake dynamics model, use a computer to perform numerical simulation calculations, and obtain the acceleration-time data at different points during the earthquake process. This method requires the accurate establishment of a suitable empirical model or earthquake dynamics model, and reasonable numerical calculations.

[0006] In the current existing technology, the current prediction methods of aftershock acceleration time history focus more on empirical intensity indicators, lacking the ability to reasonably describe and model the complex relationship between time-frequency characteristics, so the prediction is not accurate enough. In addition, since the training requires a lot of data, it also has high requirements for the model and training method. If the model training method is not good, it will result in poor model training effect, cause machine jams, and low prediction efficiency. Summary of the invention

[0007] 1. Purpose of the invention The purpose of the present invention is to provide a training method, a prediction method and a device for an aftershock time-frequency spectrum prediction model. The training method can reduce the difficulty of training and improve the efficiency of training by training the decoder and the encoder separately, thereby avoiding machine jams and improving computer performance. In addition, the model obtains the aftershock time-frequency spectrum corresponding to the mainshock based on the time-frequency spectrum corresponding to the seismological parameters of the main aftershock and the acceleration time history of the mainshock, and has the ability to reasonably describe the complex correlation between time-frequency characteristics. Compared with other models in the field, the model is more in line with the prediction law.

[0008] (II) Technical solution To solve the above problems, the first aspect of the present invention provides a training method for an aftershock time-frequency spectrum prediction model, comprising: constructing an initial prediction model, the initial prediction model comprising an encoder-decoder connected in sequence; obtaining a training data set and a test data set, each data of the training data set and the test data set comprising a time-frequency spectrum corresponding to the mainshock acceleration time history, a mainshock seismic parameter characteristic matrix and an aftershock seismic parameter characteristic matrix; training the decoder to obtain a trained decoder; using the training data set, the test data set and the trained decoder to train the encoder to obtain a trained prediction model; the prediction model is used to obtain the aftershock time-frequency spectrum corresponding to the mainshock based on the time-frequency spectrum corresponding to the mainshock acceleration time history, and the aftershock time-frequency spectrum is used to obtain the aftershock acceleration time history after phase recovery.

[0009] According to a second aspect of the present invention, there is provided a method for predicting aftershock acceleration time history, comprising: obtaining a time-frequency spectrum of the mainshock acceleration time history corresponding to the aftershock to be measured, a mainshock seismic parameter characteristic matrix and an aftershock seismic parameter characteristic matrix; obtaining a time-frequency spectrum corresponding to the aftershock acceleration time history to be measured according to the time-frequency spectrum, the mainshock seismic parameter characteristic matrix and the aftershock seismic parameter characteristic matrix; and reconstructing the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured to obtain the acceleration time history of the aftershock.

[0010] According to a third aspect of the present invention, there is provided a training device for an aftershock time-frequency spectrum prediction model, comprising: an initial prediction model construction unit, for constructing an initial prediction model, the initial prediction model comprising an encoder and a decoder connected in sequence; a data set acquisition unit, for acquiring a training data set and a test data set, each piece of data in the training data set and the test data set comprising a time-frequency spectrum corresponding to the mainshock acceleration time course, a mainshock seismological parameter characteristic matrix and an aftershock seismological parameter characteristic matrix; a decoder training unit, for training the decoder to obtain a trained decoder; an encoder training unit, for training the encoder using the training data set, the test data set and the trained decoder to obtain a trained prediction model; the prediction model is used to obtain the aftershock time-frequency spectrum corresponding to the mainshock based on the time-frequency spectrum corresponding to the mainshock acceleration time course, and the aftershock time-frequency spectrum is used to obtain the aftershock acceleration time course after phase recovery.

[0011] According to a fourth aspect of the present invention, there is provided an aftershock acceleration time-history prediction device, comprising: an acquisition unit, used to acquire the time-frequency spectrum of the mainshock acceleration time-history corresponding to the aftershock to be measured, the mainshock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix; an input unit, used to obtain the time-frequency spectrum corresponding to the aftershock acceleration time-history to be measured according to the time-frequency spectrum, the mainshock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix; and a prediction unit, used to reconstruct the time-frequency spectrum corresponding to the aftershock acceleration time-history to be measured to obtain the acceleration time-history of the aftershock.

[0012] According to a fifth aspect of the present invention, a storage medium is provided, which stores a computer program, and when the computer program is executed by a processor, it implements the training method of the aftershock time-frequency spectrum prediction model of the first aspect, or when the computer program is executed by the processor, it implements the aftershock acceleration time-history prediction method of the second aspect.

[0013] According to a sixth aspect of the present invention, there is provided an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the training method for the aftershock time-frequency spectrum prediction model of the first aspect is implemented; or, when the computer program is executed by the processor, the aftershock acceleration time-history prediction method of the second aspect is implemented.

[0014] (III) Beneficial effects The above technical solution of the present invention has the following beneficial technical effects: (1) The training method of the aftershock time-frequency spectrum prediction model provided in the embodiment of the present invention trains the decoder first and then the encoder, so that the two are trained separately. Compared with training the encoder and the decoder together, it can reduce the difficulty of model training and improve the training efficiency, thereby avoiding machine jams and improving computer performance.

[0015] (2) The aftershock acceleration time course prediction method provided in the embodiment of the present invention inputs the time-frequency spectrum corresponding to the acceleration time course of the main shock into the trained model to obtain the time-frequency spectrum corresponding to the aftershock acceleration time course, and converts the time-frequency spectrum data into the aftershock acceleration time course, which can more accurately predict the acceleration time course and improve the prediction accuracy of the aftershock acceleration time course. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of a training method for an aftershock time spectrum prediction model provided according to a first embodiment of the present invention; Figure 2 is a schematic diagram of the structure of a prediction model constructed according to a second embodiment of the present invention; Figure 3 is a schematic diagram of the structure of a decoder provided according to a second embodiment of the present invention; Figure 4 is a schematic diagram of the structure of an encoder provided according to a second embodiment of the present invention; Figure 5 is a flow chart of an aftershock acceleration time history prediction method provided according to a third embodiment of the present invention; Figure 6 is a comparison diagram of the aftershock time spectrum and acceleration time history predicted by the prediction method provided by the third embodiment of the present invention and the actual aftershock time spectrum and acceleration time history; Figure 7 2 is a schematic diagram of the structure of an aftershock time-frequency spectrum prediction model training device provided according to a fourth embodiment of the present invention; Figure 8 is a schematic structural diagram of an aftershock prediction device provided according to a fifth embodiment of the present invention; Fig. 9 is a schematic structural diagram of an electronic device provided according to a sixth embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0018] The embodiments described in the present invention are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0020] Before describing the solution of the present invention in detail, the technology related to the present invention is first introduced.

[0021] StyleGAN is a variant of the Generative Adversarial Network (GAN) that makes significant improvements in the architecture of the generator. Based on GAN, StyleGAN introduces a mapping network that maps the high-dimensional latent space to an intermediate latent space. This mapping network consists of multiple fully connected layers, which allows the model to better control the change of style during the generation process. Instead of simply receiving the latent code from the input layer, the generator of StyleGAN adjusts the style of the image according to the latent code in each convolutional layer. This design allows direct control of the strength of image features at different scales. These improvements to StyleGAN have made it a significant improvement in generating high-resolution, high-quality images, and can intuitively control the style of the generated images at different scales.

[0022] Figure 1 FIG. 1 is a flow chart of a training method for an aftershock time-frequency spectrum prediction model according to a first embodiment of the present invention. Figure 1 As shown, the training method includes: step S11-step S14.

[0023] Wherein, in step S11, an initial prediction model is constructed, and the initial prediction model includes an encoder and a decoder connected in sequence.

[0024] Please refer to Figure 2 The core idea of ​​the initial prediction model constructed in this embodiment is to embed the input features of the model into the latent space of the StyleGAN model, input the output of the encoder into the generator (decoder) of the StyleGAN model, and convert the latent space vector into the aftershock time-frequency spectrum. Among them, the decoder of the prediction model can be the generator of the StyleGAN model.

[0025] Step S12, obtaining a training data set and a test data set, wherein each data of the training data set and the test data set includes a time-frequency spectrum corresponding to the mainshock acceleration time history, a mainshock seismological parameter characteristic matrix, and an aftershock seismological parameter characteristic matrix.

[0026] In one embodiment, obtaining a training data set and a test data set includes: steps S121 to S127.

[0027] Step S121, obtaining multiple groups of earthquake acceleration time history data, each group of earthquake acceleration time history data includes the main shock acceleration time history and the aftershock acceleration time history corresponding to the main shock.

[0028] It is understandable that, due to the characteristics of earthquakes, most main shocks will have observable corresponding aftershocks, and each set of earthquake acceleration time history data includes the acceleration time history of a main shock and the aftershock acceleration time history corresponding to the main shock acceleration time history.

[0029] Optionally, each set of earthquake acceleration time history data can be obtained from the measured engineering seismic database PEER NGA-West2.

[0030] Step S122, screening the earthquake acceleration time history data according to the main shock matrix level, the main shock and aftershock recording positions, the time interval between the main shock and aftershocks, and the epicenter distance between the main shock and aftershocks.

[0031] In one embodiment, the earthquake acceleration time history data is screened according to the main shock matrix level, the main shock and aftershock recording positions, the time interval between the main shock and the aftershocks, and the epicenter distance between the main shock and the aftershocks, including: The mainshock acceleration time histories and aftershock acceleration time histories are screened out, whose mainshock moment magnitude is higher than 5.0, whose recording locations of the mainshock and aftershock are located in the same monitoring station, whose aftershock occurs within a preset time after the mainshock, and whose epicenter distances between the mainshock and aftershock are within a preset distance.

[0032] Optionally, the occurrence of aftershocks within a preset time after the main shock includes: the aftershocks occur after the main shock and the interval is within the Gardner-Knopoff time window. In other words, the preset time is the time of a time window. Optionally, when the main shock matrix magnitude is 5, the time window is 1; when the main shock moment magnitude is 5.5, the time window is 2; when the main shock moment magnitude is 5.2, the time window can be set to an interpolation of 1 and 2. Of course, technicians in this field can also choose other values, and this embodiment is not limited to this.

[0033] Optionally, the epicenter distance between the main shock and the aftershock is within a preset distance, including: the epicenter distance between the main shock and the aftershock is within a Gardner-Knopoff distance window.

[0034] Step S123, processing the screened earthquake acceleration time history data to obtain processed earthquake acceleration time history data.

[0035] In one embodiment, the filtered earthquake acceleration time history data is processed to obtain the processed earthquake acceleration time history data, including: performing the following steps in sequence: Preserve the orthogonal horizontal components of ground motion; By zero padding and resampling, the duration of all samples is unified to 160 seconds and the sampling rate is 50 Hz. Studies have shown that 160 seconds can cover the duration range of most actual earthquake samples, and 50 Hz is a more suitable value for conducting time-history response analysis in the fields of earthquake engineering or structural engineering. Therefore, this embodiment adjusts the time and sampling rate of the samples of the above-mentioned earthquake acceleration time-history data to obtain more extensive data.

[0036] Baseline correction and PGA normalization were performed, and the time alignment of the mainshock and aftershocks was performed by maximizing the dot product.

[0037] It is understandable that due to the randomness of earthquake motion and the diversity of station equipment, the original earthquake motion observation data usually has different durations, sampling rates, start and cutoff times, etc. This embodiment processes the data through the above steps, so that all earthquake data can have the same storage format and similar numerical distribution without significantly affecting the physical information contained therein. In addition, data processing can be more helpful for the subsequent training of neural network models and improve prediction accuracy.

[0038] Step S124, converting the earthquake acceleration time history data into a time-frequency spectrum.

[0039] In one embodiment, converting the earthquake acceleration time history data into a time-frequency spectrum includes: converting the earthquake acceleration time history data into a time-frequency spectrum by short-time Fourier transform.

[0040] It can be understood that the main shock acceleration time history and the aftershock acceleration time history can be converted into their respective corresponding time-frequency spectra through short-time Fourier transform.

[0041] In this embodiment, the earthquake acceleration time history data is converted into a time-frequency spectrum through short-time Fourier transform, and the high-dimensional long-time program sequence can be converted into a low-dimensional time-frequency, which can facilitate the recognition and modeling of time-frequency features in the subsequent neural network model.

[0042] In a specific embodiment, the main shock acceleration time history and the aftershock acceleration time history are respectively generated into regular matrices with a side length of 128 through short-time Fourier transform, and then normalized by 0-1 and stored as grayscale images as input features of the subsequent neural network.

[0043] Among them, the horizontal axis and vertical axis of the regular matrix with a side length of 128 correspond to the physical duration of 0-160 seconds and the frequency content of 0-12.5 Hz respectively; the window length of the short-time Fourier transform is 256, the interval between adjacent windows is 64, and the type is Hanning window.

[0044] Step S125, obtaining the seismological parameters of the main shock, each dimension of the main shock seismological parameter characteristic matrix corresponds one-to-one to the main shock seismological parameters, and the main shock seismological parameters include moment magnitude, fault type, fault distance, Joyner-Boore distance, horizontal distance from the top of the fault, fault width, fault dip, hanging plate identification (0 or 1), 30-meter shear wave velocity, and region.

[0045] Specifically, the seismological parameters of the main shock include 11 common scalar indicators summarized by PEER NGA-West2, and the seismological parameters of the aftershock also include 11 common scalar indicators summarized by PEER NGA-West2, and their symbols can be seen in Table 1 below.

[0046] Table 1 Seismological parameters Step S126, simulating the seismic parameters of the main shock based on seismology or earthquake engineering to obtain the seismic parameters of aftershocks.

[0047] Step S127, randomly dividing the processed time-frequency spectrum corresponding to the earthquake acceleration time history, the mainshock seismological parameter characteristic matrix, and the aftershock seismological parameter characteristic matrix into a training data set and a test data set according to a preset ratio.

[0048] In an optional embodiment, the preset ratio can be a ratio of training data to test data close to 9:1, for example (8.5-9):1. In a specific embodiment, the ratio of training data to test data is 2970:332. Each data in the training data set and the test data includes a time-frequency spectrum corresponding to the earthquake acceleration time history, a main shock seismological parameter matrix, and an aftershock seismological parameter matrix.

[0049] It should be noted that the more training data there is, the more accurate the prediction of the trained model will be in theory; at the same time, a certain amount of test data needs to be retained to verify the generalization ability of the model. Since the available aftershock data is relatively less than other artificial intelligence tasks, the ratio is set to approximately 9:1 to balance the above two aspects.

[0050] Step S13, training the decoder to obtain a trained decoder.

[0051] In one embodiment, training the decoder to obtain a trained decoder includes: performing unconditional generative training on the generator of StyleGAN using a time-frequency spectrum dataset of NGA-West2 to obtain a trained decoder.

[0052] In this embodiment, a decoder with fixed parameters can be obtained by performing unconditional generative training on the generator of StyleGAN through the time-frequency spectrum dataset of NGA-West2.

[0053] Figure 3 is a schematic diagram of the structure of a decoder provided according to the second embodiment of the present invention. Figure 3 As shown, the structure of the decoder is designed as a generator of StyleGAN, which consists of 16 stacked groups of convolution and upsampling layers. Compared with the conventional generative adversarial network, StyleGAN uses an extended latent space vector, which consists of 16 512-dimensional vectors; it can be understood that in related models, the latent space vector is usually input only from the first layer, but in this embodiment, the latent space vector is not input only from the first layer, but is sent to each convolution layer separately to perform interaction with the feature map, which is convenient for improving the accuracy of the model.

[0054] Step S14, using the training data set, the test data set and the trained decoder to train the encoder to obtain a trained prediction model; the prediction model is used to obtain the aftershock time spectrum corresponding to the mainshock based on the time spectrum corresponding to the mainshock acceleration time history, and the aftershock time spectrum is used to obtain the aftershock acceleration time history after phase recovery.

[0055] In one embodiment, the encoder is trained using the training data set, the test data set and the trained decoder to obtain a trained prediction model, including: step S141-step S142.

[0056] Wherein, in step S141, the encoder is trained using the training data set, the test data set and the trained decoder to obtain parameters of the encoder.

[0057] In this embodiment, by training the encoder and decoder separately, the difficulty of model training can be reduced, the efficiency of training can be improved, and machine freezes can be avoided, thereby improving computer performance.

[0058] Step S142, adjusting the parameters of the encoder according to the loss value calculated by the preset loss function until a trained encoder is obtained.

[0059] The loss function includes: Where A and A' are the real and predicted aftershock time-frequency spectra, respectively, F is the pre-trained AlexNet, and α is the weight of LPIPS loss.

[0060] It should be noted that, in this embodiment, the loss function adopts a weighted combination of L2 loss and LPIPS loss, which are used to constrain pixel similarity and perceptual similarity respectively.

[0061] Understandably, L2 loss is also known as Euclidean loss or square loss, and the way to calculate the loss is to sum the squares of the differences between the predicted value and the true value. A key feature of L2 loss is that it is very sensitive to outliers, because the square of the difference will amplify larger errors. This may cause the model to pay too much attention to those outliers during training, thereby affecting the model's generalization ability on most data.

[0062] LPIPS, the full name of which is Learned Perceptual Image Patch Similarity, is an image similarity evaluation metric based on deep learning. It uses pre-trained deep networks (such as VGG, AlexNet) to extract image features, and then calculates the distance between these features to evaluate the perceptual similarity between images. The lower the LPIPS value, the more similar the two images are, and vice versa, the greater the difference. This method is more in line with human perception than traditional pixel-based evaluation methods (such as L2 / PSNR, SSIM, FSIM) because it takes into account the global and local structural information of the image, as well as the differences in human sensitivity to different image features.

[0063] Since L2 loss can only constrain low-level pixel information, and LPIPS can use the distance of deep features of the neural network to constrain the high-level structural information of the image, the present invention adopts a weighted combination of L2 loss and LPIPS loss, which can utilize the advantage of LPIPS calculation loss in perceptual similarity, improve the accuracy of loss calculation, and make the model training effect better and the prediction results more accurate.

[0064] In a preferred embodiment, α is 0.025.

[0065] Table 2 below shows some test data about the weight α of LPIPS loss. In the test data, r is the correlation coefficient. Generally, the larger the value, the better the effect. MAE is the mean absolute error. Generally, the smaller the mean absolute error, the better the effect of model training. The lowest loss achieved in training represents the convergence of the model. It can be understood that this embodiment has been adjusted and tested many times. Studies have shown that when α is 0.025, the correlation coefficient r is the largest, the mean absolute error MAE is the smallest, and the model effect is the best. If it is lower than 0.025, it will be too small, which will cause the pixel similarity represented by L2 to dominate, and the model prediction will be easily affected by noise and displacement generated by earthquake randomness; if it exceeds 0.025, it will be too large, which will cause the perceptual similarity represented by LPIPS to dominate, and the model will focus too much on high-level structural information, resulting in the loss of key details of the prediction results.

[0066] Table 2 Partial test data of α The study shows that as the training progresses, when α is 0.025, the loss of the test data set tends to stabilize, the lowest loss is 0.00809, the model converges, and the model effect is the best.

[0067] Figure 4 is a schematic diagram of the structure of an encoder provided according to a second embodiment of the present invention.

[0068] like Figure 4 As shown, the encoder includes a fully connected network, a feature pyramid network, and a map2style network that are connected in sequence.

[0069] The fully connected network is used to perform linear transformation on the input main seismological parameter feature matrix and the residual seismological parameter feature matrix to obtain a feature map of the target dimension, and inject the obtained feature map of the target dimension into the feature pyramid network.

[0070] In one embodiment, the fully connected network may be a three-layer fully connected network, for example, a three-layer fully connected network activated by GELU.

[0071] In this embodiment, the main shock seismological parameter feature matrix and the aftershock seismological parameter feature matrix are first input into a three-layer fully connected network activated by GELU, and the three-layer fully connected network activated by GELU is sequentially subjected to three linear transformations to align the dimensions of the ResNet feature map. Among them, the main shock seismological parameter feature matrix and the aftershock seismological parameter feature matrix are used as auxiliary information for prediction.

[0072] Among them, the feature pyramid network is used to perform upsampling and sum operations based on the time-frequency spectrum corresponding to the input main shock acceleration and the feature map of the target dimension to obtain a multi-scale feature map, and input the multi-scale feature map into the map2style network.

[0073] In one embodiment, the feature pyramid network is a three-layer feature pyramid structure with a ResNet18 backbone, and the parameter settings of the feature pyramid network are the same as those of the original ResNet.

[0074] In this embodiment, the feature pyramid network performs upsampling and summation operations based on the time-frequency spectrum corresponding to the input main shock acceleration and the feature map of the target dimension output from the fully connected network to obtain feature maps of three different scales. The main seismological parameter feature matrix and the residual seismological parameter feature matrix are used as auxiliary information for prediction, so that the potential features of the auxiliary information can be injected into all convolutional layers of ResNet18, and the convolution weights are modulated by adaptive instance normalization.

[0075] The map2style network layer is used to map the multi-scale feature map to a latent space vector, thereby achieving image-to-image translation.

[0076] The Map2Style network can help achieve more refined image editing and control by extracting style features from feature maps at different scales and mapping these feature maps into latent space vectors.

[0077] In one embodiment, the number of map2style network layers is 16, each map2style network is composed of a series of stacked convolutional layers, the distribution ratio of these 16 map2style network layers is 4:6:6, and these 16 map2style network layers map feature maps of three different scales into latent space vectors.

[0078] Furthermore, these 16 map2style network layers adjust the feature maps of three different scales, and each feature map will be dynamically scaled according to the injected information and finally embedded into the latent space.

[0079] In one embodiment, the latent space vector is used to perform style modulation on feature maps of different scales. Auxiliary information such as the mainshock seismological parameter feature matrix and the aftershock seismological parameter feature matrix are integrated with the feature map through style modulation technology to obtain an aftershock latent vector of the target dimension, for example, an aftershock latent vector of 16×512 dimension.

[0080] In one embodiment, the decoder of the prediction model (eg, a StyleGAN generator) and the encoder are used to convert the obtained aftershock latent vector into a time-frequency spectrum, thereby realizing the prediction of the aftershock time-frequency spectrum.

[0081] In this embodiment, since the decoder has been trained in the previous step, the encoder can be randomly initialized and the decoder with fixed parameters after training can form a prediction model. The batch size in training can be set to 8 and the learning rate can be set to 10. -4 In this way, the trained aftershock time-frequency spectrum prediction model is obtained.

[0082] In a preferred embodiment, the prediction model training method further includes: using a Ranger optimizer to adjust the parameters of the encoder during the encoder training process.

[0083] In this embodiment, the Ranger optimizer is used in the model training process to improve the stability of training and accelerate convergence.

[0084] The training method of the aftershock time-frequency spectrum prediction model provided in the embodiment of the present invention trains the decoder first and then the decoder, so that the two are trained separately. Compared with training the encoder and the decoder together, this can reduce the difficulty of model training, improve the training efficiency, and thus avoid machine jams and improve computer performance.

[0085] Figure 5 It is a flow chart of the aftershock acceleration time history prediction method provided according to the third embodiment of the present invention.

[0086] like Figure 5 As shown, the prediction method includes steps S21 to S23.

[0087] Among them, step S21 is to obtain the time-frequency spectrum of the main shock acceleration time history corresponding to the aftershock to be detected, the main shock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix.

[0088] Step S22, obtaining the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured according to the time-frequency spectrum, the mainshock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix.

[0089] In one embodiment, a time-frequency spectrum corresponding to the aftershock acceleration time series to be measured is obtained according to the time-frequency spectrum, the mainshock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix, including: inputting the time-frequency spectrum, the mainshock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix into a prediction model pre-trained according to the method provided in the first embodiment to obtain a time-frequency spectrum corresponding to the aftershock acceleration time series to be measured.

[0090] Step S23, reconstructing the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured to obtain the acceleration time history of the aftershock.

[0091] In one embodiment, the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured is reconstructed to obtain the acceleration time history of the aftershock, including: reconstructing the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured into the acceleration time history of the aftershock through a Griffin-Lim phase retrieval algorithm.

[0092] It is understandable that the Griffin-Lim phase recovery algorithm is a method for recovering the phase spectrum from the amplitude spectrum of a signal, which is usually used as a "vocoder" in the field of speech synthesis, that is, restoring the time-course signal from the Mel spectrum of the sound signal. Similar to sound, the seismic signal also has a long duration and significant non-stationarity, so the acceleration time course can be reconstructed from the amplitude spectrum. The present invention has found that the use of the Griffin-Lim phase recovery algorithm can avoid introducing additional complex models and assumptions, so that the aftershock time spectrum can be simply and accurately reconstructed into the aftershock acceleration time course, thereby improving the accuracy of predicting the aftershock acceleration time course.

[0093] The core idea of ​​the Griffin-Lim algorithm is to use the redundant information in the amplitude spectrum to iteratively adjust the phase until the inter-frame continuity of the amplitude is met. For a transformed or generated amplitude spectrum A, it can be expressed as: Among them, ISTFT stands for inverse short-time Fourier transform, and the parameter setting is the same as STFT; represents the recovered phase spectrum, and N represents the maximum number of iterations.

[0094] In one embodiment, the number of iterations of the inverse short-time Fourier transform and the short-time Fourier transform is the same.

[0095] In a specific embodiment, the number of iterations of both methods is set to 3000 to balance the computational cost and accuracy.

[0096] Figure 6 It is a comparison diagram between the aftershock time spectrum and acceleration time history predicted by the prediction method provided by the third embodiment of the present invention and the actual aftershock time spectrum and acceleration time history.

[0097] In this embodiment, a test sample is selected, and the prediction model trained by the method provided in the first embodiment is used to predict the aftershock time spectrum of the test sample. Then, the predicted aftershock time spectrum is restored to the corresponding aftershock acceleration time history by the Griffin-Lim algorithm. Figure 6 ,The waveform generated by the prediction model is similar to the measured records, and the ,nonstationarity of amplitude and frequency over time is well reproduced, indicating that the prediction effect is ,very good and close to the measured record data.

[0098] Figure 7 2 is a schematic diagram of the structure of an aftershock time-frequency spectrum prediction model training device provided according to a fourth embodiment of the present invention; like Figure 7 As shown, the training device includes: an initial prediction model building unit, a data set acquisition unit, a decoder training unit and an encoder training unit.

[0099] The initial prediction model construction unit is used to construct an initial prediction model, and the initial prediction model includes an encoder and a decoder connected in sequence.

[0100] The data set acquisition unit is used to acquire a training data set and a test data set. Each data of the training data set and the test data set includes a time-frequency spectrum corresponding to the main shock acceleration time history, a main shock seismological parameter characteristic matrix, and an aftershock seismological parameter characteristic matrix.

[0101] In one embodiment, the data set acquisition unit includes: an acceleration time history data acquisition unit, a data screening unit, a data screening unit, a data processing unit, a time-frequency spectrum conversion unit, a mainshock seismological parameter acquisition unit, an aftershock seismological parameter calculation unit and a data set division unit.

[0102] The acceleration time history data acquisition unit is used to acquire multiple groups of earthquake acceleration time history data, each group of earthquake acceleration time history data includes the main shock acceleration time history and the aftershock acceleration time history corresponding to the main shock.

[0103] The data screening unit is used to screen the earthquake acceleration time history data according to the main shock matrix level, the main shock and aftershock recording position, the time interval between the main shock and aftershocks, and the epicenter distance between the main shock and aftershocks.

[0104] In one embodiment, the data screening unit is used to screen out the mainshock acceleration time history and aftershock acceleration time history whose mainshock moment magnitude is higher than 5.0, whose recording locations of the mainshock and aftershock are located in the same monitoring station, whose aftershock occurs within a preset time after the mainshock, and whose epicenter distance between the mainshock and aftershock is within a preset distance.

[0105] The data processing unit is used to process the screened earthquake acceleration time history data to obtain processed earthquake acceleration time history data.

[0106] In one embodiment, a data processing unit is used to process the filtered earthquake acceleration time history data according to the following steps, which include: retaining the orthogonal horizontal components of seismic motion; unifying the duration of all samples to 160 seconds and the sampling rate to 50 Hz by zero padding and resampling; performing baseline correction and PGA normalization, and aligning the main shock and aftershock time by maximizing the dot product.

[0107] The time-frequency spectrum conversion unit is used to convert the earthquake acceleration time history data into a time-frequency spectrum.

[0108] In one embodiment, the time-frequency spectrum conversion unit converts the earthquake acceleration time history data into a time-frequency spectrum by short-time Fourier transform.

[0109] A main shock seismological parameter acquisition unit, wherein each dimension of the main shock seismological parameter matrix corresponds one-to-one to the main shock seismological parameter, and the main shock seismological parameters include moment magnitude, fault type, fault data, Joyner-Boore distance, horizontal distance from the top of the fault, fault width, fault dip, hanging plate identification (0 or 1), 30-meter shear wave velocity, and region.

[0110] The aftershock seismological parameter calculation unit is used to simulate the seismological parameters of the main shock based on seismology or earthquake engineering to obtain the seismological parameters of the aftershock.

[0111] The data set division unit is used to randomly divide the time-frequency spectrum, mainshock seismological parameter matrix, and aftershock seismological parameter matrix corresponding to the processed earthquake acceleration time history into a training data set and a test data set according to a preset ratio.

[0112] The decoder training unit is used to train the decoder to obtain a trained decoder.

[0113] In one embodiment, the decoder training unit performs unconditional generative training on the generator of the StyleGAN using the time-frequency spectrum dataset of NGA-West2 to obtain a trained decoder.

[0114] An encoder training unit is used to train the encoder using the training data set, the test data set and the trained decoder to obtain a trained prediction model; the prediction model is used to obtain the aftershock time spectrum corresponding to the mainshock based on the time spectrum corresponding to the mainshock acceleration time history, and the aftershock time spectrum is used to obtain the aftershock acceleration time history after phase recovery.

[0115] In one embodiment, the encoder training unit includes: a training unit and a parameter adjustment unit.

[0116] A training unit, configured to train the encoder using the training data set, the test data set and the trained decoder to obtain parameters of the encoder; A parameter adjustment unit, used to adjust the parameters of the encoder according to the loss value calculated by the preset loss function until a trained encoder is obtained, Wherein, the loss function includes: Among them, A and A' are the real and predicted aftershock time spectra, respectively, F is the pre-trained AlexNet, and α is the weight of LPIPS loss.

[0117] In one embodiment, α is 0.025.

[0118] In one embodiment, the encoder includes a fully connected network, a feature pyramid network, and a map2style network, which are connected in sequence.

[0119] The fully connected network is used to perform linear transformation on the inputted main seismological characteristic parameter feature matrix and the residual seismological characteristic parameter feature matrix to obtain a feature map of the target dimension, and inject the obtained feature map of the target dimension into the feature pyramid network; The feature pyramid network is used to perform upsampling and summing operations based on the time-frequency spectrum input corresponding to the input main shock acceleration and the feature map of the target dimension to obtain a multi-scale feature map, and input the multi-scale feature map into the map2style network; The map2style network layer is used to map the multi-scale feature map to a latent space vector.

[0120] In one embodiment, the training device further comprises: a Ranger optimizer, configured to adjust parameters of the encoder during the encoder training process.

[0121] Figure 8 4 is a schematic diagram of the structure of an aftershock prediction device provided according to the fifth embodiment of the present invention.

[0122] like Figure 8 As shown, the aftershock acceleration time history prediction device includes: an acquisition unit, an input unit and a prediction unit.

[0123] The acquisition unit is used to acquire the time-frequency spectrum of the main shock acceleration time history corresponding to the aftershock to be detected, the main shock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix.

[0124] The input unit is used to obtain the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured according to the time-frequency spectrum, the mainshock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix.

[0125] The prediction unit is used to reconstruct the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured to obtain the acceleration time history of the aftershock.

[0126] In one embodiment, the input unit is used to input the time-frequency spectrum, the mainshock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix into the device provided by the fourth embodiment to obtain the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured.

[0127] In one embodiment, the prediction unit reconstructs the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured into the acceleration time history of the aftershock by using the Griffin-Lim phase retrieval algorithm.

[0128] Fig. 9 is a schematic structural diagram of an electronic device provided according to a sixth embodiment of the present invention.

[0129] As shown in the figure, the electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the training method of the aftershock time-frequency spectrum prediction model provided in the first embodiment is implemented; or, when the processor executes the computer program, the aftershock acceleration time-history prediction method provided in the second embodiment is implemented.

[0130] The seventh embodiment of the present invention further provides a storage medium storing a computer program, which, when executed by a processor, implements the training method of the aftershock time-frequency spectrum prediction model provided in the first embodiment, or, when executed by a processor, implements the aftershock acceleration time-history prediction method provided in the second embodiment.

[0131] It should be understood that the above specific embodiments of the present invention are only used to illustrate or explain the principles of the present invention, and do not constitute a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made without departing from the spirit and scope of the present invention should be included in the protection scope of the present invention. In addition, the appended claims of the present invention are intended to cover all changes and modifications that fall within the scope and boundaries of the appended claims, or the equivalent forms of such scope and boundaries.

[0132] The present invention has been described above with reference to the embodiments of the present invention. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention. The scope of the present invention is defined by the appended claims and their equivalents. Without departing from the scope of the present invention, a person skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

[0133] Although the embodiments of the present invention have been described in detail, it should be understood that the various changes, substitutions, and alterations could be made hereto without departing from the spirit and scope of the invention.

[0134] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.

[0135] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0136] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0137] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0138] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

Claims

1. A training method for an aftershock time-frequency spectrum prediction model, characterized in that: include: Constructing an initial prediction model, wherein the initial prediction model includes an encoder and a decoder connected in sequence; Acquire a training data set and a test data set, wherein each data item of the training data set and the test data set includes a time-frequency spectrum corresponding to the main shock acceleration time history, a main shock seismological parameter characteristic matrix, and an aftershock seismological parameter characteristic matrix; Training the decoder to obtain a trained decoder; Using the training data set, the test data set and the trained decoder to train the encoder to obtain a trained prediction model; The prediction model is used to obtain the aftershock time spectrum corresponding to the mainshock based on the time spectrum corresponding to the mainshock acceleration time history, and the aftershock time spectrum is used to obtain the aftershock acceleration time history after phase recovery.

2. The method according to claim 1, characterized in that The decoder is a generator of StyleGAN, and the training of the decoder to obtain a trained decoder includes: The StyleGAN generator is unconditionally trained using the NGA-West2 time-frequency spectrum dataset to obtain a trained decoder.

3. The method according to claim 1 or 2, characterized in that: The step of training the encoder using the training data set, the test data set and the trained decoder to obtain a trained prediction model includes: Using the training data set, the test data set and the trained decoder to train the encoder to obtain parameters of the encoder; The parameters of the encoder are adjusted according to the loss value calculated by the preset loss function until a trained encoder is obtained. The loss function includes: Where A and A' are the real and predicted aftershock time spectra, respectively, F is the pre-trained AlexNet, and α is the weight of the LPIPS loss, where α is 0.

025.

4. The method according to claim 1 or 2, characterized in that: The encoder includes a fully connected network, a feature pyramid network and a map2style network connected in sequence; The fully connected network is used to perform linear transformation on the input main shock seismological parameter feature matrix and aftershock seismological parameter feature matrix to obtain a feature vector of a target dimension, and inject the obtained feature vector of the target dimension into a feature graph of the feature pyramid network; The feature pyramid network is used to perform an adaptive instance normalization operation based on the time-frequency spectrum corresponding to the input main shock acceleration and the feature vector of the target dimension to obtain a multi-scale feature map, and input the multi-scale feature map into the map2style network; The map2style network layer is used to map the multi-scale feature map to a latent space vector.

5. The method according to claim 1 or 2, characterized in that: The obtaining of the training data set and the test data set includes: Acquire multiple groups of earthquake acceleration time history data, each group of earthquake acceleration time history data includes a main shock acceleration time history and an aftershock acceleration time history corresponding to the main shock; Screening the earthquake acceleration time history data according to the main shock matrix level, the main shock and aftershock recording location, the time interval between the main shock and aftershocks, and the epicenter distance between the main shock and aftershocks; Processing the screened earthquake acceleration time history data to obtain processed earthquake acceleration time history data; Converting the earthquake acceleration time history data into a time-frequency spectrum; Obtaining the seismological parameters of the main shock, wherein each dimension of the main shock seismological parameter characteristic matrix corresponds to the seismological parameters of the main shock, and the seismological parameters of the main shock include moment magnitude, fault type, fault distance, Joyner-Boore distance, horizontal distance of the top of the fault, fault width, fault dip, hanging plate identification 0 or 1, 30-meter shear wave velocity, and region; Simulating the seismological parameters of the main shock based on seismology or earthquake engineering to obtain the seismological parameters of the aftershock; The time-frequency spectrum corresponding to the earthquake acceleration time history, the main shock seismological parameter characteristic matrix, and the aftershock seismological parameter characteristic matrix are randomly divided into a training data set and a test data set according to a preset ratio.

6. The method according to claim 5, characterized in that in, The screening of the earthquake acceleration time history data according to the main shock matrix level, the main shock and aftershock recording positions, the time interval between the main shock and aftershocks, and the epicenter distance between the main shock and aftershocks includes: Screen out the main shock acceleration time histories and aftershock acceleration time histories of the main shock moment magnitude greater than 5.0, the recording locations of the main shock and aftershock are located in the same monitoring station, the aftershock occurs within a preset time after the main shock, and the epicenter distance between the main shock and aftershock is within a preset distance; or, The step of processing the filtered earthquake acceleration time history data to obtain the processed earthquake acceleration time history data comprises: sequentially performing the following steps: Preserve the orthogonal horizontal components of ground motion; All samples are unified to a duration of 160 seconds and a sampling rate of 50 Hz by zero padding and resampling; Baseline correction and PGA normalization were performed, and the main shock and aftershock time were aligned by maximizing the dot product; Alternatively, the converting the earthquake acceleration time history data into a time-frequency spectrum includes: converting the earthquake acceleration time history data into a time-frequency spectrum by short-time Fourier transform.

7. A method for predicting aftershock acceleration time history, characterized in that: include: Obtaining the time-frequency spectrum of the main shock acceleration time history corresponding to the aftershock to be measured, the main shock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix; Obtaining a time-frequency spectrum corresponding to the aftershock acceleration time history to be measured according to the time-frequency spectrum, the mainshock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix; The time-frequency spectrum corresponding to the aftershock acceleration time history to be measured is reconstructed to obtain the acceleration time history of the aftershock.

8. The method according to claim 7, characterized in that The step of obtaining the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured according to the time-frequency spectrum, the main shock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix comprises: Inputting the time-frequency spectrum, the main shock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix into the prediction model trained according to the method described in any one of claims 1 to 6, to obtain the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured; Alternatively, reconstructing the time-frequency spectrum corresponding to the aftershock acceleration time history to be measured to obtain the acceleration time history of the aftershock includes: The time-frequency spectrum corresponding to the aftershock acceleration time history to be measured is reconstructed into the acceleration time history of the aftershock through the Griffin-Lim phase retrieval algorithm.

9. An aftershock acceleration time history prediction device, characterized in that: include: An acquisition unit, used for acquiring the time-frequency spectrum of the main shock acceleration time history corresponding to the aftershock to be detected, the main shock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix; An input unit, used for obtaining a time-frequency spectrum corresponding to the aftershock acceleration time history to be measured according to the time-frequency spectrum, the main shock seismological parameter characteristic matrix and the aftershock seismological parameter characteristic matrix; The prediction unit is used to reconstruct the time-frequency spectrum corresponding to the aftershock acceleration time history to obtain the acceleration time history of the aftershock.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the training method of the aftershock time-frequency spectrum prediction model as described in any one of claims 1-6, or, when the computer program is executed by the processor, it implements the aftershock acceleration time-history prediction method as described in claim 7 or 8.

Images