A P-wave detection method for seismic signals based on PPNet network
By building a lightweight PPNet network in seismic signal processing, using encoder, feature filter and decoder modules, the problems of poor picking accuracy and large errors at the time of seismic P wave are solved, and high-precision and low-error pickup are achieved and improve the generalization and robustness of the model.
Patent Information
- Application Number
- CN202210908537.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-29
- Publication Date
- 2025-02-25
- Estimated Expiration
- 2042-07-29
AI Technical Summary
The prior art has poor accuracy and large errors in picking up seismic P waves when they arrive, and the deep learning model lacks the ability to extract data features, resulting in low picking efficiency.
A seismic signal P-wave detection method based on PPNet network is proposed. By constructing a lightweight P-wave arrival pickup network PPNet, the encoder module is used to perform deep feature extraction, and the feature filter module refines the P-wave arrival features and in A feature fusion mechanism is added to the decoder module to realize high-precision and low-error P-wave pickup at the time.
The accuracy of picking up seismic P waves at that time is improved, the error is reduced, the feature extraction capability of the model is improved, and the detection of higher accuracy and lower errors is achieved. The network has strong generalization and robustness.
Smart Images
Figure CN115436993B_ABST
Abstract
Description
Technical Field
[0001] The invention discloses a seismic signal P wave detection method based on a PPNet network, belonging to the technical field of seismic signal processing. Background Art
[0002] Seismic phase analysis is the research basis for earthquake activity analysis and evaluation, earthquake tectonic research and earthquake early warning. In seismic phase research, the earthquake P wave is the earliest waveform to reach the surface during earthquake activity. It is also the most frequently appearing waveform in the field of earthquake signal research. Therefore, the study of P wave arrival time picking is a key step in earthquake early warning and monitoring work.
[0003] At present, there are three ways to pick up earthquake P-wave arrival times: manual picking, traditional picking methods, and deep learning picking methods. The manual picking method is difficult in identifying and picking microseismic phases, which can easily cause false detections and lead to poor picking efficiency. The picking effect of traditional P-wave picking methods depends largely on the signal-to-noise ratio and detection interval. The picking accuracy is greatly affected by the set threshold, and the generalization ability and robustness are poor, resulting in low picking efficiency.
[0004] In addition to using traditional methods to pick up earthquake P waves, deep learning methods have become the mainstream technology in the field of processing seismic signals. At present, there are many seismic phase recognition algorithms such as CRED, EQTransformer, ConvNetQuake, and PhaseNet. PhaseNet refers to the construction of the UNet model. It recognizes the three-component earthquake data, uses jump connections for feature fusion, and obtains the P-wave arrival information by outputting a probability sequence of equal length to the input data. Due to the low complexity of its network model, the signal fitting degree needs to be improved; Unet-cea optimizes PhaseNet to achieve better picking performance; the picking network proposed by Li Yu et al., which integrates the spatiotemporal attention mechanism, extracts and fuses multi-scale features of the data, and redistributes weights to achieve a high picking rate; the regression model constructed by Cai Zhenyu et al. does not need to scan the original waveform point by point, and can effectively pick up the first arrival time of the earthquake aftershock P wave. The research on seismic phase identification and arrival time picking of these networks still uses deep neural networks to automatically extract abstract features of complex waveforms. After supervised learning with labeled data, the trained model can outperform traditional methods on test data sets similar to training data. However, the key to its practical application lies in whether the neural network has higher generalization.
[0005] At present, some networks can achieve high generalization. For example, Ross et al. proposed a method for picking P-wave arrival time and determining the initial polarity using deep learning; Perol et al. proposed a method for phase identification and arrival picking based on U-type neural network; Lee et al. proposed an energy ratio method for arrival picking. In recent years, artificial intelligence technology has developed rapidly, and the deep learning algorithms derived from it have also been widely used in many fields. By taking advantage of the advantages of deep learning algorithms in the field of natural language processing, good results have been achieved in phase identification and arrival picking. In addition to the above-mentioned picking models, Earthquake Transformer and others have good P-wave picking effects. However, due to the insufficient ability of deep learning models to extract data features, effective information is often ignored, resulting in low arrival picking efficiency. Therefore, improving the accuracy of P-wave arrival picking is an urgent problem to be solved in the current field of seismic signal research. Summary of the invention
[0006] The present invention overcomes the shortcomings of the prior art and proposes a seismic signal P-wave detection method based on the PPNet network. In view of the problems of poor accuracy and large error in the existing seismic P-wave arrival picking, the feature extraction capability of the model is improved, the deep hidden features of the seismic signal are mined, and the seismic P-wave is detected and picked up with higher accuracy and lower error.
[0007] To achieve the above object, the present invention is implemented by the following technical solution: a seismic signal P wave detection method based on PPNet network, comprising the following steps:
[0008] Step S1: Use manually annotated earthquake P-wave arrival time data as raw data for preprocessing, and divide the data set into a training set and a test set;
[0009] Step S2: Construct a lightweight P-wave arrival time picking network PPNet with the ability to analyze and fuse seismic phase characteristics to train the seismic signal dataset;
[0010] The PPNet includes an encoder module, a feature filter module, and a decoder module. The encoder module performs deep feature extraction on the input waveform data; the feature filter module refines the P wave arrival time feature and completes the feature information; the decoder module processes the output feature sequence to realize the recognition of the P wave arrival time in the feature sequence;
[0011] Step S3: Pick up the earthquake P wave through the PPNet network: share parameters for the multi-layer convolution features output by the network, identify and output the feature sequence output by the network point by point through shared weight information, and accurately pick up the arrival time information of the seismic phase.
[0012] The encoder module consists of 1 initial convolution module and 4 downsampling modules, wherein the initial convolution module consists of 2 convolution layers with a step size of 1 and a channel number of 8, and the downsampling modules are composed of 2 convolution layers. The step size of the first convolution layer is 4, and the number of channels is consistent with the number of channels of the input features. The step size of the second convolution layer is 1, and the number of channels is greater than that of the first convolution layer, and deep feature extraction is performed on the seismic signal input to the network.
[0013] The feature filter module only processes the feature sequences output by the last three downsampling modules of the encoder module, and does not process the initial convolution of the shallow layer of the encoder and the output features of the first downsampling module, thereby avoiding feature loss due to deconvolution and increasing feature diversity.
[0014] The decoder module includes 4 upsampling modules, 3 feature fusion modules and 1 time step fully connected module. The upsampling module consists of an upsampling layer with a step size of 4 and a Cropping layer. The Cropping layer crops the feature sequence output by the upsampling layer.
[0015] The feature fusion module first fuses the decoder features with their corresponding feature fuser features in a concatenated manner, then uses a convolutional layer with a scale of 1 for dimensionality reduction and uses the ReLU function for activation;
[0016] The time-step fully connected module serves as the output layer of the network. It performs fully connected operations on all channels in each time step and outputs a length feature sequence. After the operation is completed, an activation function is used to map the feature sequence to between 0 and 1, and finally outputs a P-wave arrival probability sequence.
[0017] The shared parameters described in step S3 are the shared parameters of the multi-layer convolution features output by the network by the time step fully connected module. The feature sequence output by the network is identified and output point by point through the shared weight information, and the arrival time information of the seismic phase is accurately picked up.
[0018] The beneficial effects of the present invention relative to the prior art are as follows: the present invention combines UNet++ encoding and decoding, integrates feature filters to design a lightweight P-wave arrival time picking network PPNet (P-wave Pyramid Network), the network adopts a convolution layer with a large convolution kernel and a low number of channels in the encoder module, performs deep feature extraction on the input waveform data, effectively extracts the earthquake P-wave arrival time features while suppressing and filtering out interference information, and effectively controls the model parameter quantity; the decoder module adds a feature fusion mechanism in the process of feature restoration to complete the feature information, prevent the problem of poor network performance caused by the lack of feature details, and realize accurate recognition of each time step in the seismic signal sequence; in order to avoid introducing a large number of calculation parameters and affecting the calculation efficiency, the present invention only adds a feature fusion device to the three downsampling modules after the encoder to process the output feature sequence, does not process the shallow initial convolution of the encoder and the output features of the first downsampling module, refines the P-wave arrival time features, improves the arrival time picking accuracy, and accelerates the model recognition efficiency. The method of the present invention improves the picking rate of picking earthquake P waves and reduces the picking error. The network model of the present invention has strong generalization and robustness, and has better comprehensive performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] The present invention will be further described below in conjunction with the accompanying drawings.
[0020] Figure 1 It is a flow chart of the seismic P wave picking method based on PPNet of the present invention.
[0021] Figure 2 This is the network structure diagram of the PPNet of the present invention.
[0022] Figure 3 Schematic diagram of processing results at each stage of data processing in an embodiment of the present invention. DETAILED DESCRIPTION
[0023] like Figures 1 to 3As shown, in view of the problems of low precision and large error of the existing seismic P-wave arrival time picking network, the present invention combines UNet++ encoding and decoding, integrates feature filters to design a lightweight P-wave arrival time picking network PPNet (P-wave Pyramid Network) with seismic phase feature analysis and fusion capabilities, and realizes high-precision and low-error picking of seismic P waves. This network uses a convolution layer with a large convolution kernel and a low number of channels in the encoder module to perform deep feature extraction on the input seismic signal, enhance the weight of the phase arrival time in the signal sequence, and suppress and filter other interference information. In the feature restoration process of the decoder module, a feature fusion mechanism is added to complete the feature information, and accurate identification of each time step in the seismic signal sequence is realized. Only the feature filters are added to the last three downsampling modules of the encoder to process the output feature sequence, and the output features of the shallow initial convolution of the encoder and the first downsampling module are not processed, and the P-wave arrival features are refined to improve the arrival time picking accuracy.
[0024] like Figure 2 The network structure diagram of the PPNet of the present invention is shown, which is composed of an encoder module, a decoder module and a feature filter module. The encoder module is constructed by referring to PhaseNet, and adopts a convolution layer with a large convolution kernel and a low number of channels, including an initial convolution module and a downsampling module, which can effectively extract the characteristics of the earthquake phase arrival time and suppress the filtering of a large amount of interference information, and the low number of channels can effectively control the parameter amount of the model; the decoder module is composed of an upsampling module and a time step fully connected module, the deconvolution layer is used to restore the feature scale, the Cropping layer is used to control the scale change caused by upsampling, and the time step fully connected layer shares parameters for the multi-layer convolution features output by the network, and identifies and outputs the feature sequence output by the network point by point through the shared weight information, accurately picking up the earthquake phase arrival information; the feature filter module further extracts the output features of the encoder at different stages according to the feature granularity, and performs feature fusion with the corresponding decoder, avoiding the feature loss caused by deconvolution, and increasing the diversity of features. The network module structure and its parameters, input and output scales are shown in Table 1:
[0025]
[0026] Table 1 Parameters, input and output scales of each module of the PPNet network.
[0027] The three modules of the present invention are introduced in sequence below.
[0028] Encoder module:
[0029] The encoder module can perform deep feature extraction on the seismic signal input to the network, enhance the arrival weight of the seismic phase in the signal sequence, and suppress and filter other interfering information.
[0030] The encoder module of the present invention is composed of 1 initial convolution module and 4 downsampling modules. The convolution kernel size of all convolution layers is 7, and the activation function is the ReLU function. The encoder with a large-scale convolution kernel has a larger receptive field in each convolution operation, which is conducive to comprehensive global feature analysis, greatly improving the model's ability to analyze signal sequences, effectively extracting the arrival time characteristics of the seismic phase, enhancing the weight characteristics near the arrival time, and suppressing interference features such as waveform peaks.
[0031] The initial convolution module consists of two convolution layers with a step size of 1 and a channel number of 8. This module performs preliminary feature extraction on the input seismic signal waveform data, while increasing the number of data channels and enriching the feature details.
[0032] The downsampling modules are composed of 2 convolutional layers. The step size of the first convolutional layer is 4, and the number of channels is consistent with the number of channels of the input features. The step size of the second convolutional layer is 1, and the number of channels is increased compared with the first convolutional layer. As the depth of the encoder network increases, the features are filtered by each convolutional layer and nonlinearly fitted by the pooling layer. The seismic phase arrival feature is greatly enhanced, the feature dimension is improved, and the hidden features of the data are deeply mined and expressed.
[0033] Feature Filter Module:
[0034] The feature filter module has two functions: First, it further refines the features extracted by the encoder to obtain more accurate arrival information and improve the recognition accuracy of the network; second, it completes the feature information of the decoder's feature sequence to prevent the feature sequence from being insufficiently characterized due to upsampling.
[0035] The feature filter module of the present invention uses a convolution layer with a small convolution kernel to process the feature sequence output by each downsampling module of the encoder, refines the seismic phase arrival feature, and improves the arrival time picking accuracy. The encoder module performs sequence analysis and phase arrival feature extraction on the seismic signal data, roughly determines the range of the phase arrival time, and the feature filter has a smaller calculation granularity, which can refine the analysis of the feature sequence and obtain more accurate arrival time information.
[0036] According to the different positions of the downsampling module in the encoder, the number of convolution layers and the size of the convolution kernel of the corresponding feature filter are also different. The deeper the position of the downsampling module in the encoder, the higher the feature condensation, the fewer the number of convolution layers of the feature filter used, and the smaller the convolution kernel. When the position is shallow, the shallow feature characterization ability obtained is weak, and a deeper and larger-scale feature filter is required to perform feature filtering and sequence analysis on the output features. However, due to the limited filtering degree and insufficient nonlinear fitting degree in the early stage of the encoder, the output feature sequence has more noise and insufficient seismic phase information, and a deep feature filter needs to be constructed to process it, which will introduce a large number of parameters and the calculation efficiency will be greatly affected. Therefore, the present invention only adds feature filters to the last three downsampling modules of the encoder to process the output feature sequence, and does not process the initial convolution of the shallow layer of the encoder and the output features of the last downsampling module. In this way, the network parameter amount can be effectively reduced and the model recognition efficiency can be improved under the premise of ensuring the network recognition performance. In the seismic waveform signal data, the time feature is weaker than the peak feature. If it is directly fused with the decoder feature, it will interfere with the feature that has completed the deep filtering and affect the picking performance. Therefore, in the process of encoding and decoding feature fusion, the present invention first uses a feature filter to deeply explore and extract the encoder features, and effectively filters out redundant and noise information. The obtained feature sequence has distinct features and rich detail features. After fusing this sequence with the features obtained by upsampling the decoder, the detailed features of the sequence are effectively supplemented, avoiding the pollution problem of invalid features on the later feature sequence, and improving the network recognition performance.
[0037] Decoder module:
[0038] The decoder module can restore the feature sequence to the same length as the input data, identify each time step, and realize the recognition of the arrival time of the P wave in the feature sequence. The present invention adds a feature fusion mechanism in the process of feature restoration to complete the feature information, while preventing the problem of poor network performance caused by the lack of feature details.
[0039] The decoder module of the present invention includes 4 upsampling modules, 3 feature fusion modules and 1 time step full connection module. The upsampling module is composed of an upsampling layer with a step length of 4 and a Cropping layer. The upsampling layer restores the length of the feature sequence, and the Cropping layer clips the feature sequence output by the upsampling to solve the length change problem caused by rounding up the downsampling feature sequence; the feature fusion module first fuses the decoder feature with its corresponding feature fuser feature in a Concate manner, and then uses a convolutional layer with a scale of 1 for dimensionality reduction and uses the ReLU function for activation; the time step full connection is used as the output layer of the network, and all channels in each time step are fully connected. The output length is a feature sequence with a t dimension of 1. After the operation is completed, the Sigmoid function is used to map the feature sequence to between 0 and 1, and finally the P wave arrival probability sequence is output. The time step full connection realizes the accurate identification of each time step in the seismic signal sequence through point-by-point prediction.
[0040] The method of the present invention can realize high-precision and low-error picking of P-wave arrival time, has strong data feature extraction and fitting capabilities, and the network model has strong generalization and robustness.
[0041] Regarding the specific structure of the present invention, it should be noted that the connection relationship between the various component modules adopted in the present invention is definite and feasible. Except for special instructions in the embodiments, the specific connection relationship can bring about corresponding technical effects, and solve the technical problems raised by the present invention without relying on the execution of corresponding software programs. The components, modules, models of specific components, the connection methods between each other, and the conventional use methods and expected technical effects brought about by the above-mentioned technical features, except for specific instructions, all belong to the disclosed contents in patents, journal articles, technical manuals, technical dictionaries, and textbooks that can be obtained by technical personnel in this field before the application date, or belong to the existing technologies such as conventional technologies and common knowledge in this field, and there is no need to elaborate, so that the technical solution provided in this case is clear, complete, and feasible, and the corresponding physical products can be reproduced or obtained according to the technical means.
[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A seismic signal P wave detection method based on PPNet network, characterized by: The steps include: Step S1: Use manually annotated earthquake P-wave arrival time data as raw data for preprocessing, and divide the data set into a training set and a test set; Step S2: Construct a lightweight P-wave arrival time picking network PPNet (P-wave Pyramid Network) with the ability to analyze and fuse seismic phase characteristics to train the seismic signal dataset; The PPNet network includes an encoder module, a feature filter module, and a decoder module, and the encoder module performs deep feature extraction on the input waveform data; The feature filter module refines the P wave arrival time feature and completes the feature information; the decoder module processes the output feature sequence to realize the recognition of the P wave arrival time in the feature sequence; The encoder module is composed of 1 initial convolution module and 4 downsampling modules, wherein the initial convolution module is composed of 2 convolution layers with a step size of 1 and a channel number of 8, and the downsampling modules are composed of 2 convolution layers, the step size of the first convolution layer is 4, and the number of channels is consistent with the number of channels of the input features, and the step size of the second convolution layer is 1, and the number of channels is greater than the first convolution layer, and deep feature extraction is performed on the seismic signal input to the network; The feature filter module only processes the feature sequences output by the last three downsampling modules of the encoder module, and does not process the initial convolution of the shallow layer of the encoder and the output features of the first downsampling module; The decoder module includes 4 upsampling modules, 3 feature fusion modules and 1 time step fully connected module. The upsampling module is composed of upsampling layers. The upsampling layers include deconvolution layers and Cropping layers. The deconvolution layers restore the feature scales, and the Cropping layers crop the feature sequences output by upsampling. The feature fusion module first fuses the decoder features with their corresponding feature fuser features in a concatenated manner, then uses a convolutional layer with a scale of 1 for dimensionality reduction and uses the ReLU function for activation; The time-step fully connected module serves as the output layer of the network. It performs a fully connected operation on all channels in each time step and outputs a length feature sequence. After the operation is completed, an activation function is used to map the feature sequence to between 0 and 1, and finally outputs a P-wave arrival probability sequence. Step S3: Pick up the earthquake P wave through the PPNet network: The time step fully connected module in the decoder module shares parameters for the multi-layer convolution features output by the network, identifies and outputs the feature sequence output by the network point by point through shared weight information, and accurately picks up the arrival time information of the seismic phase.
2. The method for detecting P waves of seismic signals based on PPNet network according to claim 1, characterized in that: The shared parameters described in step S3 are the shared parameters of the multi-layer convolution features output by the network by the time step fully connected module. The feature sequence output by the network is identified and output point by point through the shared weight information, and the arrival time information of the seismic phase is accurately picked up.