A lightweight seismic phase picking method and system suitable for embedded systems

Abstract

The present invention discloses a lightweight seismic phase picking method and system suitable for embedded systems, comprising the following steps: collecting seismic signal data, and performing noise reduction and filtering on the seismic signal data; constructing a seismic phase picking model for the embedded system; inputting the seismic signal data after noise reduction and filtering into the seismic phase picking model to obtain seismic phase picking results. The present invention combines the inductive learning ability of convolutional neural networks with the global perception advantage of Transformer, aiming to provide a lightweight solution for seismic phase picking networks on low-end embedded processors. This hybrid model not only realizes the fusion of two different network structures at the design level, but also maintains sensitivity to local features while improving global information processing capabilities, providing a novel and effective tool for seismic data processing.

Description

Technical Field

[0001] The present invention belongs to the field of earthquake event detection, and in particular relates to a lightweight earthquake phase picking method and system suitable for embedded systems. Background Art

[0002] In the field of seismic data processing, phase identification is regarded as a crucial and challenging task, which aims to provide researchers with objective and accurate data references in order to improve the accuracy of seismic data inversion. Although high-performance phase identification neural network models can significantly improve performance, this performance improvement is often accompanied by an increase in model size and a longer inference time. In this context, the demand for instant seismic wave phase identification at the acquisition end is increasing, especially with the increasing popularity of distributed systems and Internet of Things technologies. Therefore, it has become an urgent need to research and develop algorithms that are both lightweight and can maintain model performance in resource-constrained environments, such as embedded systems. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention proposes a lightweight seismic phase picking method and system suitable for embedded systems, which combines the inductive learning ability of convolutional neural networks with the global perception advantage of Transformer, aiming to provide a lightweight solution for seismic phase picking networks on low-end embedded processors. This hybrid model not only realizes the fusion of two different network structures at the design level, but also maintains sensitivity to local features while improving global information processing capabilities, providing a novel and effective tool for seismic data processing.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A lightweight seismic phase picking method suitable for embedded systems comprises the following steps:

[0006] Collecting seismic signal data, and performing noise reduction and filtering on the seismic signal data;

[0007] Build a seismic phase picking model for embedded systems;

[0008] The seismic signal data after noise reduction and filtering is input into the seismic phase picking model to obtain the seismic phase picking result.

[0009] Preferably, the seismic phase picking model includes: an initial feature extraction module, a dual-branch attention feature extraction module and a deconvolution module; wherein the dual-branch attention feature extraction module includes: a recursive self-attention module and a skip connection module.

[0010] Preferably, the method of inputting the de-noised and filtered seismic signal data into the seismic phase picking model to obtain the seismic phase picking result comprises:

[0011] Based on the initial feature extraction module, initial features of the seismic signal data after noise reduction and filtering are obtained;

[0012] Based on the dual-branch attention feature extraction module, the initial features are processed to obtain fused features;

[0013] Based on the deconvolution module, the fused features are mapped to a probability sequence associated with the P phase and S phase of the seismic signal at each time point to obtain a seismic phase picking result.

[0014] Preferably, based on the dual-branch attention feature extraction module, the method for processing the initial features to obtain fused features includes:

[0015] By default, X∈R L×C represents the input feature map of the dual-branch attention feature extraction module, through X PW =PW(X) performs point-by-point convolution on the input primary features to obtain the output X PW ;

[0016] Through the bottleneck layer, the X PW Perform feature dimensionality reduction and output local feature X Bottle =BottleNeck(X PW );

[0017] For the local feature X Bottle =BottleNeck(X PW ) performs point-by-point convolution kernel depth convolution and outputs X DW =DW(X Bottle ); the X DW =DW(X Bottle ) Through the recursive self-attention module, the global feature X is obtained attention =RMSA(X DW );

[0018] The local feature X Bottle =BottleNeck(X PW ) and the global feature X attention =RMSA(X DW ) and add them together to get X Feature =X Bottle +X attention ;

[0019] The global feature X attention =RMSA(X DW) and X Feature =X Bottle +X attention Splice along the channel to get X Concat =X attention +X Feature ;

[0020] The output X of the point-by-point convolution kernel depth convolution is converted through the skip connection module DW =DW(X Bottle ) are added, and then fused through 1x1 convolution, and then the fused information is combined with the X Concat =X attention +X Feature Add together to obtain the fused feature X Block =Bag(X PW +X DW )+Fuse(X Concat ).

[0021] Preferably, the recursive self-attention module is composed of N identical enhanced self-attention modules stacked together.

[0022] Preferably, the output of the enhanced self-attention module is:

[0023] X attention =X+SiLU(UP(X"))

[0024] Among them, SiLU is the activation function, and X″ is defined as:

[0025] X"=Concat(X,Attention(X' Q ,X' k ,X' v ))

[0026] Among them, Concat means concatenating two tensors by channel, X′ Q , X′ k , X′ v The input to the enhanced self-attention module consists of query Q, key K and value V;

[0027] Attention is defined as:

[0028] Where d is the number of channels.

[0029] The present invention also provides a lightweight seismic phase picking system suitable for embedded systems, comprising: an acquisition module, a model building module and a phase picking module;

[0030] The acquisition module is used to acquire seismic signal data, and perform noise reduction and filtering on the seismic signal data;

[0031] The model building module is used to build a seismic phase picking model of the embedded system;

[0032] The phase picking module is used to input the seismic signal data after noise reduction and filtering into the seismic phase picking model to obtain the seismic phase picking result.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention proposes a seismic phase picking method and system suitable for embedded systems. The present invention aims to provide an efficient and accurate deep learning solution for embedded systems. It realizes accurate processing of seismic signals by integrating the local feature extraction capability of convolutional neural networks with the global information aggregation advantage of Transformer. Among them, the seismic phase picking model pays special attention to the optimization of computing resources. By adopting lightweight design and dual-branch attention strategy, it greatly reduces processing time and energy consumption while maintaining high inference accuracy. In addition, its recursively modified self-attention mechanism effectively captures subtle dependencies between features and improves the performance of the model in complex data analysis. Jump connections and feature fusion further enhance the robustness of the model. These characteristics make the network of the present invention not only suitable for applications with high real-time requirements, but also easy to deploy on various embedded devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0036] Figure 1 It is a block diagram of a seismic phase picking model applicable to an embedded system according to an embodiment of the present invention;

[0037] Figure 2 It is the framework of the dual-branch attention mechanism implemented in the present invention;

[0038] Figure 3 It is the framework of the recursive self-attention module of the embodiment of the present invention;

[0039] Figure 4 It is a flow chart of a lightweight seismic phase picking method suitable for embedded systems according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the 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.

[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Embodiment 1

[0043] like Figure 4 As shown, the present invention provides a lightweight seismic phase picking method suitable for embedded systems, comprising the following steps:

[0044] Collect seismic signal data, and use hardware circuits to perform noise reduction and filtering on the seismic signal data;

[0045] Build a seismic phase picking model for embedded systems;

[0046] The seismic signal data after noise reduction and filtering is input into the seismic phase picking model to obtain the seismic phase picking results.

[0047] In this embodiment, if Figure 1 As shown, the seismic phase picking model (deep neural network model) includes: an initial feature extraction module, a dual-branch attention feature extraction module and a deconvolution module; wherein the dual-branch attention feature extraction module includes: a recursive self-attention module and a jump connection module.

[0048] In this embodiment, the method of inputting the de-noised and filtered seismic signal data into the seismic phase picking model to obtain the seismic phase picking result includes:

[0049] Based on the initial feature extraction module, the initial features of the seismic signal data after noise reduction and filtering are obtained;

[0050] Based on the dual-branch attention feature extraction module, the initial features are processed to obtain fused features;

[0051] Based on the deconvolution module, the fused features are mapped to the probability sequence associated with the P phase and S phase of the seismic signal at each time point to obtain the seismic phase picking results.

[0052] Specifically: The initial feature extraction module in the embedded system adopts a multi-layer convolutional network structure. The initial features of the input data are extracted by using a set of convolutional layers. For seismic data, these initial features may include basic seismic signal shape and texture information such as edges and corners of various waveforms. These layers have a 3x3 convolution kernel and a step size of 2, which can capture the local features of seismic data while maintaining a low computational cost. This is different from the linear mapping strategy used in the traditional Transformer model. The method of the present invention can more effectively learn useful information from the original data.

[0053] The processed digital signal is transmitted to the embedded processing unit for deep learning model calculation. The feature vector is then fed into a dual-branch attention feature extraction module. This module consists of two branches: a through branch responsible for maintaining local information and an attention branch responsible for extracting global dependencies. The through branch uses residual connections to transfer features, while the attention branch processes global information through an enhanced recursive self-attention module.

[0054] The features after the dual-branch attention feature extraction process will be sent to the deconvolution module. The function of the deconvolution module is to map the high-dimensional feature vector back to the time scale of the original signal, generate a probability sequence corresponding to the P phase and S phase of the seismic signal at each time point, and thus realize the phase picking of the signal.

[0055] The method of processing the initial features based on the dual-branch attention feature extraction module to obtain fused features includes:

[0056] like Figure 2 As shown in the figure, the dual-branch attention feature extraction module adopts a dual-branch strategy. One branch processes local features, which may appear as correlation features of certain continuous time points in the signal. The self-attention mechanism can capture these local dependencies and help the model understand the local dynamics of the seismic signal without performing attention operations. The other branch extracts global features through an enhanced attention mechanism. Since the self-attention mechanism is not limited by the sequence distance, it can extract global context information within the entire sequence. In seismic data, this means that the inter-wave relationship, energy distribution, and propagation effects in the entire seismic event can be captured. This design allows the model to reduce the computational burden caused by attention operations without sacrificing the ability to aggregate global information.

[0057] By default, X∈R L×C represents the input feature map of the dual-branch attention feature extraction module, through X PW =PW(X) performs point-by-point convolution on the input primary features to obtain the output X PW ;

[0058] Through the bottleneck layer, the X PW Perform feature dimensionality reduction, increase the nonlinear expression ability of the model, and output local features X Bottle =BottleNeck(X PW );

[0059] Updated feature X Bottle Moved to two different branches, local and global;

[0060] The local branch directly converts X Bottle Delivered as residuals;

[0061] The global branch is used for the local feature X Bottle =BottleNeck(X PW ) performs point-by-point convolution kernel depth convolution and outputs X DW =DW(X Bottle );Then X DW =DW(X Bottle ) Through the recursive self-attention module, the global feature X is obtained attention =RMSA(X DW );

[0062] The local feature X Bottle =BottleNeck(X PW ) and the global feature X attention =RMSA(X DW ) and add them together to get X Feature =X Bottle +Xa ttention ;

[0063] In order to retain more information, the global feature X attention =RMSA(X DW ) and X Feature =X Bottl e+X attention Splice along the channel to get X Concat =X attention +X Feature ;

[0064] The output X of the point-by-point convolution kernel depth convolution is converted through the skip connection module DW =DW(X Bottle ) are added, and then fused through 1x1 convolution to further extract more delicate features, and the jump connection mechanism is used to strengthen the propagation of features to avoid information loss in the deep network. The fused information is then combined with the X Concat =X attention +X Feature Add together to obtain the fused feature X Block=Bag(X PW +X DW )+Fuse(X Concat ).

[0065] Among them, the recursive self-attention module is composed of multiple identical self-attention units, which are stacked by N identical enhanced self-attention modules, such as Figure 3 shown.

[0066] The processing flow of a single enhanced self-attention module is as follows:

[0067] First define Represents the input of the mechanism, using the standard calculation output of the convolutional neural network as the input of the attention, specifically a 5X5 standard convolution. The input of the attention consists of the query Q, key K and value V, specifically expressed as:

[0068]

[0069] Among them, reshape transforms the number of channels of the input features, and split splits the input tensor in half according to the number of channels.

[0070] The output of the enhanced attention mechanism is

[0071] X attention =X+SiLU(UP(X"))

[0072] Where SiLU is the activation function and X' is defined as:

[0073] X"=Concat(X,Attention(X' Q ,X' k ,X v '))

[0074] Concat means concatenating two tensors by channel.

[0075] Attention is defined as:

[0076] Where d is the number of channels.

[0077] The seismic phase picking model is trained using supervised learning. The input is the preprocessed seismic signal, and the label is the time point of the calibrated P phase and S phase. To adapt to the resource constraints of the embedded system, the model training uses quantization and model pruning technology to reduce the model's memory usage and computational complexity. The trained model is deployed on the embedded device through the conversion tool chain.

[0078] Embodiment 2

[0079] The present invention also provides a lightweight seismic phase picking system suitable for embedded systems, comprising: an acquisition module, a model building module and a phase picking module;

[0080] The acquisition module is used to acquire seismic signal data and perform noise reduction and filtering on the seismic signal data;

[0081] The model building module is used to build a seismic phase picking model for embedded systems;

[0082] The phase picking module is used to input the seismic signal data after noise reduction and filtering into the seismic phase picking model to obtain the seismic phase picking result.

[0083] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A lightweight seismic phase picking method suitable for embedded systems, characterized in that: The following steps are involved: Collecting seismic signal data, and performing noise reduction and filtering on the seismic signal data; Build a seismic phase picking model for embedded systems; Inputting the de-noised and filtered seismic signal data into the seismic phase picking model to obtain a seismic phase picking result; The seismic phase picking model includes: an initial feature extraction module, a dual-branch attention feature extraction module and a deconvolution module; wherein the dual-branch attention feature extraction module includes: a recursive self-attention module and a skip connection module; The method of inputting the de-noised and filtered seismic signal data into the seismic phase picking model to obtain the seismic phase picking result comprises: Based on the initial feature extraction module, initial features of the seismic signal data after noise reduction and filtering are obtained; Based on the dual-branch attention feature extraction module, the initial features are processed to obtain fused features; Based on the deconvolution module, the fused features are mapped to probability sequences associated with the P phase and S phase of the seismic signal at each time point to obtain seismic phase picking results; Based on the dual-branch attention feature extraction module, the method for processing the initial features to obtain fused features includes: By default, X∈R L×C represents the input feature map of the dual-branch attention feature extraction module, through X PW =PW(X) performs point-by-point convolution on the input primary features to obtain the output X PW ; Through the bottleneck layer, the X PW Perform feature dimensionality reduction and output local feature X Bottle =BottleNeck(X PW ); For the local feature X Bottle =BottleNeck(X PW ) performs point-by-point convolution kernel depth convolution and outputs X DW =DW(X Bottle );Then X DW =DW(X Bottle ) Through the recursive self-attention module, the global feature X is obtained attention =RMSA(X DW ); The local feature X Bottle =BottleNeck(X PW ) and the global feature X attention =RMSA(X DW ) and add them together to get X Feature =X Bottle +X attention ; The global feature X attention =RMSA(X DW ) and X Feature =X Bottle +X attention Splice along the channel to get X Concat =X attention +X Feature ; The output X of the point-by-point convolution kernel depth convolution is converted through the skip connection module DW =DW(X Bottle ) are added, and then fused through 1x1 convolution, and then the fused information is combined with the X Concat =X attention +X Feature Add together to obtain the fused feature X Block =Bag(X PW +X DW )+Fuse(X Concat ).

2. The lightweight seismic phase picking method suitable for embedded systems according to claim 1, characterized in that: The recursive self-attention module is composed of N identical enhanced self-attention modules stacked together.

3. The lightweight seismic phase picking method suitable for embedded systems according to claim 2, characterized in that: The output of the enhanced self-attention module is: X attention =X+SiLU(UP(X”)) Among them, SiLU is the activation function, and X″ is defined as: X”=Concat(X,Attention(X' Q ,X' k ,X v ')) Among them, Concat means concatenating two tensors by channel, X′ Q , X′ k , X′ v The input to the enhanced self-attention module consists of query Q, key K and value V; Attention is defined as: Where d is the number of channels.

Images