Earthquake seismic phase pickup method and system based on meta-knowledge memory bank enhancement
By constructing a meta-knowledge memory base and integrating earthquake regional features using graph neural networks and memory enhancement modules, the problems of cross-regional adaptability and knowledge reuse of deep learning models are solved, thereby improving the intelligence and deployment efficiency of earthquake phase picking.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INSTITUTE OF PETROCHEMICAL TECHNOLOGY
- Filing Date
- 2026-02-02
- Publication Date
- 2026-05-12
AI Technical Summary
Existing deep learning models tend to forget the knowledge features of the original region when adapting to new regional data, resulting in a decline in seismic phase picking performance. Furthermore, the lack of cross-regional knowledge reuse mechanisms leads to a waste of data resources and learning experience, insufficient continuous learning ability of the model, and low deployment efficiency.
A meta-knowledge memory bank is constructed, and a graph neural network is used to learn the relationship graph of different regions to enhance the representation of earthquake regional features. The memory enhancement module is used to fuse basic temporal features and memory vectors to achieve efficient accumulation and transfer of knowledge, which is suitable for the continuous construction of earthquake monitoring networks.
It enables efficient accumulation and reuse of cross-regional earthquake knowledge, enhances the intelligent capability of earthquake phase picking, adapts to the phase picking performance of new regions, and reduces model training costs and computational costs.
Smart Images
Figure CN122017974A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of earthquake phase picking, specifically relating to an earthquake phase picking method and system based on a meta-knowledge memory base. Background Technology
[0002] Deep learning technology, with its powerful feature extraction capabilities, excels in processing large amounts of complex data and has been widely applied in seismic phase picking tasks. However, existing deep learning models still face the following challenges: To adapt to data from new regions, models require fine-tuning, often forgetting knowledge and features learned in existing regions. This leads to a significant decline in seismic phase picking performance in those regions, making it difficult for models to simultaneously serve stations in multiple different areas. This severely limits the deployment efficiency of deep learning methods in large-scale earthquake monitoring. Furthermore, different seismic regions may share inherent similarities in geological structure types, focal mechanisms, and waveform propagation characteristics, but existing methods lack effective knowledge reuse mechanisms, failing to fully exploit and utilize these cross-regional common features, resulting in a waste of data resources and learning experience. In addition, the models' continuous learning capabilities are insufficient. Traditional methods employ a static training paradigm, requiring retraining the entire model when acquiring labeled data from new regions. This is not only computationally expensive but also fails to achieve gradual knowledge accumulation and transfer.
[0003] In recent years, meta-knowledge has shown significant potential in complex data reasoning. By constructing a sustainably evolving meta-knowledge memory, efficient accumulation, retrieval, and reuse of cross-regional earthquake knowledge can be achieved, providing a new approach for earthquake phase picking. Summary of the Invention
[0004] This invention provides an earthquake phase picking method and system based on a meta-knowledge memory base, which solves the problems of forgetting caused by traditional fine-tuning methods and the inability to reuse knowledge for similar areas. It is particularly suitable for the continuous construction of earthquake monitoring networks, and can continuously accumulate and utilize knowledge as new stations are added, providing earthquake early warning systems with increasingly intelligent phase picking capabilities.
[0005] To achieve the above objectives, the present invention provides the following solution: A method for seismic phase picking based on a meta-knowledge memory base, the method comprising: Acquire seismic waveform data; The seismic waveform data is processed by the basic model encoder of the phase picking to obtain the basic time series characteristics; Construct a meta-knowledge memory base; An index is generated by encoding regional features, and a memory vector is obtained by retrieving it from the meta-knowledge memory base. The memory enhancement module is used to fuse the basic temporal features and the memory vector to obtain the enhanced feature representation; The enhanced feature representation is decoded using a decoder to output P-wave and S-wave probability sequences, thus completing the phase picking task.
[0006] Preferred methods for constructing a meta-knowledge memory include: By using graph neural networks to learn and fuse neighboring information based on the relationship graphs of different regions, the representation of earthquake region characteristics is enhanced, and a meta-knowledge memory is constructed.
[0007] Preferably, the method of generating an index through region feature encoding and retrieving it from the meta-knowledge memory base to obtain the memory vector includes: ; ; in, , , This indicates the weight of the memory meta-knowledge in the fusion process. This represents the retrieved historical metadata related to the features of the target region. Indicates sharpness, This represents a memory vector, which incorporates retrieved meta-knowledge for fusion. For the enhanced region feature representation, For the weighted attention coefficient, As a regional feature, It is a non-linear activation function. i and j They represent different regions. Indicates the region i Number of support set samples.
[0008] Preferably, the method for fusing the basic temporal features and the memory vector using a memory enhancement module to obtain the enhanced feature representation includes: ; ; in, Indicates a gating mechanism. For modulation signal, It is a temporal characteristic.
[0009] Preferably, the method of using a decoder to decode the enhanced feature representation and output P-wave and S-wave probability sequences to complete the phase picking task includes: ; ; in, Indicates wave type, hup This represents the waveform sequence obtained after upsampling. This indicates the computation of the fully connected layer. Represents the probability sequences of P-waves and S-waves. l Indicates the number of floors.
[0010] The present invention also provides an earthquake phase picking system based on a meta-knowledge memory base, the system being used to implement the aforementioned method, the system comprising: an acquisition module, an encoding module, a construction module, a retrieval module, a fusion module, and a decoding module; The acquisition module is used to acquire seismic waveform data; The encoding module is used to process seismic waveform data through a seismic phase picking basic model encoder to obtain basic time series characteristics; The construction module is used to construct a meta-knowledge memory library; The retrieval module is used to generate an index through regional feature encoding, retrieve it from the meta-knowledge memory base, and obtain a memory vector. The fusion module is used to fuse the basic temporal features and the memory vector using the memory enhancement module to obtain the enhanced feature representation; The decoding module is used to decode the enhanced feature representation using a decoder and output P-wave and S-wave probability sequences to complete the phase picking task.
[0011] Preferably, the process of constructing a meta-knowledge memory includes: By using graph neural networks to learn and fuse neighboring information based on the relationship graphs of different regions, the representation of earthquake region characteristics is enhanced, and a meta-knowledge memory is constructed.
[0012] Preferably, the process of generating an index through regional feature encoding, retrieving it from the meta-knowledge memory base, and obtaining the memory vector includes: ; ; in, , , This indicates the weight of the memory meta-knowledge in the fusion process. This represents the retrieved historical metadata related to the features of the target region. Indicates sharpness, This represents a memory vector, which incorporates retrieved meta-knowledge for fusion. For the enhanced region feature representation, For the weighted attention coefficient, As a regional feature, It is a non-linear activation function. i and jThey represent different regions. Indicates the region i Number of support set samples.
[0013] Preferably, the process of fusing the basic temporal features and the memory vector using a memory enhancement module to obtain the enhanced feature representation includes: ; ; in, Indicates a gating mechanism. For modulation signal, It is a temporal characteristic.
[0014] Preferably, the process of using a decoder to decode the enhanced feature representation and output P-wave and S-wave probability sequences to complete the phase picking task includes: ; ; in, Indicates wave type, h up This represents the waveform sequence obtained after upsampling. This indicates the computation of the fully connected layer. Represents the probability sequences of P-waves and S-waves. l Indicates the number of floors.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a seismic phase picking method based on a meta-knowledge memory base. By constructing a meta-knowledge memory base, relevant historical meta-knowledge is retrieved and updated. In the base model, the memory vectors of the memory enhancement module apply relevant historical meta-knowledge to the feature representation, which solves the problems of forgetting and the inability to reuse knowledge for similar areas in traditional fine-tuning methods. This method is particularly suitable for the continuous construction of earthquake monitoring networks, enabling the continuous accumulation and utilization of knowledge as new stations are added, providing increasingly intelligent seismic phase picking capabilities for earthquake early warning systems. Attached Figure Description
[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of the method flow in an embodiment of the present invention; Figure 2This is a schematic diagram of the basic model for seismic phase picking in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the construction of the meta-knowledge memory in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the contents of the meta-knowledge memory in an embodiment of the present invention; Figure 5 This is a schematic diagram of historical knowledge retrieval in the meta-knowledge memory database in an embodiment of the present invention; Figure 6 This is a schematic diagram of model reasoning in an embodiment of the present invention; Figure 7 This is a schematic diagram of knowledge updating in the meta-knowledge memory in an embodiment of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0020] Example 1 This invention provides a method for seismic phase picking based on a meta-knowledge memory base, the method comprising: Acquire seismic waveform data; The seismic waveform data is processed by the basic model encoder of the phase picking to obtain the basic time series characteristics; Construct a meta-knowledge memory base; An index is generated by encoding regional features, and a memory vector is obtained by retrieving it from the meta-knowledge memory base. The memory enhancement module is used to fuse the basic temporal features and the memory vector to obtain the enhanced feature representation; The enhanced feature representation is decoded using a decoder to output P-wave and S-wave probability sequences, thus completing the phase picking task.
[0021] In this embodiment, the method for constructing a meta-knowledge memory includes: By using graph neural networks to learn and fuse neighboring information based on the relationship graphs of different regions, the representation of earthquake region characteristics is enhanced, and a meta-knowledge memory is constructed.
[0022] In this embodiment, the method for generating an index through region feature encoding and retrieving it from the meta-knowledge memory base to obtain the memory vector includes: ; ; in, , , This indicates the weight of the memory meta-knowledge in the fusion process. This represents the retrieved historical metadata related to the features of the target region. Indicates sharpness, This represents a memory vector, which incorporates retrieved meta-knowledge for fusion. For the enhanced region feature representation, For the weighted attention coefficient, As a regional feature, It is a non-linear activation function. i and j They represent different regions. Indicates the region i Number of support set samples.
[0023] In this embodiment, the method for fusing the basic temporal features and the memory vector using a memory enhancement module to obtain the enhanced feature representation includes: ; ; in, Indicates a gating mechanism. For modulation signal, It is a temporal characteristic.
[0024] In this embodiment, the method for decoding the enhanced feature representation using a decoder to output P-wave and S-wave probability sequences to complete the phase picking task includes: ; ; in, Indicates wave type, h up This represents the waveform sequence obtained after upsampling. This indicates the computation of the fully connected layer. Represents the probability sequences of P-waves and S-waves. l Indicates the number of floors.
[0025] Example 2 like Figure 1As shown, this embodiment provides a seismic phase picking method based on a meta-knowledge memory. The method consists of three parts: an encoder S1, a meta-knowledge memory S2, and a decoder S3.
[0026] During the training phase of the earthquake phase picking basic model, only encoder S1 and decoder S3 participate in the training.
[0027] The meta-knowledge memory S2 is constructed as an independent module to store meta-knowledge information generated during the historical earthquake phase picking process.
[0028] During the integration phase, the encoder S1, the meta-knowledge memory S2, and the decoder S3 are combined to form a complete phase picking model.
[0029] By using a meta-knowledge memory in the earthquake phase picking model task, the efficient utilization of historical earthquake meta-knowledge is achieved, thereby enhancing knowledge transfer in the earthquake phase picking task.
[0030] like Figure 2 As shown, a seismic phase picking model based on a deep neural network is designed, including an encoder S1 and a decoder S3.
[0031] Input three-component seismic waveform data ,in T This represents the number of sampling points.
[0032] The seismic waveform data is encoded using encoder S1, and a multi-layer one-dimensional convolutional network is used to extract local feature representations of the seismic waveform data. A bidirectional approach is employed. LSTM Modeling the long-range time-series dependency of seismic waveforms yields the following time-series characteristics: in l Indicates the number of floors. d Representing feature dimension, This indicates the time series length after downsampling.
[0033] The aforementioned temporal features are input into the memory enhancement module (the memory vector v is then used). mem The enhanced feature representation is obtained by fusing it into this basic temporal feature module. During the pre-training phase of the basic model, the memory enhancement module is empty and degenerates into an identity mapping. The temporal features are decoded using decoder S3, and the probability sequences of P-waves and S-waves are output through an upsampling layer and two independent fully connected layers: in Indicates wave type, hup This represents the waveform sequence obtained after upsampling. This indicates the computation of the fully connected layer. This represents the probability sequence of P-waves and S-waves.
[0034] By constructing relationships between different regions, seismic waveform features of each region are extracted. A graph neural network is then used to learn and fuse adjacent information based on the relationship graph of different regions, thereby enhancing the representation of seismic region features and constructing a meta-knowledge memory S2. For example... Figure 3 As shown, for each region or station i Historical earthquake waveforms are collected to form the support set for the region, and this support set is defined as follows: in Indicates the region i Number of supporting samples It is the first n A three-component waveform.
[0035] For each support set The region feature encoding E is performed through residual convolution at four different scales, constructing feature representations at four scales: Concatenate vectors of four scales and then... MLP Projecting yields the initial feature representations for each region: in d It is the embedded dimension. This represents the feature vector after concatenation across four scales. This represents the initial feature representation after MLP projection. For the ... i The feature representations within each region are averaged and pooled to obtain... .
[0036] After obtaining the initial feature representations of each region, each region is represented as a node, and a regional relationship graph is constructed to model the spatial and geological relationships between regions: in Indicates the weighting coefficient. Indicates geographical proximity. Indicating the same geological structure, Indicates the region i and region j Edge weights between them; dist ( loc i , loc j ) indicates a region i and j Geographical distance between them; Representing regional characteristics and Cosine similarity between them; , Indicates the region i and j The geological structure to which it belongs. i and j They represent different regions.
[0037] Graph attention networks are used to learn and fuse information from neighboring regions: in Indicates the attention coefficient. This represents the learnable weight matrix. Represents the attention parameter vector. Represents the set of adjacent nodes. The edge weights are... Integrating this into the attention mechanism yields the weighted attention coefficients: in The weighting coefficients are expressed. Node feature representations are updated through the fusion of information from neighboring regions as follows: in It is a non-linear activation function. The enhanced region features are represented... , serving as an index Q in the memory knowledge of each region, is used for meta-knowledge retrieval of new regions or new stations.
[0038] The meta-knowledge memory S2 is divided into three subsets as described above, such as... Figure 4 The first subset shown memorizes global meta-knowledge such as the general characteristic patterns of P-wave arrivals and the typical manifestations of high S-wave amplitude; specifically: in For global feature vectors, For universal memory vectors, For reliability scoring, This represents the number of global regions.
[0039] The second subset memorizes regional meta-knowledge such as waveform propagation characteristics, noise patterns, and focal depth distribution features caused by regional geological structures; specifically: in For the region feature vector, For region memory vectors, Score the confidence level. This is the set of adjacent regions of this region. This represents the number of geological regions.
[0040] The third subset focuses on memorizing station instrument response characteristics, local site effects, and station-specific noise baselines as meta-knowledge of the stations. Specifically: in For station feature vectors, Station memory vector To rate the quality, This is an index of the region to which this station belongs. The number of stations.
[0041] Meta-knowledge memory: S2= + + .
[0042] When new waveform data for the target area appears, such as Figure 5 As shown, firstly, in the station memory subset , This represents the memorization of historical knowledge about the stations; Indicates the station's memory bank; Indicates the index of the target station; This involves retrieving station memory vectors. If the target region already has historical knowledge in the current station's memory, the station's adaptive meta-knowledge is directly utilized. If the target region has no historical knowledge in the memory, station-level retrieval is skipped, and the region memory subset retrieval begins. During the region memory subset retrieval stage, weights are assigned based on the similarity between the target region's features and other regions in the memory bank to determine the memory influence of each region. Through weighted calculation, the adaptive meta-knowledge representation of the target region in the current region is obtained. Specifically: in, Represents the query vector for the target region; Representation domain feature encoder; This represents the support set samples of the target region; This represents the query vector after graph enhancement; This represents the enhanced feature representation of the j-th region in the memory bank; Indicates attention weight; Indicates the scaling factor; Indicates the memory of regional historical knowledge; (·) represents a graph neural network; Represents a diagram showing the relationships between regions; This represents the region memory vector of the j-th region; This represents the feature representation of the j-th region in the memory bank; This indicates the confidence score.
[0043] If the target region differs significantly from other regions, retrieval is performed in the global memory subset. This stage primarily handles cases where the target region differs greatly from historical data. During global memory subset retrieval, the global common features of the target region are compared with those of other regions in the memory bank, and the global common meta-knowledge at the region level is integrated using a weighted average method. Specifically: in, This represents the attention weight for global retrieval; Represents the memory of global historical meta-knowledge; Represents the global feature vector; Indicates reliability score; This represents the global memory vector.
[0044] In the final meta-knowledge fusion stage, based on the retrieval results of each memory subset, the system automatically adjusts and merges the meta-knowledge representations of each subset. The system weights the memory meta-knowledge representations of different subsets according to reliability and importance, forming a comprehensive memory vector. Specifically: in, (.) represents the activation function.
[0045] By combining encoder S1, meta-knowledge memory S2, and decoder S3, a complete phase picking model is constructed. The feature representation obtained by encoder S1 is enhanced by meta-knowledge memory S2, and then decoded by decoder S3 to output the probability sequences of P-waves and S-waves.
[0046] The complete seismic phase picking model is fine-tuned by using a small number of labeled samples in the target area, and a small gradient update is used to achieve adaptive adjustment of the complete model to waveform data in the new area.
[0047] like Figure 6 As shown, when a new area or station needs to perform phase picking, the phase picking basic model encoder S1 is used on the seismic waveform data of the new area to obtain the time series characteristics: ; Based on the characteristics of the target region, region feature encoding E is performed to obtain the feature representation of each region. Using Q as an index, matching knowledge is sequentially retrieved from the meta-knowledge memory S2, and then fused with the confidence level to obtain the memory vector: in , , This indicates the weight of the memory meta-knowledge in the fusion process. This represents the retrieved historical metadata related to the features of the target region. Indicates sharpness, This represents a memory vector, which incorporates retrieved meta-knowledge for fusion.
[0048] memory vector Temporal features obtained from the base model Fusion is achieved through a memory enhancement module. In this process, Injected into the memory enhancement module, it is converted into a modulated signal using a gating mechanism: An attention mechanism is used to fuse the memory vector with temporal features to obtain an enhanced feature representation: The enhanced feature representations described above are decoded using decoder S3 to output the probability sequences of P-waves and S-waves: in, The enhanced feature representation indicates that the temporal resolution feature sequence has been recovered after upsampling.
[0049] With more and more new regions or sites being used for earthquake phase picking tasks, the knowledge memory needs to be updated with new memory vectors, such as... Figure 7 As shown, the enhanced feature representation is extracted from the memory enhancement module after fine-tuning the seismic phase picking of the new region or new site, and then transformed into a dimension reduction mapping. p 3D memory vectors are written into the meta-knowledge memory: in This indicates the memory meta-knowledge adapted to the new region after fine-tuning. Indicates vectorization operation, Represents a learnable matrix. Indicates bias. This represents the knowledge increment of the memory vector; This represents the knowledge increment after vectorization; This indicates that a new memory vector is written into the memory bank.
[0050] At the same time, a practicality and timeliness evaluation was conducted on all memory vectors in the memory bank: in Indicates the utility factor. This indicates the number of times the memory vector is used. This represents the average performance change when using the memory vector model. Indicates the time interval used. Indicates the aging decay factor. Represents the time constant. Indicates the current time value; This indicates the time elapsed since the last use.
[0051] A comprehensive forgetting score is constructed using the aforementioned practical and time-sensitive factors. When the memory bank reaches its capacity limit, entries with high forgetting scores are removed. in Indicates the weight of timeliness; Indicates the practicality weight; This represents the forgetting score of the k-th memory item.
[0052] Example 3 The present invention also provides an earthquake phase picking system based on a meta-knowledge memory base, the system being used to implement the method described in Embodiment 1, the system comprising: an acquisition module, an encoding module, a construction module, a retrieval module, a fusion module, and a decoding module; The acquisition module is used to acquire seismic waveform data; The encoding module is used to process seismic waveform data through the seismic phase picking base model encoder to obtain the base time series characteristics; Build modules are used to construct the meta-knowledge memory. The retrieval module is used to generate an index through regional feature encoding, retrieve it from the meta-knowledge memory base, and obtain a memory vector. The fusion module is used to fuse the basic temporal features and the memory vector using the memory enhancement module to obtain the enhanced feature representation; The decoding module is used to decode the enhanced feature representation using a decoder and output P-wave and S-wave probability sequences to complete the phase picking task.
[0053] In this embodiment, the process of constructing the meta-knowledge memory includes: By using graph neural networks to learn and fuse neighboring information based on the relationship graphs of different regions, the representation of earthquake region characteristics is enhanced, and a meta-knowledge memory is constructed.
[0054] In this embodiment, the process of generating an index through region feature encoding, retrieving it from the meta-knowledge memory base, and obtaining the memory vector includes: ; ; in, , , This indicates the weight of the memory meta-knowledge in the fusion process. This represents the retrieved historical metadata related to the features of the target region. Indicates sharpness, This represents a memory vector, which incorporates retrieved meta-knowledge for fusion. For the enhanced region feature representation, For the weighted attention coefficient, As a regional feature, It is a non-linear activation function. i and j They represent different regions. Indicates the region i Number of support set samples.
[0055] In this embodiment, the process of fusing the basic temporal features and the memory vector using the memory enhancement module to obtain the enhanced feature representation includes: ; ; in, Indicates a gating mechanism. For modulation signal, It is a temporal characteristic.
[0056] In this embodiment, the process of using a decoder to decode the enhanced feature representation and output P-wave and S-wave probability sequences to complete the phase picking task includes: ; ; in, Indicates wave type, h upThis represents the waveform sequence obtained after upsampling. This indicates the computation of the fully connected layer. This represents the probability sequence of P-waves and S-waves.
[0057] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for seismic phase picking based on a meta-knowledge memory base, characterized in that, The method includes: Acquire seismic waveform data; The seismic waveform data is processed by the basic model encoder of the phase picking to obtain the basic time series characteristics; Construct a meta-knowledge memory base; An index is generated by encoding regional features, and a memory vector is obtained by retrieving it from the meta-knowledge memory base. The memory enhancement module is used to fuse the basic temporal features and the memory vector to obtain the enhanced feature representation; The enhanced feature representation is decoded using a decoder to output P-wave and S-wave probability sequences, thus completing the phase picking task.
2. The method according to claim 1, characterized in that, Methods for constructing meta-knowledge memories include: By using graph neural networks to learn and fuse neighboring information based on the relationship graphs of different regions, the representation of earthquake region characteristics is enhanced, and a meta-knowledge memory is constructed.
3. The method according to claim 2, characterized in that, Methods for generating an index through region feature encoding and retrieving it from a meta-knowledge memory base to obtain a memory vector include: ； ； in, , , This indicates the weight of the memory meta-knowledge in the fusion process. This represents the retrieved historical metadata related to the characteristics of the target region. Indicates sharpness, This represents a memory vector, which incorporates retrieved meta-knowledge for fusion. For the enhanced region feature representation, For the weighted attention coefficient, As a regional feature, It is a non-linear activation function. i and j They represent different regions. Indicates the area i Number of support set samples.
4. The method according to claim 3, characterized in that, The method of fusing the basic temporal features and the memory vector using a memory enhancement module to obtain the enhanced feature representation includes: ； ； in, Indicates a gating mechanism. For modulation signal, It is a temporal characteristic.
5. The method according to claim 4, characterized in that, Methods for using a decoder to decode the enhanced feature representation and output P-wave and S-wave probability sequences to complete the phase picking task include: ； ； in, Indicates wave type, hup This represents the waveform sequence obtained after upsampling. This indicates the computation of the fully connected layer. Represents the probability sequences of P-waves and S-waves. l Indicates the number of floors.
6. A seismic phase picking system based on a meta-knowledge memory base, the system being used to implement the method described in any one of claims 1-5, characterized in that, The system includes: an acquisition module, an encoding module, a construction module, a retrieval module, a fusion module, and a decoding module; The acquisition module is used to acquire seismic waveform data; The encoding module is used to process seismic waveform data through a seismic phase picking basic model encoder to obtain basic time series characteristics; The construction module is used to construct a meta-knowledge memory library; The retrieval module is used to generate an index through regional feature encoding, retrieve it from the meta-knowledge memory base, and obtain a memory vector. The fusion module is used to fuse the basic temporal features and the memory vector using the memory enhancement module to obtain the enhanced feature representation; The decoding module is used to decode the enhanced feature representation using a decoder and output P-wave and S-wave probability sequences to complete the phase picking task.
7. The system according to claim 6, characterized in that, The process of constructing a meta-knowledge memory includes: By using graph neural networks to learn and fuse neighboring information based on the relationship graphs of different regions, the representation of earthquake region characteristics is enhanced, and a meta-knowledge memory is constructed.
8. The system according to claim 7, characterized in that, The process of generating an index through region feature encoding and retrieving it from the meta-knowledge memory base to obtain the memory vector includes: ； ； in, , , This indicates the weight of the memory meta-knowledge in the fusion process. This represents the retrieved historical metadata related to the characteristics of the target region. Indicates sharpness, This represents a memory vector, which incorporates retrieved meta-knowledge for fusion. For the enhanced region feature representation, For the weighted attention coefficient, As a regional feature, It is a non-linear activation function. i and j They represent different regions. Indicates the area i Number of support set samples.
9. The system according to claim 8, characterized in that, The process of fusing the basic temporal features and the memory vector using the memory enhancement module to obtain the enhanced feature representation includes: ； ； in, Indicates a gating mechanism. For modulation signal, It is a temporal characteristic.
10. The system according to claim 9, characterized in that, The process of using a decoder to decode the enhanced feature representation and output P-wave and S-wave probability sequences to complete the phase picking task includes: ； ； in, Indicates wave type, hup This represents the waveform sequence obtained after upsampling. This indicates the computation of the fully connected layer. Represents the probability sequences of P-waves and S-waves. l Indicates the number of floors.