Multi-source monitoring data analysis method and system based on earthquake active source

Abstract

The invention provides a multi-source monitoring data analysis method and system based on an earthquake active source, and relates to the technical field of artificial intelligence. A pre-trained earthquake wave strengthening analysis network is introduced, the earthquake wave strengthening analysis network is composed of a plurality of cascaded model function layers, and a multi-stage knowledge migration unit is embedded in each layer; and deep learning and efficient analysis of seismic wave data are realized. The multi-stage knowledge transfer learning mechanism not only can effectively extract key features in time period waveform data, but also gradually enhances the ability to understand and identify seismic wave sequences through layer-by-layer knowledge transfer and node fusion. The method not only improves the accuracy and robustness of seismic wave data analysis, but also significantly enhances the sensitivity and prediction precision of seismic monitoring. And more reliable and data support is provided for earthquake early warning, seismological research and earthquake disaster assessment by generating enhanced waveform data and an enhanced learning seismic wave sequence.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a multi-source monitoring data analysis method and system based on active seismic sources. Background Art

[0002] Earthquakes are a natural disaster, and their prediction and monitoring are of vital importance to reducing casualties and property losses. Traditional earthquake monitoring methods mainly rely on hardware equipment such as seismographs to directly record and analyze seismic waves. However, these methods are often limited by equipment accuracy, deployment density, and data processing capabilities, making it difficult to achieve comprehensive and accurate monitoring of earthquake activities.

[0003] With the advancement of science and technology, especially the rapid development of big data and artificial intelligence technology, earthquake monitoring data analysis methods have been significantly improved. Among them, multi-source monitoring technology based on active seismic sources has become an important means of earthquake monitoring. This technology actively stimulates seismic waves and uses multiple monitoring stations to simultaneously record the propagation of seismic waves, thereby obtaining richer and more comprehensive earthquake information.

[0004] However, efficient and accurate analysis of multi-source monitoring data remains an urgent problem to be solved. Traditional data analysis methods often have difficulty in effectively processing large-scale, multi-dimensional seismic wave data, resulting in omissions or misjudgments of key information. In addition, the complexity and nonlinear characteristics of seismic wave data also pose great challenges to data analysis. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a multi-source monitoring data analysis method based on seismic active sources, the method comprising:

[0006] Acquire waveform data of multiple time periods of candidate seismic wave sequences of multi-source monitoring data of seismic active sources, and for each time period of waveform data, load the waveform data of the time period into a pre-trained seismic wave enhancement analysis network, wherein the seismic wave enhancement analysis network includes multiple model function layers, each model function layer is cascaded, and each of the model function layers includes a multi-stage knowledge transfer unit, and the multi-stage knowledge transfer unit includes more than two stages of knowledge transfer learning;

[0007] In the multi-stage knowledge transfer unit of the target model function layer, multi-stage knowledge transfer learning is performed on the loading data of the target model function layer, and the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of each stage is node-fused to generate a multi-stage knowledge transfer learning vector corresponding to the target model function layer, and the multi-stage knowledge transfer learning vector is used to generate the reinforcement learning knowledge vector data corresponding to the target model function layer; wherein the first stage knowledge transfer learning is used to learn the loading data of the target model function layer, and the knowledge transfer learning of each remaining stage is used to learn the remaining node knowledge vector data except the key node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous stage; wherein, when the target model function layer is the first model function layer among the multiple model function layers, the loading data of the target model function layer is the basic knowledge vector data of the waveform data of the time period; when the target model function layer is the model function layer after the first model function layer among the multiple model function layers, the loading data of the target model function layer is the reinforcement learning knowledge vector data corresponding to the previous model function layer of the target model function layer;

[0008] Based on the reinforcement learning knowledge vector data corresponding to the last model functional layer, the reinforcement waveform data corresponding to the waveform data of the time period is generated; based on the reinforcement waveform data corresponding to each of the waveform data of the time period, the reinforcement learning seismic wave sequence corresponding to the candidate seismic wave sequence is generated.

[0009] On the other hand, an embodiment of the present invention also provides a multi-source monitoring data analysis system based on active seismic sources, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0010] Based on the above aspects, the embodiment of the present application introduces a pre-trained seismic wave reinforcement analysis network, which is composed of multiple cascaded model function layers, and each layer is embedded with a multi-stage knowledge transfer unit, which realizes deep learning and efficient analysis of seismic wave data. The multi-stage knowledge transfer learning mechanism can not only effectively extract the key features in the waveform data of the time period, but also gradually enhance the understanding and recognition ability of the seismic wave sequence through layer-by-layer knowledge transfer and node fusion. This method not only improves the accuracy and robustness of seismic wave data analysis, but also significantly enhances the sensitivity and prediction accuracy of earthquake monitoring. By generating enhanced waveform data and reinforcement learning seismic wave sequences, more reliable and data support is provided for earthquake early warning, seismological research and earthquake disaster assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the execution flow of the multi-source monitoring data analysis method based on active seismic sources provided by an embodiment of the present invention.

[0012] Figure 2 It is a schematic diagram of the hardware architecture of a multi-source monitoring data analysis system based on active seismic sources provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a multi-source monitoring data analysis method based on active seismic sources provided by an embodiment of the present invention. The multi-source monitoring data analysis method based on active seismic sources is introduced in detail below.

[0014] Step S110, obtain waveform data of multiple time periods of candidate seismic wave sequences of multi-source monitoring data of seismic active sources, and for each time period of waveform data, load the waveform data of the time period into a pre-trained seismic wave enhancement analysis network, wherein the seismic wave enhancement analysis network includes multiple model function layers, each model function layer is cascaded, and each of the model function layers includes a multi-stage knowledge transfer unit, and the multi-stage knowledge transfer unit includes more than two stages of knowledge transfer learning.

[0015] In this embodiment, in the scenario of earthquake monitoring, the active seismic source refers to a device that obtains underground structure information by artificially exciting seismic waves, and can stimulate seismic waves at different locations and times to obtain multi-source monitoring data. These source monitoring data contain a large number of candidate seismic wave sequences, and each candidate seismic wave sequence is composed of waveform data of multiple time periods.

[0016] For example, in a specific earthquake monitoring area, multiple active seismic source transmitting devices are set up. These devices send seismic wave signals underground according to the predetermined time interval and excitation method. When these seismic waves propagate underground and reflect back, they are received by multiple seismic monitoring sensors distributed in the area. The data collected by the server from these sensors is the multi-source monitoring data of the active seismic source.

[0017] There are many candidate seismic wave sequences in these multi-source monitoring data. Taking a candidate seismic wave sequence as an example, it may contain waveform data of different periods in the complete process of seismic waves from excitation to propagation, reflection and reception. These waveform data of different periods reflect the characteristics of seismic waves at different times, such as amplitude, frequency, phase and other information.

[0018] After the server obtains the waveform data of these time periods, it loads the waveform data of each time period into the pre-trained seismic wave enhancement analysis network. This seismic wave enhancement analysis network is a complex neural network structure, which consists of multiple model function layers. Each model function layer is cascaded. And each model function layer contains a multi-stage knowledge transfer unit, which has more than two stages of knowledge transfer learning capabilities. This multi-stage knowledge transfer learning capability allows the network to better utilize existing knowledge and experience when processing time period waveform data, thereby improving the analysis and processing capabilities of seismic wave data.

[0019] For example, suppose that the seismic wave enhancement analysis network has three model functional layers, namely functional layer A, functional layer B, and functional layer C. When the server obtains a period of waveform data, it first loads the period of waveform data into functional layer A. The multi-stage knowledge transfer unit in functional layer A will begin to perform preliminary knowledge transfer learning on the period of waveform data, and use its internal multi-stage knowledge transfer learning mechanism to analyze the characteristics of the period of waveform data from different angles and levels, laying the foundation for subsequent processing and analysis.

[0020] Step S120, in the multi-stage knowledge transfer unit of the target model function layer, multi-stage knowledge transfer learning is performed on the loading data of the target model function layer, and the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of each stage is node-fused to generate a multi-stage knowledge transfer learning vector corresponding to the target model function layer, and the multi-stage knowledge transfer learning vector is used to generate the reinforcement learning knowledge vector data corresponding to the target model function layer. Among them, the first stage knowledge transfer learning is used to learn the loading data of the target model function layer, and the knowledge transfer learning of each remaining stage is used to learn the remaining node knowledge vector data except the key node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the previous stage knowledge transfer learning. Among them, when the target model function layer is the first model function layer among the multiple model function layers, the loading data of the target model function layer is the basic knowledge vector data of the waveform data of the period. When the target model function layer is a model function layer after the first model function layer among the multiple model function layers, the loading data of the target model function layer is the reinforcement learning knowledge vector data corresponding to the previous model function layer of the target model function layer.

[0021] In this embodiment, when operating in the multi-stage knowledge migration unit of the target model functional layer, it is assumed that the target model functional layer is functional layer B (in the aforementioned seismic wave enhancement analysis network composed of functional layer A, functional layer B and functional layer C).

[0022] First, since functional layer B is not the first model functional layer (the first is functional layer A), its loaded data is the reinforcement learning knowledge vector data corresponding to functional layer A. After processing the period waveform data, functional layer A generates reinforcement learning knowledge vector data, which contains various feature information and learned knowledge about the period waveform data after being processed by functional layer A.

[0023] In the multi-stage knowledge transfer unit of the functional layer B, multi-stage knowledge transfer learning begins. Assume that this multi-stage knowledge transfer unit has three stages of knowledge transfer learning, namely, stage one, stage two, and stage three.

[0024] In stage 1, functional layer B performs first-order knowledge transfer learning on the reinforcement learning knowledge vector data obtained from functional layer A. For example, this first-order knowledge transfer learning in functional layer B may further refine and analyze certain features of seismic waves output by functional layer A (such as energy distribution in a specific frequency range) to generate a first-order knowledge transfer learning vector. This first-order knowledge transfer learning vector contains the new feature information and knowledge learned in this stage.

[0025] Next, we enter the second stage, which is the transition knowledge transfer unit. In this stage, we first need to obtain the remaining node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level (that is, stage one). Node splitting is to decompose the knowledge transfer learning vector generated in stage one into different node knowledge vector data according to certain rules. For example, assuming that the knowledge transfer learning vector of stage one contains comprehensive information about the frequency, amplitude and phase of seismic waves, it can be split into node knowledge vector data about frequency, node knowledge vector data about amplitude and node knowledge vector data about phase through node splitting. Then, knowledge transfer learning is performed on these remaining node knowledge vector data (here, it is assumed that the node knowledge vector data about amplitude and phase except for some key nodes). This process is similar to in-depth learning of part of the content in a certain knowledge system during the learning process. Through the learning of this stage, a transition knowledge transfer learning vector is generated, which contains more in-depth knowledge about seismic wave amplitude and phase learned in this transition stage.

[0026] Finally, in stage three, which is the last knowledge transfer unit, first obtain the remaining node knowledge vector data generated by node splitting of the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level (stage two). Then, the last stage of knowledge transfer learning is performed on these remaining node knowledge vector data (assuming that they are more detailed partial node knowledge vector data about the phase) to generate the key node knowledge vector data generated by the last stage of knowledge transfer learning. This key node knowledge vector data may contain the most critical feature information about the seismic wave phase learned at this stage.

[0027] Then, the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning in each stage (stage one, stage two and stage three) is node fused. Node fusion can be understood as combining the most important parts of knowledge learned in each stage to form a complete and more valuable knowledge set. In this way, a multi-stage knowledge transfer learning vector corresponding to functional layer B is generated. This multi-stage knowledge transfer learning vector is the comprehensive knowledge result of functional layer B after multi-stage knowledge transfer learning, which will be used to generate reinforcement learning knowledge vector data corresponding to functional layer B.

[0028] Assume that the target model functional layer is functional layer A (i.e., the first model functional layer). At this time, the loading data of the target model functional layer is the basic knowledge vector data of the period waveform data. The acquisition of this basic knowledge vector data is achieved through the regularization conversion unit and the node expansion unit in the seismic wave enhancement analysis network.

[0029] For example, the server first performs regular conversion on the time period waveform data through a regular conversion unit. Assume that the time period waveform data is an irregular sequence of seismic wave amplitude and frequency data containing various noises and interferences. The regular conversion unit generates regular waveform data according to predetermined rules, such as adjusting the amplitude value range to a specific interval, normalizing the frequency data, etc. Then, the regular waveform data is node expanded through the node expansion unit. For example, the regular waveform data originally only contains the amplitude and frequency values ​​at a few simple time points. Through the node expansion unit, more nodes can be added between these time points according to a certain algorithm (such as an interpolation algorithm), thereby generating basic knowledge vector data containing more detailed information.

[0030] In the multi-stage knowledge transfer unit of functional layer A, it is assumed that it has two stages of knowledge transfer learning. In the first stage of knowledge transfer learning, functional layer A directly learns the basic knowledge vector data, learns the basic characteristics of the waveform data of this period, such as the initial amplitude change trend, frequency distribution range, etc., and generates the knowledge transfer learning vector of the first stage. Then in the second stage of knowledge transfer learning, the first stage knowledge transfer learning vector is node-split to obtain the remaining node knowledge vector data except the key node knowledge vector data, and then learn these remaining node knowledge vector data. Finally, the key node knowledge vector data of the two stages are node-fused to generate the multi-stage knowledge transfer learning vector corresponding to functional layer A. This multi-stage knowledge transfer learning vector will be used to generate the reinforcement learning knowledge vector data corresponding to functional layer A.

[0031] Step S130, based on the reinforcement learning knowledge vector data corresponding to the last model functional layer, generate the reinforcement waveform data corresponding to the waveform data of the time period, and based on the reinforcement waveform data corresponding to each of the waveform data of the time period, generate the reinforcement learning seismic wave sequence corresponding to the candidate seismic wave sequence.

[0032] In the process of generating enhanced waveform data corresponding to the time period waveform data based on the reinforcement learning knowledge vector data corresponding to the last model functional layer, it is assumed that the last model functional layer is functional layer C (in the aforementioned seismic wave enhancement analysis network composed of functional layer A, functional layer B and functional layer C).

[0033] After the previous multi-stage knowledge transfer learning and related processing, the functional layer C generates the corresponding reinforcement learning knowledge vector data. This data contains rich information about the period waveform data after being processed by the entire seismic wave reinforcement analysis network.

[0034] For example, the server wants to generate enhanced waveform data corresponding to the time period waveform data. First, the server combines the basic knowledge vector data of the time period waveform data and the reinforcement learning knowledge vector data corresponding to the functional layer C. Assume that the basic knowledge vector data of the time period waveform data contains some basic feature information of the initial seismic wave (such as the original amplitude and frequency information), and the reinforcement learning knowledge vector data corresponding to the functional layer C contains more accurate and detailed feature information about the seismic wave after the entire network processing (such as the corrected amplitude and frequency information, more accurate phase information, etc.). Combine these two vector data together to generate the target reinforcement learning knowledge vector data.

[0035] Then, the enhanced waveform data corresponding to the time period waveform data is determined based on the target enhanced learning knowledge vector data. In this process, the seismic wave enhanced analysis network also includes an inverse regularization conversion unit and a node compression unit.

[0036] The server first performs node compression on the target reinforcement learning knowledge vector data through the node compression unit. Assuming that the target reinforcement learning knowledge vector data is a vector containing a large amount of detailed node information after multi-layer processing, the node compression unit will compress the number of nodes in this vector to the same number of nodes as the period waveform data according to a certain algorithm based on factors such as the original number of nodes in the period waveform data. For example, if the original period waveform data has 100 nodes, and the target reinforcement learning knowledge vector data has 200 nodes after processing, the node compression unit will select the most important 100 nodes or merge and compress the 200 nodes into 100 nodes to generate the target knowledge vector data.

[0037] Next, the target knowledge vector data is subjected to inverse regularization conversion by the inverse regularization conversion unit. Inverse regularization conversion is the inverse process of regularization conversion. If in step S110, the regularization conversion unit adjusts the amplitude value range to a specific interval and normalizes the frequency data on the original time period waveform data, then the inverse regularization conversion unit will restore the amplitude and frequency data in the target knowledge vector data to a representation similar to the original time period waveform data, thereby generating enhanced waveform data corresponding to the time period waveform data.

[0038] Finally, based on the enhanced waveform data corresponding to the waveform data of each time period, a reinforcement learning seismic wave sequence corresponding to the candidate seismic wave sequence is generated. For example, for a candidate seismic wave sequence containing waveform data of 10 time periods, the server generates corresponding enhanced waveform data for each time period waveform data according to the above process. Then these 10 enhanced waveform data are combined together in the original order to generate the reinforcement learning seismic wave sequence corresponding to the candidate seismic wave sequence. Compared with the original candidate seismic wave sequence, this reinforcement learning seismic wave sequence has more accurate and reliable seismic wave characteristic information, and can be better used for earthquake-related analysis and research, such as the study of seismic wave propagation characteristics and the inference of underground geological structures.

[0039] Based on the above steps, the embodiment of the present application introduces a pre-trained seismic wave enhanced analysis network, which is composed of multiple cascaded model function layers, and each layer is embedded with a multi-stage knowledge transfer unit, which realizes deep learning and efficient analysis of seismic wave data. The multi-stage knowledge transfer learning mechanism can not only effectively extract the key features in the period waveform data, but also gradually enhance the understanding and recognition ability of seismic wave sequences through layer-by-layer knowledge transfer and node fusion. This method not only improves the accuracy and robustness of seismic wave data analysis, but also significantly enhances the sensitivity and prediction accuracy of earthquake monitoring. By generating enhanced waveform data and strengthening learning seismic wave sequences, more reliable and data support is provided for earthquake early warning, seismological research and earthquake disaster assessment.

[0040] In a possible implementation, the seismic wave enhancement analysis network further includes a regularization conversion unit and a node expansion unit, and the step of determining the basic knowledge vector data of the time period waveform data includes:

[0041] The regularized conversion unit performs regularized conversion on the time period waveform data to generate regularized waveform data.

[0042] The node expansion unit performs node expansion on the regularized waveform data to generate basic knowledge vector data of the time period waveform data.

[0043] In this embodiment, firstly, the period waveform data is regularized and converted by the regularization conversion unit to generate regularized waveform data. For example, after obtaining the period waveform data of the candidate seismic wave sequence in the multi-source monitoring data of the seismic active source, these period waveform data may have various different representation forms and value ranges. The regularization conversion unit will process these data according to predefined rules. If the value range of the amplitude data in the period waveform data is between -100 and 100, and the amplitude value range of the waveform data in different periods is relatively different, the regularization conversion unit will uniformly convert it to a value range of 0 to 1, so that the amplitude data of the waveform data in different periods are comparable. For frequency data, if it is in Hertz and the value is relatively scattered, the regularization conversion unit will normalize it according to a certain proportional relationship, so that the frequency data also follows specific rules. After these operations, the original period waveform data is converted into regularized waveform data, which is more regular in data format and value range, and is convenient for subsequent processing.

[0044] Next, the node expansion unit is used to perform node expansion on the regularized waveform data to generate the basic knowledge vector data of the period waveform data. Since the regularized waveform data may not have dense data points after the regularization conversion, the node expansion unit will supplement it. For example, the regularized waveform data records a data point every 1 second on the time axis, which may not be detailed enough for seismic wave analysis. The node expansion unit will insert new data points between adjacent data points according to certain algorithms, such as linear interpolation algorithms or interpolation algorithms based on seismic wave propagation models. If the original regularized waveform data has 10 data points in a certain period, it may increase to 20 data points after node expansion. These newly added data points and the original data points together constitute a richer information set. After further processing and organization, this regularized waveform data containing more data points forms the basic knowledge vector data of the period waveform data. This basic knowledge vector data contains more detailed and comprehensive information about the period waveform data, which is the basic data for subsequent multi-stage knowledge transfer learning in the first model function layer of the seismic wave enhancement analysis network. In this process, the server processes the data strictly according to the functional logic of the regularized conversion unit and the node expansion unit to ensure the accuracy and effectiveness of the generated basic knowledge vector data, so as to provide a reliable data foundation for the entire seismic wave enhancement analysis process.

[0045] In a possible implementation, when the target model function layer is a model function layer subsequent to the first model function layer among the multiple model function layers, step S120 includes:

[0046] Step S121, in the first-order knowledge transfer unit of the multi-stage knowledge transfer unit, first-order knowledge transfer learning is performed on the reinforcement learning knowledge vector data corresponding to the previous model function layer of the target model function layer to generate a first-order knowledge transfer learning vector.

[0047] Step S122, in the transition knowledge transfer unit of the multi-stage knowledge transfer unit, obtain the remaining node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level, perform knowledge transfer learning on the remaining node knowledge vector data, and generate a transition knowledge transfer learning vector.

[0048] Step S123, in the last knowledge transfer unit of the multi-stage knowledge transfer unit, obtain the remaining node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level, perform the last stage of knowledge transfer learning on the remaining node knowledge vector data, and generate the key node knowledge vector data generated by the last stage of knowledge transfer learning.

[0049] Step S124, performing node fusion on the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning in each stage, and generating a multi-stage knowledge transfer learning vector corresponding to the functional layer of the target model.

[0050] In a possible implementation, based on the above step S121 and step S124, the method further includes:

[0051] Step A110, performing node splitting on the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level, and generating node knowledge vector data of each feature node.

[0052] Step A120, based on the arrangement order of each feature node, starting from the first feature node, the node knowledge vector data of the target number of nodes with a later node order are used as the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level, and the node knowledge vector data of the remaining nodes are used as the remaining node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level.

[0053] In this embodiment, first, in the first-order knowledge transfer unit of the multi-stage knowledge transfer unit, the server will perform first-order knowledge transfer learning on the reinforcement learning knowledge vector data corresponding to the previous model function layer of the target model function layer, thereby generating a first-order knowledge transfer learning vector. Taking the functional layers A, B and C in the seismic wave reinforcement analysis network mentioned above as an example, when the target model function layer is function layer B, its previous model function layer is function layer A. After a series of processing, functional layer A generates reinforcement learning knowledge vector data, which contains various information after functional layer A processes the period waveform data, such as characteristic information after adjustment and analysis of the amplitude, frequency, phase, etc. of the seismic wave. The first-order knowledge transfer unit in functional layer B will receive the reinforcement learning knowledge vector data from functional layer A. First-order knowledge transfer learning is a process of in-depth mining and analysis of these data. It will analyze, extract and recombine the various elements in the reinforcement learning knowledge vector data of functional layer A according to the learning mechanism and parameter settings of functional layer B itself. For example, if the reinforcement learning knowledge vector data of functional layer A contains information such as the amplitude change trend of seismic waves in different time periods and the distribution range of frequencies, the first-order knowledge transfer unit may focus on the new representation and potential relationship of this information in functional layer B, thereby generating a first-order knowledge transfer learning vector. This first-order knowledge transfer learning vector may contain the characteristic information about seismic waves after preliminary adjustment and reorganization, providing a new starting point for subsequent learning.

[0054] Next, in the transition knowledge transfer unit of the multi-stage knowledge transfer unit, the server needs to obtain the remaining node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level, and then perform knowledge transfer learning on these remaining node knowledge vector data to generate transition knowledge transfer learning vectors. Taking functional layer B as an example, the previous level is the first-order knowledge transfer learning vector generated by the first-order knowledge transfer unit. This vector contains a variety of characteristic information about seismic waves, and the server will perform node splitting on it. Node splitting is to decompose the data in this vector according to different characteristics or attributes according to pre-set rules. For example, the first-order knowledge transfer learning vector contains relevant information such as amplitude, frequency, phase and propagation time of seismic waves, and node splitting may split this information into different node knowledge vector data, such as amplitude-related node knowledge vector data, frequency-related node knowledge vector data, etc. Then, according to a specific algorithm or rule, a part of the node knowledge vector data is determined as the key node knowledge vector data, and the rest is the remaining node knowledge vector data. For these remaining node knowledge vector data, the transition knowledge transfer unit will perform knowledge transfer learning. This learning process will use the specific learning algorithm and parameters of the transition knowledge transfer unit in the functional layer B to further analyze and process the information in the remaining node knowledge vector data. For example, for the remaining node knowledge vector data related to frequency, it may be adjusted and optimized based on some general laws about frequency learned in the entire seismic wave enhancement analysis network and the characteristics of the waveform data in the current period, so as to generate a transition knowledge transfer learning vector. This vector contains the results of learning the remaining node knowledge vector data in the transition knowledge transfer unit, and it focuses more on certain specific seismic wave feature information to a certain extent.

[0055] Then, in the last knowledge transfer unit of the multi-stage knowledge transfer unit, the server also needs to obtain the remaining node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the previous level of knowledge transfer learning, and then perform the last stage of knowledge transfer learning on these remaining node knowledge vector data, and then generate the key node knowledge vector data generated by the last stage of knowledge transfer learning. Continuing with functional layer B as an example, the previous level is the transition knowledge transfer learning vector generated by the transition knowledge transfer unit. This vector will be node split again to obtain different remaining node knowledge vector data. These data will be received by the last knowledge transfer unit, and then the last stage of knowledge transfer learning will be performed. This learning process will perform a final deep mining and adjustment on these remaining node knowledge vector data based on the special algorithm of the last knowledge transfer unit in functional layer B and the previously accumulated knowledge about seismic waves. For example, if the previous transition knowledge transfer learning vector obtains the remaining node knowledge vector data about the change of seismic wave phase in a specific time period after node splitting, the last knowledge transfer unit may accurately analyze and adjust the remaining node knowledge vector data according to the overall cognition of the entire seismic wave enhancement analysis network on phase change and the special requirements of the waveform data in the current period, thereby generating the key node knowledge vector data generated by the last stage of knowledge transfer learning. This key node knowledge vector data contains the very critical information about seismic wave characteristics learned in the last stage.

[0056] Finally, the server performs node fusion on the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning in each stage, thereby generating a multi-stage knowledge transfer learning vector corresponding to the functional layer of the target model. For example, for functional layer B, the key node knowledge vector data in the first-order knowledge transfer learning vector, the key node knowledge vector data in the transition knowledge transfer learning vector, and the key node knowledge vector data generated by the last stage of knowledge transfer learning are node fused. This fusion process is not a simple addition, but an organic combination based on the importance, relevance and other factors of each key node knowledge vector data. Through a specific algorithm, the information in these key node knowledge vector data is integrated together to form a complete multi-stage knowledge transfer learning vector that integrates the key information of each stage. This vector is an important result of the multi-stage knowledge transfer learning of functional layer B. It will serve as the basis for further processing to generate the reinforcement learning knowledge vector data corresponding to functional layer B.

[0057] In this process, the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level is also split into nodes to generate node knowledge vector data of each feature node, and the node knowledge vector data of the target number of nodes with a later node order are used as the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level according to the arrangement order of each feature node, starting from the first feature node, and the node knowledge vector data of the remaining nodes are used as the remaining node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level. Taking the first-order knowledge transfer learning vector in the functional layer B as an example, this vector contains information of multiple feature nodes, such as the amplitude feature nodes, frequency feature nodes, phase feature nodes, etc. of the seismic wave arranged in a certain order. Assume that according to a predetermined rule, the node knowledge vector data of a certain number (target number) of nodes with a later node order are used as key node knowledge vector data. If there are 10 characteristic nodes in the first-order knowledge transfer learning vector, the node knowledge vector data of the three nodes with the later node order (such as characteristic nodes related to phase, propagation time, and energy attenuation) are specified as key node knowledge vector data, and the node knowledge vector data of the remaining seven nodes (such as characteristic nodes related to amplitude, frequency, etc.) are the remaining node knowledge vector data. This method of determining the key node knowledge vector data and the remaining node knowledge vector data is helpful to process the node information of different importance in a targeted manner in the knowledge transfer learning at different stages, thereby improving the efficiency and effect of the entire multi-stage knowledge transfer learning, ensuring that the multi-stage knowledge transfer learning vector corresponding to the target model functional layer generated in the end can accurately reflect the characteristic information of the seismic wave, and provide strong support for the subsequent seismic wave analysis and processing.

[0058] In a possible implementation, each model functional layer also includes a node nonlinear conversion unit based on a feature self-focusing strategy, and the method also includes: loading the multi-stage knowledge transfer learning vector generated by the multi-stage knowledge transfer unit in the target model functional layer into the node nonlinear conversion unit in the target model functional layer, performing node nonlinear conversion processing on the multi-stage knowledge transfer learning vector according to the feature self-focusing strategy, and generating a multi-stage knowledge transfer learning vector for node nonlinear conversion.

[0059] The step of generating the reinforcement learning knowledge vector data corresponding to the target model functional layer includes: in the target model functional layer, based on the multi-stage knowledge transfer learning vector corresponding to the target model functional layer, the multi-stage knowledge transfer learning vector of the node nonlinear conversion and the reinforcement learning knowledge vector data corresponding to the previous model functional layer of the target model functional layer, generating the reinforcement learning knowledge vector data corresponding to the target model functional layer.

[0060] In a possible implementation, the multi-stage knowledge transfer learning vector corresponding to the target model function layer includes node knowledge transfer learning vectors of each feature node, and performing node nonlinear conversion processing on the multi-stage knowledge transfer learning vector according to the feature self-focusing strategy to generate a multi-stage knowledge transfer learning vector of node nonlinear conversion includes:

[0061] The multi-stage knowledge transfer learning vector is subjected to a focusing process using a feature self-focusing strategy, and the attention weights corresponding to the nodes in the multi-stage knowledge transfer learning vector are determined.

[0062] The attention weights and node knowledge transfer learning vectors corresponding to each feature node are respectively multiplied to generate a multi-stage knowledge transfer learning vector for node nonlinear transformation.

[0063] In a possible implementation, the generating of the reinforcement learning knowledge vector data corresponding to the target model functional layer based on the multi-stage knowledge transfer learning vector corresponding to the target model functional layer, the multi-stage knowledge transfer learning vector of the node nonlinear conversion, and the reinforcement learning knowledge vector data corresponding to the previous model functional layer of the target model functional layer includes:

[0064] The multi-stage knowledge transfer learning vector corresponding to the functional layer of the target model and the multi-stage knowledge transfer learning vector of the node nonlinear transformation are combined to generate combined knowledge vector data.

[0065] Perform node conversion on the combined knowledge vector data to generate converted knowledge vector data, and combine the converted knowledge vector data with the reinforcement learning knowledge vector data corresponding to the previous model function layer of the target model function layer to generate the reinforcement learning knowledge vector data corresponding to the target model function layer.

[0066] In this embodiment, first, the server loads the multi-stage knowledge transfer learning vector generated by the multi-stage knowledge transfer unit in the target model functional layer into the node nonlinear conversion unit in the target model functional layer, and then performs node nonlinear conversion processing on the multi-stage knowledge transfer learning vector according to the feature self-focusing strategy to generate a multi-stage knowledge transfer learning vector for node nonlinear conversion. Taking the functional layer in the seismic wave enhancement analysis network mentioned above as an example, it is assumed that the target model functional layer is functional layer B. After a series of knowledge transfer learning, the multi-stage knowledge transfer unit in functional layer B generates a multi-stage knowledge transfer learning vector. This multi-stage knowledge transfer learning vector contains rich feature information about seismic waves, which exists in the form of node knowledge transfer learning vectors of each feature node. When this multi-stage knowledge transfer learning vector is loaded into the node nonlinear conversion unit of functional layer B, it begins to be processed according to the feature self-focusing strategy.

[0067] In this process, the multi-stage knowledge transfer learning vector is focused by the feature self-focusing strategy to determine the attention weights corresponding to each node in the multi-stage knowledge transfer learning vector. For example, the feature nodes in the multi-stage knowledge transfer learning vector may include the amplitude feature node, frequency feature node, phase feature node, and propagation time feature node of the seismic wave. For the amplitude feature node, the server will analyze the importance of this node in the entire seismic wave analysis according to the feature self-focusing strategy. If the change in amplitude is of high importance for the identification and analysis of seismic waves in the current period of waveform data, then this amplitude feature node will be given a higher attention weight. The determination of this attention weight is based on the knowledge about seismic wave characteristics accumulated by the seismic wave enhancement analysis network in the previous learning and processing process, as well as the characteristics of the waveform data in the current period. Similarly, for feature nodes such as frequency, phase, and propagation time, corresponding attention weights will be given according to their respective importance.

[0068] Next, the server performs dot multiplication on the attention weights and node knowledge transfer learning vectors corresponding to each feature node, thereby generating a multi-stage knowledge transfer learning vector for node nonlinear transformation. For example, for the amplitude feature node, assuming that its attention weight is 0.8, its node knowledge transfer learning vector contains a series of numerical information about the amplitude, such as the amplitude value in different time periods. Perform dot multiplication on this attention weight 0.8 and each value in the node knowledge transfer learning vector of the amplitude feature node to obtain the numerical information about the amplitude after attention adjustment. The same operation is performed on other feature nodes such as frequency, phase, and propagation time, and the respective attention weights are dot multiplied with the corresponding node knowledge transfer learning vector. Finally, the feature node information after dot multiplication is combined to generate a multi-stage knowledge transfer learning vector for node nonlinear transformation. Compared with the original multi-stage knowledge transfer learning vector, this vector highlights the information of important feature nodes more, and the representation of seismic wave characteristics is more in line with the key needs of current analysis.

[0069] Afterwards, in the target model functional layer, the reinforcement learning knowledge vector data corresponding to the target model functional layer is generated based on the multi-stage knowledge transfer learning vector corresponding to the target model functional layer, the multi-stage knowledge transfer learning vector of the node nonlinear conversion, and the reinforcement learning knowledge vector data corresponding to the previous model functional layer of the target model functional layer. Still taking functional layer B as an example, the multi-stage knowledge transfer learning vector corresponding to functional layer B contains the characteristic information of the seismic wave after multi-stage knowledge transfer learning, and the multi-stage knowledge transfer learning vector of the node nonlinear conversion is the information that focuses more on important features after the node nonlinear conversion processing. The reinforcement learning knowledge vector data corresponding to the previous model functional layer of functional layer B, namely functional layer A, contains the result information of the seismic wave processing by functional layer A.

[0070] The server first combines the multi-stage knowledge transfer learning vector corresponding to the target model functional layer and the multi-stage knowledge transfer learning vector of the node nonlinear conversion to generate the combined knowledge vector data. For example, the information about the amplitude, frequency, phase, etc. of the seismic wave in the multi-stage knowledge transfer learning vector corresponding to the functional layer B and the information about the amplitude, frequency, phase, etc. after attention adjustment in the multi-stage knowledge transfer learning vector of the node nonlinear conversion are combined according to the corresponding feature nodes. If the amplitude information in the multi-stage knowledge transfer learning vector is [1, 2, 3], and the amplitude information in the multi-stage knowledge transfer learning vector of the node nonlinear conversion is [0.8, 1.6, 2.4] after attention adjustment, the combined amplitude information may be [1.8, 3.6, 5.4] (this is only a simple example, and the actual combination method may be more complicated). The same combination operation is performed on other feature nodes such as frequency and phase to obtain the combined knowledge vector data.

[0071] Then, the server performs node conversion on the combined knowledge vector data to generate converted knowledge vector data. Node conversion is to adjust each node in the combined knowledge vector data according to the conversion rules within the functional layer B. For example, for the amplitude information [1.8, 3.6, 5.4] in the combined knowledge vector data, it may be converted to [2, 4, 6] according to specific conversion rules (this is just an example). Other feature nodes such as frequency and phase are also converted accordingly to obtain the converted knowledge vector data. Finally, the server combines the converted knowledge vector data with the reinforcement learning knowledge vector data corresponding to the previous model functional layer of the target model functional layer to generate the reinforcement learning knowledge vector data corresponding to the target model functional layer. For example, the amplitude, frequency, phase and other information in the converted knowledge vector data of functional layer B and the relevant information in the reinforcement learning knowledge vector data corresponding to functional layer A are combined according to certain rules. If the amplitude information in the reinforcement learning knowledge vector data corresponding to functional layer A is [0.5, 1, 1.5], and the amplitude information in the knowledge vector data after conversion of functional layer B is [2, 4, 6], then the combined amplitude information may be [2.5, 5, 7.5] (this is just an example). The same combination operation is performed on other characteristic nodes such as frequency and phase, and finally the reinforcement learning knowledge vector data corresponding to functional layer B is obtained. This reinforcement learning knowledge vector data contains the complete analysis information about seismic waves of functional layer B after comprehensively considering its own multi-stage knowledge transfer learning, node nonlinear conversion and the processing results of functional layer A, providing a more accurate and comprehensive basis for the subsequent seismic wave analysis process. As a result, in each model functional layer of the seismic wave enhancement analysis network, the ability to understand and analyze seismic wave characteristics is gradually improved, so that seismic wave data can be processed more effectively, providing strong support for earthquake research, monitoring and other related work.

[0072] In a possible implementation, step S130 includes: combining the basic knowledge vector data of the time period waveform data and the reinforcement learning knowledge vector data corresponding to the last model function layer to generate target reinforcement learning knowledge vector data, and determining the reinforcement waveform data corresponding to the time period waveform data based on the target reinforcement learning knowledge vector data.

[0073] The seismic wave enhancement analysis network further includes an inverse regularization conversion unit and a node compression unit, and the enhanced waveform data corresponding to the waveform data of the time period is determined according to the target reinforcement learning knowledge vector data, including:

[0074] The target reinforcement learning knowledge vector data is node compressed by the node compression unit to generate target knowledge vector data, and the number of nodes of the target knowledge vector data is the same as the number of nodes of the period waveform data.

[0075] The target knowledge vector data is subjected to an inverse regularization conversion by the inverse regularization conversion unit to generate enhanced waveform data corresponding to the time period waveform data.

[0076] In this embodiment, the basic knowledge vector data of the period waveform data and the reinforcement learning knowledge vector data corresponding to the last model functional layer are first combined to generate the target reinforcement learning knowledge vector data. Take the functional layers A, B and C in the seismic wave reinforcement analysis network as an example, assuming that the functional layer C is the last model functional layer. The basic knowledge vector data of the period waveform data is obtained after the period waveform data is initially regularized and expanded. It contains relatively basic but comprehensive information about the period waveform data, such as the detailed vector representation of the amplitude, frequency, phase and other information of the seismic wave in the initial state. The reinforcement learning knowledge vector data corresponding to the functional layer C is the high-level feature information about the seismic wave obtained after the comprehensive processing of the multi-stage knowledge transfer learning, node nonlinear conversion and the previous functional layer (functional layer B) related information within the functional layer C. The combination of these two vector data is not a simple addition, but the corresponding seismic wave feature information in the two vector data is integrated according to the rules pre-set by the seismic wave reinforcement analysis network. For example, the amplitude information in the basic knowledge vector data may be the amplitude values ​​of the original seismic waves at different times without in-depth analysis, while the amplitude information in the reinforcement learning knowledge vector data may be the result of adjustment, supplementation or re-evaluation of these amplitude values ​​after analysis by functional layer C. When combined, these two types of information about amplitude will be fused according to a certain algorithm, such as weighted summation according to their respective weights or merging according to specific logical rules. Similarly, similar operations are performed on other seismic wave characteristic information such as frequency and phase, and finally the target reinforcement learning knowledge vector data is generated. This target reinforcement learning knowledge vector data combines the basic information of the time period waveform data and the advanced information after in-depth processing by the last functional layer of the entire seismic wave enhancement analysis network, providing a comprehensive data basis for determining the enhanced waveform data corresponding to the time period waveform data.

[0077] Next, the server determines the enhanced waveform data corresponding to the time period waveform data based on the target enhanced learning knowledge vector data. This process involves the inverse regularization conversion unit and the node compression unit in the seismic wave enhanced analysis network.

[0078] The server first performs node compression on the target reinforcement learning knowledge vector data through a node compression unit to generate the target knowledge vector data. After the previous combination step, the target reinforcement learning knowledge vector data may contain more node information than the original period waveform data, which is generated for detailed analysis and processing in the seismic wave enhancement analysis network. However, in order to generate enhanced waveform data corresponding to the period waveform data, the number of nodes needs to be adjusted to the same number of nodes as the period waveform data. For example, the target reinforcement learning knowledge vector data may contain 100 nodes of various characteristic information about seismic waves, while the original period waveform data has only 50 nodes. The node compression unit will compress the target reinforcement learning knowledge vector data according to certain algorithms and rules. This algorithm may merge or discard relatively unimportant node information based on an evaluation of the importance of seismic wave features. For example, for some nodes corresponding to subtle fluctuation information that has little impact on the overall characteristics of seismic waves, a merge or removal operation can be performed. After node compression, the number of nodes of the generated target knowledge vector data is the same as the number of nodes of the period waveform data, for example, it becomes 50 nodes, and this target knowledge vector data still retains the key characteristic information of the seismic wave, but exists in a more compact form that matches the number of nodes of the original period waveform data.

[0079] Then, the server performs an inverse regularization conversion on the target knowledge vector data through an inverse regularization conversion unit to generate enhanced waveform data corresponding to the time period waveform data. In the previous steps, the time period waveform data was subjected to a regularization conversion by the regularization conversion unit, such as normalizing the amplitude and frequency, adjusting the data to a specific value range, and other operations. Now, the inverse regularization conversion unit is to perform the opposite operation. If the amplitude value range is normalized from the original -100 to 100 to 0 to 1 during the regularization conversion, then during the inverse regularization conversion, it will be restored to the original value range or a reasonable value range corresponding thereto according to the amplitude information in the target knowledge vector data and the previous regularization conversion algorithm. Similarly, corresponding inverse regularization conversion operations are also performed for other seismic wave characteristic information such as frequency and phase. After the inverse regularization conversion, the target knowledge vector data is converted into enhanced waveform data corresponding to the time period waveform data. This enhanced waveform data is based on the waveform data of the original time period. After deep processing by the seismic wave enhancement analysis network, it is more accurate and can better reflect the true characteristics of the seismic wave. It has more important value in the analysis and research of seismic waves and possible subsequent applications (such as earthquake early warning, underground structure detection, etc.) because it integrates the basic information of the original data and the advanced feature information after network learning analysis, and can provide more reliable data support for related technologies and research.

[0080] In one possible implementation, the method includes:

[0081] Step S101 , obtaining sample learning data, each of the sample learning data includes waveform data of a first template period and waveform data of a second template period, and the data value of the waveform data of the first template period is greater than the data value of the waveform data of the second template period.

[0082] Step S102, loading the waveform data of the second template period into the initialized seismic wave enhancement analysis network, the seismic wave enhancement analysis network includes multiple model function layers, each model function layer is cascaded, each of the model function layers includes a multi-stage knowledge transfer unit, and the multi-stage knowledge transfer unit includes more than two orders of knowledge transfer learning.

[0083] Step S103, in the multi-stage knowledge transfer unit of the target model function layer, multi-stage knowledge transfer learning is performed on the loading data of the target model function layer, and the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of each stage is node-fused to generate a multi-stage knowledge transfer learning vector corresponding to the target model function layer, and the multi-stage knowledge transfer learning vector is used to generate the reinforcement learning knowledge vector data corresponding to the target model function layer. Among them, the first stage knowledge transfer learning is used to learn the loading data of the target model function layer, and the knowledge transfer learning of each remaining stage is used to learn the remaining node knowledge vector data except the key node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the previous stage knowledge transfer learning. Among them, when the target model function layer is the first model function layer among the multiple model function layers, the loading data of the target model function layer is the basic knowledge vector data of the waveform data of the second template period. When the target model function layer is a model function layer after the first model function layer among the multiple model function layers, the loading data of the target model function layer is the reinforcement learning knowledge vector data corresponding to the previous model function layer of the target model function layer.

[0084] Step S104, based on the reinforcement learning knowledge vector data corresponding to the last model functional layer, generate the template enhanced waveform data corresponding to the second template period waveform data, generate the target training cost according to the error between the template enhanced waveform data and the first template period waveform data, and optimize the network parameters of the seismic wave enhancement analysis network based on the target training cost.

[0085] Wherein, step S101 includes:

[0086] Step S1011, obtaining a first template seismic wave sequence, wherein the first template seismic wave sequence includes a plurality of first template time period waveform data.

[0087] Step S1012: performing noise disturbance processing on the first template seismic wave sequence to generate a seismic wave sequence after noise disturbance processing.

[0088] Step S1013, performing feature compression on the seismic wave sequence after the noise disturbance processing to generate a second template seismic wave sequence.

[0089] Step S1014: Generate sample learning data based on first template period waveform data and second template period waveform data in the same period identifier obtained from the first template seismic wave sequence and the second template seismic wave sequence.

[0090] In this embodiment, in the training process of the seismic wave analysis and seismic wave enhancement analysis network, the first step is to obtain sample learning data. For example, a first template seismic wave sequence can be obtained first, and this first template seismic wave sequence contains multiple first template time period waveform data. For example, in a specific seismic monitoring area, relatively pure, accurate and representative seismic wave data collected by high-precision seismic monitoring equipment constitute the first template seismic wave sequence. These first template time period waveform data have high data value, and they can more accurately reflect the true characteristics of the seismic waves in the area at different time periods, including accurate amplitude, frequency, phase and other information.

[0091] Then, the server performs noise disturbance processing on the first template seismic wave sequence. This operation is to simulate various interference factors that may exist in actual seismic wave monitoring. In the actual environment, seismic waves will be disturbed by many factors during propagation, such as the unevenness of underground geological structures, weak vibrations caused by human activities, etc. The server adds noise to the first template seismic wave sequence through a specific algorithm to generate a seismic wave sequence after noise disturbance processing. For example, according to a predetermined noise model, random fluctuations within a certain range may be added to the amplitude of the original seismic wave, or a small offset may be introduced in the frequency, so that this seismic wave sequence is closer to the complex seismic wave conditions that may actually be received.

[0092] Next, the server performs feature compression on the seismic wave sequence after noise disturbance processing. Since the seismic wave sequence after noise disturbance processing may contain more redundant information or overly complex features, feature compression is required to facilitate subsequent learning and analysis. The feature compression process is to extract the most representative and key feature information in the seismic wave sequence through a specific algorithm, thereby generating a second template seismic wave sequence. For example, when processing the frequency characteristics of seismic waves, frequency information that fluctuates frequently within a small range but has little effect on the overall seismic wave characteristics may be merged or simplified, retaining those frequency characteristics that have an important impact on the overall shape and propagation characteristics of the seismic waves.

[0093] Finally, the server generates sample learning data based on the first template period waveform data and the second template period waveform data in the same period identifier obtained from the first template seismic wave sequence and the second template seismic wave sequence. This means that for each specific period, there is a corresponding set of first template period waveform data (with high data value) and second template period waveform data (relatively low value data after noise disturbance and feature compression), and such a pair of data constitutes a sample learning data. These sample learning data will be used to train the seismic wave enhancement analysis network.

[0094] After obtaining the sample learning data, the server starts to use the data to train the initialized seismic wave enhancement analysis network. The server loads the waveform data of the second template period into the initialized seismic wave enhancement analysis network, which contains multiple model function layers, each of which is cascaded, and each model function layer includes a multi-stage knowledge transfer unit, which includes more than two levels of knowledge transfer learning.

[0095] When the waveform data of the second template period is loaded into the network, in the multi-stage knowledge transfer unit of the target model functional layer, multi-stage knowledge transfer learning is performed on the loaded data of the target model functional layer. Take a seismic wave enhancement analysis network containing three model functional layers (functional layer A, functional layer B, functional layer C) as an example. If the target model functional layer is functional layer A (i.e., the first model functional layer), the loaded data of the target model functional layer is the basic knowledge vector data of the waveform data of the second template period. This basic knowledge vector data is obtained by the aforementioned operations such as regularization conversion and node expansion of the waveform data of the second template period. In the multi-stage knowledge transfer unit of functional layer A, it is assumed that this unit has three stages of knowledge transfer learning.

[0096] In the first stage of knowledge transfer learning, functional layer A directly learns the basic knowledge vector data of the waveform data of the second template period. This learning process analyzes various seismic wave feature information in this vector data, such as the initial amplitude distribution, frequency range, etc., and generates the first stage of knowledge transfer learning vector through internal calculation and analysis mechanisms. This vector contains the results of learning the basic knowledge vector data at this stage, such as the preliminary understanding and analysis results of the amplitude and frequency characteristics.

[0097] Then, in the subsequent stage of knowledge transfer learning, for the second stage, the knowledge transfer learning vector generated by the knowledge transfer learning in the first stage must first be node-splitting to obtain the remaining node knowledge vector data except for the key node knowledge vector data. Assuming that the knowledge transfer learning vector in the first stage contains comprehensive information about the amplitude, frequency, phase, etc. of the seismic wave, through node splitting, it can be split into node knowledge vector data about amplitude, node knowledge vector data about frequency, etc. according to different characteristics. Then, these remaining node knowledge vector data are learned. This learning process will further mine and analyze the information in these remaining node knowledge vector data according to the algorithm and parameter settings within the functional layer A, such as conducting a more in-depth study of the specific frequency intervals in the frequency node knowledge vector data, and generating the knowledge transfer learning vector for the second stage.

[0098] For the third stage, the knowledge transfer learning vector generated by the knowledge transfer learning in the second stage is first split into nodes to obtain the remaining node knowledge vector data, and then these data are learned to generate the key node knowledge vector data generated by the knowledge transfer learning in the third stage. This key node knowledge vector data may contain important information learned in this stage about a key feature of the seismic wave (such as the law of phase change under specific conditions).

[0099] Finally, the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning in each stage (three stages) are node fused. Node fusion is to combine the most important parts of knowledge learned in each stage according to certain rules. For example, the key information about amplitude in the first stage, the key information about frequency in the second stage, and the key information about phase in the third stage are organically combined to form a multi-stage knowledge transfer learning vector corresponding to functional layer A. This vector will be used to generate reinforcement learning knowledge vector data corresponding to functional layer A.

[0100] If the target model functional layer is functional layer B (i.e., the model functional layer after the first model functional layer), the loaded data of the target model functional layer is the reinforcement learning knowledge vector data corresponding to functional layer A. In the multi-stage knowledge transfer unit of functional layer B, multi-stage knowledge transfer learning is also performed. In the first-order knowledge transfer unit of the multi-stage knowledge transfer unit, first-order knowledge transfer learning is performed on the reinforcement learning knowledge vector data corresponding to functional layer A to generate a first-order knowledge transfer learning vector. This process is similar to the knowledge transfer learning in functional layer A, except that the input data becomes the reinforcement learning knowledge vector data of functional layer A. Functional layer B will analyze and process this data according to its own learning mechanism, extract new feature information, and generate a first-order knowledge transfer learning vector.

[0101] In the transition knowledge transfer unit of the multi-stage knowledge transfer unit, the remaining node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level (the first-order knowledge transfer unit) is obtained, and knowledge transfer learning is performed on these remaining node knowledge vector data to generate a transition knowledge transfer learning vector. For example, the first-order knowledge transfer learning vector contains more complex feature information about the seismic wave after being processed by the functional layer A. The node knowledge vector data about different features is obtained by node splitting, and then some of the node knowledge vector data (the remaining node knowledge vector data) are deeply learned, such as a more detailed analysis of the node knowledge vector data related to the energy attenuation characteristics of the seismic wave after being processed by the functional layer A, to generate a transition knowledge transfer learning vector.

[0102] In the last knowledge transfer unit of the multi-stage knowledge transfer unit, the remaining node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level (transition knowledge transfer unit) is obtained, and the last stage of knowledge transfer learning is performed on these remaining node knowledge vector data to generate the key node knowledge vector data generated by the last stage of knowledge transfer learning. For example, the last stage of learning is performed on the remaining node knowledge vector data related to the change of the seismic wave phase in a specific time period in the transition knowledge transfer learning vector to obtain the key information about this phase change at this stage, forming the key node knowledge vector data generated by the last stage of knowledge transfer learning.

[0103] Then, the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning in each stage (first stage, transition, and last stage) are fused to generate a multi-stage knowledge transfer learning vector corresponding to functional layer B. This vector is used to generate reinforcement learning knowledge vector data corresponding to functional layer B.

[0104] The same process will also be carried out in functional layer C (assuming it is the last model functional layer), and finally the reinforcement learning knowledge vector data corresponding to functional layer C will be generated.

[0105] Based on the reinforcement learning knowledge vector data corresponding to the last model functional layer (functional layer C), the server generates template reinforcement waveform data corresponding to the waveform data of the second template period. This process is similar to the process of generating reinforcement waveform data when processing normal period waveform data, including combining the basic knowledge vector data of the waveform data of the second template period and the reinforcement learning knowledge vector data corresponding to the functional layer C to generate the target reinforcement learning knowledge vector data, and then compressing the target reinforcement learning knowledge vector data through the node compression unit to generate the target knowledge vector data, and finally performing inverse regularization conversion on the target knowledge vector data through the inverse regularization conversion unit to obtain the template reinforcement waveform data corresponding to the waveform data of the second template period.

[0106] Next, the server generates the target training cost based on the error between the template enhanced waveform data and the waveform data of the first template period. Since the waveform data of the first template period has a higher data value and is a more accurate seismic wave data, and the template enhanced waveform data is the result of the seismic wave enhancement analysis network processing the relatively low-value waveform data of the second template period, the difference between the two can reflect the current processing effect of the seismic wave enhancement analysis network. Through a specific error calculation method, such as calculating the mean square error of the two in terms of amplitude, frequency, phase, etc., a quantized error value is obtained, and this error value is the target training cost.

[0107] Finally, the server optimizes the network parameters of the seismic wave enhancement analysis network based on this target training cost. According to the size and direction of the target training cost (whether it is too large or too small, and which features have large errors, etc.), the parameters of each model function layer in the seismic wave enhancement analysis network are adjusted. For example, if it is found that the error in amplitude between the template enhancement waveform data and the waveform data of the first template period is large when calculating the target training cost, then the parameters in the model function layer related to amplitude analysis (such as the parameter part involving amplitude analysis in function layer A, function layer B, and function layer C) will be adjusted, which may be to adjust the weights in knowledge transfer learning, the coefficients in node nonlinear conversion, etc., so that the network can reduce this error when processing similar data next time, thereby improving the analysis and processing capabilities of seismic wave data. After multiple such training processes, the seismic wave enhancement analysis network can continuously optimize its own parameters to process seismic wave data more accurately.

[0108] Figure 2 The hardware structure of the multi-source monitoring data analysis system 100 based on active seismic sources for implementing the multi-source monitoring data analysis method based on active seismic sources provided by the embodiment of the present invention is shown as follows: Figure 2As shown, the seismic active source-based multi-source monitoring data analysis system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0109] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the multi-source monitoring data analysis system 100 based on active seismic sources uses to execute or use to complete the exemplary method described in the present invention.

[0110] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the multi-source monitoring data analysis method based on seismic active source in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0111] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the multi-source monitoring data analysis system 100 based on the above-mentioned active seismic source. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0112] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above multi-source monitoring data analysis method based on active seismic sources is implemented.

[0113] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A multi-source monitoring data analysis method based on active seismic sources, characterized in that: The method comprises: Acquire waveform data of multiple time periods of candidate seismic wave sequences of multi-source monitoring data of seismic active sources, and for each time period of waveform data, load the waveform data of the time period into a pre-trained seismic wave enhancement analysis network, wherein the seismic wave enhancement analysis network includes multiple model function layers, each model function layer is cascaded, and each of the model function layers includes a multi-stage knowledge transfer unit, and the multi-stage knowledge transfer unit includes more than two stages of knowledge transfer learning; In the multi-stage knowledge transfer unit of the target model function layer, multi-stage knowledge transfer learning is performed on the loading data of the target model function layer, and the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of each stage is node-fused to generate a multi-stage knowledge transfer learning vector corresponding to the target model function layer, and the multi-stage knowledge transfer learning vector is used to generate the reinforcement learning knowledge vector data corresponding to the target model function layer; wherein the first stage knowledge transfer learning is used to learn the loading data of the target model function layer, and the knowledge transfer learning of each remaining stage is used to learn the remaining node knowledge vector data except the key node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous stage; wherein, when the target model function layer is the first model function layer among the multiple model function layers, the loading data of the target model function layer is the basic knowledge vector data of the waveform data of the time period; when the target model function layer is the model function layer after the first model function layer among the multiple model function layers, the loading data of the target model function layer is the reinforcement learning knowledge vector data corresponding to the previous model function layer of the target model function layer; Based on the reinforcement learning knowledge vector data corresponding to the last model functional layer, the reinforcement waveform data corresponding to the waveform data of the time period is generated; based on the reinforcement waveform data corresponding to each of the waveform data of the time period, the reinforcement learning seismic wave sequence corresponding to the candidate seismic wave sequence is generated.

2. The multi-source monitoring data analysis method based on active seismic sources according to claim 1 is characterized in that: The seismic wave enhancement analysis network further includes a regularization conversion unit and a node expansion unit. The step of determining the basic knowledge vector data of the time period waveform data includes: The regularized conversion unit performs regularized conversion on the time period waveform data to generate regularized waveform data; The node expansion unit performs node expansion on the regularized waveform data to generate basic knowledge vector data of the time period waveform data.

3. The multi-source monitoring data analysis method based on active seismic sources according to claim 1 is characterized in that: When the target model function layer is a model function layer after the first model function layer among the multiple model function layers, performing multi-stage knowledge transfer learning on the loaded data of the target model function layer, performing node fusion on key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of each stage, and generating a multi-stage knowledge transfer learning vector corresponding to the target model function layer, includes: In the first-order knowledge transfer unit of the multi-stage knowledge transfer unit, first-order knowledge transfer learning is performed on the reinforcement learning knowledge vector data corresponding to the previous model function layer of the target model function layer to generate a first-order knowledge transfer learning vector; In the transition knowledge transfer unit of the multi-stage knowledge transfer unit, the remaining node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level is obtained, and the knowledge transfer learning is performed on the remaining node knowledge vector data to generate a transition knowledge transfer learning vector; In the last knowledge transfer unit of the multi-stage knowledge transfer unit, the remaining node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level is obtained, and the last stage of knowledge transfer learning is performed on the remaining node knowledge vector data to generate the key node knowledge vector data generated by the last stage of knowledge transfer learning; The key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning in each stage are node-fused to generate a multi-stage knowledge transfer learning vector corresponding to the functional layer of the target model.

4. The multi-source monitoring data analysis method based on active seismic sources according to claim 3 is characterized in that: The method further comprises: Perform node splitting on the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level to generate node knowledge vector data of each feature node; According to the arrangement order of each feature node, starting from the first feature node, the node knowledge vector data of the target number of nodes with a later node order are used as the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level, and the node knowledge vector data of the remaining nodes are used as the remaining node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of the previous level.

5. The multi-source monitoring data analysis method based on active seismic sources according to claim 1 is characterized in that: Each model function layer also includes a node nonlinear conversion unit based on a feature self-focusing strategy, and the method also includes: Loading the multi-stage knowledge transfer learning vector generated by the multi-stage knowledge transfer unit in the target model function layer into the node nonlinear conversion unit in the target model function layer, performing node nonlinear conversion processing on the multi-stage knowledge transfer learning vector according to the feature self-focusing strategy, and generating a multi-stage knowledge transfer learning vector for node nonlinear conversion; The step of generating the reinforcement learning knowledge vector data corresponding to the target model function layer includes: In the target model functional layer, reinforcement learning knowledge vector data corresponding to the target model functional layer is generated based on the multi-stage knowledge transfer learning vector corresponding to the target model functional layer, the multi-stage knowledge transfer learning vector of the node nonlinear transformation, and the reinforcement learning knowledge vector data corresponding to the previous model functional layer of the target model functional layer.

6. The multi-source monitoring data analysis method based on active seismic sources according to claim 5 is characterized in that: The multi-stage knowledge transfer learning vector corresponding to the target model function layer includes node knowledge transfer learning vectors of each feature node, and the node nonlinear conversion processing is performed on the multi-stage knowledge transfer learning vector according to the feature self-focusing strategy to generate a multi-stage knowledge transfer learning vector of node nonlinear conversion, including: Performing a focusing process of a feature self-focusing strategy on the multi-stage knowledge transfer learning vector, and determining the attention weights corresponding to each node in the multi-stage knowledge transfer learning vector; The attention weights and node knowledge transfer learning vectors corresponding to each feature node are respectively multiplied to generate a multi-stage knowledge transfer learning vector for node nonlinear transformation.

7. The multi-source monitoring data analysis method based on active seismic sources according to claim 5 is characterized in that: The step of generating the reinforcement learning knowledge vector data corresponding to the target model functional layer based on the multi-stage knowledge transfer learning vector corresponding to the target model functional layer, the multi-stage knowledge transfer learning vector of the node nonlinear conversion, and the reinforcement learning knowledge vector data corresponding to the previous model functional layer of the target model functional layer comprises: Combining the multi-stage knowledge transfer learning vector corresponding to the functional layer of the target model and the multi-stage knowledge transfer learning vector of the node nonlinear conversion to generate combined knowledge vector data; Perform node conversion on the combined knowledge vector data to generate converted knowledge vector data, and combine the converted knowledge vector data with the reinforcement learning knowledge vector data corresponding to the previous model function layer of the target model function layer to generate the reinforcement learning knowledge vector data corresponding to the target model function layer.

8. The multi-source monitoring data analysis method based on active seismic sources according to claim 1 is characterized in that: The step of generating the enhanced waveform data corresponding to the time period waveform data based on the enhanced learning knowledge vector data corresponding to the last model function layer includes: Combining the basic knowledge vector data of the period waveform data and the reinforcement learning knowledge vector data corresponding to the last model function layer to generate target reinforcement learning knowledge vector data, and determining the reinforcement waveform data corresponding to the period waveform data based on the target reinforcement learning knowledge vector data; The seismic wave enhancement analysis network further includes an inverse regularization conversion unit and a node compression unit, and the enhanced waveform data corresponding to the waveform data of the time period is determined according to the target reinforcement learning knowledge vector data, including: Performing node compression on the target reinforcement learning knowledge vector data by the node compression unit to generate target knowledge vector data, wherein the number of nodes of the target knowledge vector data is the same as the number of nodes of the period waveform data; The target knowledge vector data is subjected to an inverse regularization conversion by the inverse regularization conversion unit to generate enhanced waveform data corresponding to the time period waveform data.

9. The multi-source monitoring data analysis method based on active seismic sources according to claim 1 is characterized in that: The method comprises: Acquire sample learning data, each of the sample learning data includes waveform data of a first template period and waveform data of a second template period, and the data value of the waveform data of the first template period is greater than the data value of the waveform data of the second template period; Loading the second template period waveform data into an initialized seismic wave enhancement analysis network, wherein the seismic wave enhancement analysis network includes a plurality of model function layers, each of which is cascaded, and each of which includes a multi-stage knowledge transfer unit, and the multi-stage knowledge transfer unit includes knowledge transfer learning of more than two orders; In the multi-stage knowledge transfer unit of the target model function layer, multi-stage knowledge transfer learning is performed on the loading data of the target model function layer, and the key node knowledge vector data in the knowledge transfer learning vector generated by the knowledge transfer learning of each stage is node-fused to generate a multi-stage knowledge transfer learning vector corresponding to the target model function layer, and the multi-stage knowledge transfer learning vector is used to generate the reinforcement learning knowledge vector data corresponding to the target model function layer; wherein the first stage knowledge transfer learning is used to learn the loading data of the target model function layer, and the knowledge transfer learning of each remaining stage is used to learn the remaining node knowledge vector data except the key node knowledge vector data generated by node splitting the knowledge transfer learning vector generated by the knowledge transfer learning of the previous stage; wherein, when the target model function layer is the first model function layer among the multiple model function layers, the loading data of the target model function layer is the basic knowledge vector data of the waveform data of the second template period; when the target model function layer is the model function layer after the first model function layer among the multiple model function layers, the loading data of the target model function layer is the reinforcement learning knowledge vector data corresponding to the previous model function layer of the target model function layer; Based on the reinforcement learning knowledge vector data corresponding to the last model function layer, generate template reinforcement waveform data corresponding to the second template period waveform data, generate a target training cost according to the error between the template reinforcement waveform data and the first template period waveform data, and optimize the network parameters of the seismic wave reinforcement analysis network based on the target training cost; The step of obtaining sample learning data includes: Acquire a first template seismic wave sequence, wherein the first template seismic wave sequence includes a plurality of first template time period waveform data; Performing noise disturbance processing on the first template seismic wave sequence to generate a seismic wave sequence after noise disturbance processing; Performing feature compression on the seismic wave sequence after the noise disturbance processing to generate a second template seismic wave sequence; Based on obtaining the first template period waveform data and the second template period waveform data in the same period identifier from the first template seismic wave sequence and the second template seismic wave sequence, sample learning data is generated.

10. A multi-source monitoring data analysis system based on active seismic sources, characterized in that: The multi-source monitoring data analysis system based on active seismic sources includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the multi-source monitoring data analysis method based on active seismic sources as described in any one of claims 1 to 9.

Images