A seismic electromagnetic multi-field data fusion method and system

Abstract

The application discloses a kind of shock electromagnetic multi-field data fusion method and system, it is related to data processing technical field, solve the technical problem that cannot utilize multi-field data to monitor the state of support structure body, its technical scheme main point is that feature encoder is extracted respectively from sensor data microseismic field data, electric field data and electromagnetic radiation field data features.Through field data relationship learning module and inter-field data relationship learning module, global timing features in single-field data and local interaction features between multi-field data are learned respectively, and consistency loss is used to constrain local interaction features between multi-field data, and finally, the fusion of intra-field features and multi-field interaction features is realized through multi-feature fusion module, to early rupture of support structure body is early warned.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a seismic and electromagnetic multi-field data fusion method and system. BACKGROUND

[0002] Four-dimensional support of deep underground tunnel requires to adjust and control the deformation process of surrounding rock to ensure the safety and stability of underground tunnel. Monitoring and early warning technology can judge the stability of surrounding rock by monitoring the dynamic changes of four-dimensional support structure, and provide decision basis for decision makers to realize disaster early warning.

[0003] The current monitoring means starts from the aspects of seismic field, geoelectric field and electromagnetic field, uses sensor technology to collect microseismic, electric field and electromagnetic radiation signals of four-dimensional support structure of deep tunnel in real time, reflects the health status of the structure through signal changes, and forms a real-time monitoring system of surrounding rock. The above monitoring system has achieved certain effect, but is based on single field source, ignores the multi-field relationship in the deformation and fracture process of surrounding rock structure, cannot fully utilize multi-field information, and cannot mine the correlation between multi-field relationship and structure fracture, so as to accurately realize early warning. Therefore, a method capable of fusing multi-field data is needed to realize accurate early warning of support structure. SUMMARY

[0004] The present application provides a seismic and electromagnetic multi-field data fusion method and system, which aims to capture the relationship of multi-field data and realize the fusion of seismic and electromagnetic multi-field data to realize accurate early warning of the state of support structure.

[0005] The above technical purpose of the present application is realized by the following technical scheme:

[0006] A seismic and electromagnetic multi-field data fusion method, comprising:

[0007] S1: extracting data features of microseismic field, electric field and electromagnetic radiation field from sensor data through a feature encoder, respectively;

[0008] S2: learning the data relationship in the field of the data features of microseismic field, electric field and electromagnetic radiation field, respectively, to obtain global time sequence features in single-field data;

[0009] S3: learning the data relationship between fields by combining the data features of microseismic field, electric field and electromagnetic radiation field in pairs, respectively, to obtain local interaction features between multi-field data;

[0010] S4: constraining the local interaction features between multi-field data by consistency loss to obtain the local interaction features between multi-field data after constraint;

[0011] S5: The global temporal features within a single field data are fused with the local interaction features between constrained multi-field data to obtain fused features, thereby realizing the fusion of seismomagnetic multi-field data.

[0012] Furthermore, in step S2, learning the in-field data relationships using a graph attention network includes:

[0013] S21: Construct fully connected graph attention networks for the microseismic field, electric field, and electromagnetic radiation field, respectively. These fully connected graph attention networks will integrate the data features X of the microseismic field, electric field, and electromagnetic radiation field. μ The system is divided into several nodes, and each node is adjacent to other nodes in the field, resulting in fully connected graphs of the microseismic field, electric field, and electromagnetic radiation field. The attention network of the fully connected graph adopts the self-attention mechanism of the transformer network. Nodes in the field that are adjacent to each other are connected by edges.

[0014] S22: Perform neighbor aggregation on each node and all its adjacent nodes in the fully connected graph within the field to obtain the aggregated feature. The aggregated feature is then represented as:

[0015]

[0016] in, This represents the weight of an edge; the edge weight is the self-attention factor in the transformer network.

[0017] Self-attention coefficient value calculated by force mechanism; X μ Indicates features before aggregation; The aggregated features are represented by i = 1, ..., h, where h represents the number of self-attention heads, i.e., the number of aggregations; z, d, and c represent the microseismic field, electric field, and electromagnetic radiation field, respectively; d k express or In the dimension of , softmax represents the activation function. Both represent weight matrices, and || represents a cascade operation;

[0018] S23: Update the state of each node in the fully connected graph within the field to obtain the updated features. The state update is represented as follows:

[0019]

[0020] in, Represents the updated feature; ⊙ represents the Hadamard product; This is the weight matrix;

[0021] S24: repeating steps S21-S23 to perform multi-layer graph learning on the updated features, and finally obtaining the global temporal features within single field data by combining adaptive residuals, which is represented as:

[0022]

[0023] wherein, represents the global temporal features within single field data; represents the node feature matrix after L times of neighbor aggregation and state update; represents the weight matrix.

[0024] Further, in step S3, the inter-field data relationship learning is performed by a graph attention network, including:

[0025] S31: constructing an inter-field sparse graph attention network between the microseismic field, the electric field and the electromagnetic radiation field, which divides the data features X μ of the microseismic field, the electric field and the electromagnetic radiation field into a plurality of nodes, calculates the inter-field attention between the nodes of different fields, respectively retains C attention coefficient values of other two fields from large to small for each node, and sets the remaining attention coefficient values to zero to obtain a sparse graph in which C nodes in different fields are adjacent to any one node, and in this way, three inter-field sparse graphs, i.e., a microseismic field-electric field inter-field sparse graph, a microseismic field-electromagnetic radiation field inter-field sparse graph and an electric field-electromagnetic radiation field inter-field sparse graph, are obtained; wherein the inter-field nodes having adjacent relationship are connected by edges, and the weight of the edge is the self-attention coefficient value;

[0026] S32: performing multi-head neighbor aggregation on each node and all adjacent nodes in the inter-field sparse graph to obtain aggregated features, and the aggregated features are represented as:

[0027]

[0028]

[0029] wherein μ j , μ j' ∈{z,d,c}, and μ j ≠ μ j' , z, d, and c respectively represent the microseismic field, the electric field and the electromagnetic radiation field; represents the feature of the node in the μ j' field aggregated with the nodes having adjacent relationship in the μ j' field; represents the feature of the node in the μ j field aggregated with the nodes having adjacent relationship in the μ j field. and both represent the pre-aggregation features; and both represent the weight matrix;

[0030] S33: performing state update on each node in the inter-field sparse graph to obtain updated features, and the state update is represented as:

[0031]

[0032] wherein, and both represent the updated features, both represent the weight matrix;

[0033] S34: repeating steps S21-S23 on the updated features to perform multi-layer graph learning, and after completing the multi-layer graph learning, introducing adaptive residuals to obtain final inter-field local interaction features and are represented as:

[0034]

[0035] wherein, and represent the features obtained through L-layer graph learning, and both represent the weight matrix.

[0036] Further, in step S4, the constrained inter-field local interaction features are represented as:

[0037]

[0038] wherein, μ∈{z,d,c}, μ j ,μ j' ,μ j” ∈{z,d,c}, and μ j ≠μ j' ≠μ j” ,μ j =μ.

[0039] Further, in step S5, the fusion features X are represented as:

[0040]

[0041] A seismic electromagnetic multi-field data fusion system, comprising:

[0042] a feature encoder configured to extract data features of a microseismic field, an electric field and an electromagnetic radiation field from sensor data, respectively;

[0043] The field-in data relationship learning module learns the data features of the microseismic field, the electric field and the electromagnetic radiation field, and obtains global time sequence features in single-field data respectively;

[0044] The field-in data relationship learning module learns the data features of the microseismic field, the electric field and the electromagnetic radiation field, and obtains global time sequence features in single-field data respectively;

[0045] The consistency loss module constrains the local interaction features between the multi-field data, and obtains the constrained local interaction features between the multi-field data;

[0046] The multi-feature fusion module fuses the global time sequence features in single-field data and the constrained local interaction features between the multi-field data, obtains fusion features, and realizes the fusion of the seismo-electromagnetic multi-field data.

[0047] A device includes a memory and a processor, the memory is used to store a computer program, and the processor is used to realize the seismo-electromagnetic multi-field data fusion method.

[0048] A computer readable storage medium, characterized in that the computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize the seismo-electromagnetic multi-field data fusion method.

[0049] The seismo-electromagnetic multi-field data fusion method and system have the following beneficial effects: the feature encoder is used to extract the data features of the microseismic field, the electric field and the electromagnetic radiation field from the original sensor data respectively, the feature encoder learns the time relationship in each field data through the bidirectional LSTM, and each field feature is extracted. The field-in data relationship learning module constructs a single-field graph according to each field data, learns the hidden space-time relationship between the data by using a graph attention network, and obtains global time sequence features in single-field data. The field-in data relationship learning module combines the multi-field data in pairs, constructs an inter-field sparse graph, uses a graph attention network to mine the relationship between the multi-field data, and obtains local interaction features between the multi-field data. The consistency loss module can constrain the local interaction features between the multi-field data, so as to guide the learning of the local interaction features between the multi-field data to be more accurate. The multi-feature fusion module fuses the global time sequence features in single-field data and the constrained local interaction features between the multi-field data, and the fusion features are used as the final features to participate in the structure damage identification.

[0050] The fusion of the seismo-electromagnetic multi-field data can monitor the state of the structure and early warn the structure damage as much as possible, so as to feedback the reinforcement of the structure and avoid the consequences caused by the structure damage. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1A flow framework diagram of the seismic-electromagnetic multi-field data fusion method in the embodiments of the present application is provided.

[0052] Figure 2 A field data relationship learning structure diagram is provided for the embodiments of the present application.

[0053] Figure 3 A field data relationship learning structure diagram is provided for the embodiments of the present application.

[0054] Figure 4 A confusion matrix result diagram of model classification is provided for the embodiments of the present application.

[0055] Figure 5 A sample prediction class distribution diagram along the time axis of model classification is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0056] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings.

[0057] As shown in the drawings, Figure 1 a seismic-electromagnetic multi-field data fusion method described in the present application comprises:

[0058] S1: Extracting data features of microseismic field, electric field and electromagnetic radiation field from sensor data through a feature encoder.

[0059] Specifically, three bidirectional LSTM networks are used to extract data features of microseismic field, electric field and electromagnetic radiation field from original sensor data; wherein the data collected by sensors at different positions are taken as features at each time, and average pooling is used to balance the outputs at different times to complete feature extraction.

[0060] S2: Learning the intra-field data relationship of the data features of the microseismic field, the electric field and the electromagnetic radiation field, and obtaining the global time sequence features in the single field data respectively.

[0061] In the embodiments of the present application, the intra-field data relationship is learned through a graph attention network, as shown in the drawings, Figure 2 specifically comprising:

[0062] S21: Building a full-connection graph attention network about the microseismic field, the electric field and the electromagnetic radiation field respectively, and the full-connection graph attention network is a multi-layer graph learning network. The full-connection graph attention network divides the data features X μ of the microseismic field, the electric field and the electromagnetic radiation field into a plurality of nodes, and any one node has an adjacent relationship with other nodes in the field, and the intra-field global full-connection graph of the microseismic field, the electric field and the electromagnetic radiation field is obtained respectively; wherein the full-connection graph attention network adopts the self-attention mechanism of the transformer network; the intra-field nodes with adjacent relationship are connected through edges.

[0063] Intra-field refers to the same field, for example, the data characteristics of a microseismic field are divided into several nodes by a full-connected graph attention network, wherein any node has an adjacent relationship with other nodes, if the data characteristics of a microseismic field are divided into 3 nodes (a, b, c), there are three adjacent relationships (a and b, c, b and a, c, c and a, b) to form the intra-field full-connected graph of the microseismic field. The intra-field full-connected graph of the electric field and the electromagnetic radiation field is also obtained in the same way.

[0064] S22: multi-head neighbor aggregation is performed on each node and all adjacent nodes in the intra-field full-connected graph to obtain the aggregated feature, and the aggregated feature is represented as:

[0065]

[0066] wherein, represents the weight of the edge, that is, the self-attention coefficient value calculated by the self-attention mechanism in the transformer network; X μ represents the pre-aggregation feature; represents the post-aggregation feature; i = 1,..., h, h represents the number of self-attention heads; μ ∈ {z, d, c}, z, d, and c represent the microseismic field, the electric field, and the electromagnetic radiation field, respectively; d k represents or the dimension of or, and softmax represents an activation function, both represent weight matrices, and || represents a concatenation operation.

[0067] S23: state updating is performed on each node in the intra-field full-connected graph to obtain an updated feature, and the state updating is represented as:

[0068]

[0069] wherein, represents the updated feature; and ⊙ represents Hadamard product. is a weight matrix.

[0070] S24: since the full-connected graph attention network is a multi-layer graph learning network, the updated feature needs to be repeatedly subjected to steps S21 to S23 for multi-layer graph learning. With the deepening of the graph neural network, over-smoothing is prone to occur, and in view of this problem, the application designs an adaptive residual to obtain the final single-field intra-data global time sequence feature, which is represented as:

[0071]

[0072] wherein, represents the single-field intra-data global time sequence feature. denotes a node feature matrix after L times of neighbor aggregation and state update; denotes a weight matrix.

[0073] S3: learn the inter-field data relationship by two-by-two combination of data features of the microseismic field, the electric field and the electromagnetic radiation field, to obtain local interaction features between multi-field data.

[0074] In the embodiments of the present application, the inter-field data relationship is learned through a graph attention network, as shown in Figure 3 , and specifically includes:

[0075] S31: construct an inter-field sparse graph attention network between the microseismic field, the electric field and the electromagnetic radiation field, which is a multi-layer graph learning network. The inter-field sparse graph attention network divides the data features X μ of the microseismic field, the electric field and the electromagnetic radiation field into a plurality of nodes, calculates the inter-field attention between the nodes of different fields, retains C attention coefficient values of each node from large to small with the other two fields respectively, and sets the remaining attention coefficient values to zero, to obtain a sparse graph in which any one node and C nodes in different fields have an adjacent relationship. In this way, three inter-field sparse graphs are obtained, i.e. a microseismic field-electric field inter-field sparse graph, a microseismic field-electromagnetic radiation field inter-field sparse graph and an electric field-electromagnetic radiation field inter-field sparse graph; wherein the inter-field nodes having an adjacent relationship are connected by edges, and the weight of the edge is the attention coefficient value.

[0076] The inter-field refers to the different fields, such as the microseismic field and the electric field, the microseismic field and the electromagnetic radiation field. For example, the data in the microseismic field, the electric field and the electromagnetic radiation field are all divided into three nodes, which are (a1, b1, c1), (a2, b2, c2) and (a3, b3, c3). The inter-field attention of the nodes a1, b1, c1 in the microseismic field and all nodes (a2, b2, c2) in the electric field is calculated, and if each node retains two self-attention coefficient values from large to small, there are six attention coefficient values, such as the attention coefficient values between the edges a1a2, a1c2, b1a2, b1b2, c1b2 and c1c2. At the same time, the inter-field attention of the nodes a2, b2, c2 in the electric field and all nodes (a1, b1, c1) in the microseismic field is calculated, and each node retains two attention coefficient values from large to small, and there are six attention coefficient values, such as the attention coefficient values between the edges a2a1, a2b1, b2b1, b2c1, c2a1 and c2c1. Finally, the inter-field sparse graph of the microseismic field-electric field is composed of the six bidirectional edges (a1a2, a1c2) (b1a2, b1b2) (c1b2, c1c2) (a2a1, a2b1) (b2b1, b2c1) (c2a1, c2c1). The inter-field sparse graph of the microseismic field-electromagnetic radiation field and the inter-field sparse graph of the electric field-electromagnetic radiation field are also obtained according to the above method.

[0077] S32: Multi-head neighbor aggregation is performed on each node and all adjacent nodes in the inter-field sparse graph to obtain the aggregated features, and the aggregated features are represented as:

[0078]

[0079] Wherein, μ j , μ j' ∈{z, d, c}, and μ j ≠ μ j' , z, d, c represent the microseismic field, the electric field and the electromagnetic radiation field, respectively; represents the feature of the node pair in the μ j field after aggregation with the nodes adjacent thereto in the μ j' field; represents the feature of the node pair in the μ j' field after aggregation with the nodes adjacent thereto in the μ j field; and both represent the pre-aggregation features; and both represent the weight matrix;

[0080] S33: State updating is performed on each node in the inter-field sparse graph to obtain the updated features, and the state updating is represented as:

[0081]

[0082] wherein, and represent updated features, all represent weight matrices;

[0083] S34: repeating steps S21-S23 for the updated features to perform multi-layer graph learning, and introducing adaptive residuals after completing the multi-layer graph learning to obtain final local interaction features between multi-field data and are represented as:

[0084]

[0085] wherein, and represent features obtained through L-layer graph learning, and all represent weight matrices.

[0086] S4: constraining the local interaction features between multi-field data through a consistency loss to obtain constrained local interaction features between multi-field data.

[0087] Specifically, the output of the inter-field data relationship learning is constrained, and the output features of the same field are added to obtain the constrained local interaction features between multi-field data X z' , X d' and X c' for the microseismic field, the electric field and the electromagnetic radiation field, respectively, which are represented as:

[0088]

[0089] The constrained local interaction features between multi-field data for each field can be used to identify and classify the damage of the structure, thereby constraining the process of inter-field data relationship learning. The constrained local interaction features between multi-field data for each field are classified by three fully connected layers F z , F d and F c , and the classification process is as follows:

[0090] H z = F z (X z′ )

[0091] H d = F d (X d′ );

[0092] H c = F c(X c′ )

[0093] wherein, H z ,H d ,H c respectively represent the classification output of three fields, participating in the decision of the structural damage condition as part of the final classification.

[0094] S5: fuse the global time sequence features in single field data with the local interaction features between the constrained multi-field data to obtain fusion features, and realize the fusion of the seismic and electromagnetic multi-field data.

[0095] Specifically, the fusion feature X is expressed as:

[0096]

[0097] wherein, X is used as the final fusion feature for monitoring the health state of the structure. A fully connected layer is needed for classifying the fusion feature, and the classification process is as follows:

[0098] H = F(X);

[0099] wherein, H represents the classification output of the fusion feature. The structural damage condition is jointly decided by H z ,H d ,H c .

[0100] According to the monitoring data of the deep underground engineering structure, the seismic and electromagnetic multi-field data set is formed, the damage state of the structure is divided into four levels, which are initial compaction, elastic deformation, slight rupture and serious rupture, in order to achieve the effect of early warning, the first 10% and the last 10% of the data set corresponding to each stage are taken as the test set, and the pytorch neural network framework is used to build the seismic and electromagnetic multi-field data fusion system for four-dimensional support health monitoring of deep roadway, the recognition accuracy and F1 score of which both reach more than 92%, in order to further explore the recognition effect of each damage state of the structure, the confusion matrix diagram is drawn, as shown in Figure 4 From Figure 4 , it can be seen that the accuracy of recognizing the first rupture reaches 96.67%, which plays a good early warning role.

[0101] Figure 5 is the rupture prediction result drawn along the time axis, from Figure 5 , it can be seen that when the slight rupture just occurs, it can be basically accurately predicted, which provides positive feedback and decision support for support reinforcement and early warning.

[0102] The seismic and electromagnetic multi-field data fusion system described in the application is built by using the pytorch neural network framework, which specifically comprises:

[0103] a feature encoder configured to extract data features of a microseismic field, an electric field and an electromagnetic radiation field respectively from sensor data;

[0104] a field-intra data relationship learning module configured to learn field-intra data relationship of the data features of the microseismic field, the electric field and the electromagnetic radiation field, and obtain single-field data-intra global time sequence features respectively;

[0105] a field-inter data relationship learning module configured to learn field-inter data relationship of the data features of the microseismic field, the electric field and the electromagnetic radiation field in pairs, and obtain multi-field data-inter local interaction features respectively;

[0106] a consistency loss module configured to constrain the multi-field data-inter local interaction features by a consistency loss, and obtain constrained multi-field data-inter local interaction features;

[0107] a multi-feature fusion module configured to fuse the single-field data-intra global time sequence features and the constrained multi-field data-inter local interaction features, and obtain fusion features, so as to realize seismic-electromagnetic multi-field data fusion.

[0108] The above is a demonstrative embodiment of the present application, and the protection scope of the present application is defined by the claims and their equivalents.

Claims

1. A seismic electromagnetic multi-field data fusion method, characterized in that, The method comprises the following steps: S1: extracting data features of microseismic field, electric field and electromagnetic radiation field from sensor data respectively through a feature encoder; S2: learning the data relationship in the field to obtain global time sequence features in single-field data respectively; S3: learning the data relationship between fields in pairs to obtain local interaction features between multi-field data respectively; S4: constraining the local interaction features between multi-field data through consistency loss to obtain constrained local interaction features between multi-field data; S5: fusing the global time sequence features in single-field data and the constrained local interaction features between multi-field data to obtain fusion features, and realizing the fusion of seismo-electromagnetic multi-field data.

2. The method of claim 1, wherein, In the step S2, the data relationship in the field is learned through a graph attention network, comprising: S21: respectively construct a full connectivity graph attention network about the microseismic field, the electric field and the electromagnetic radiation field, which will data features X of the microseismic field, the electric field and the electromagnetic radiation field μ are divided into a plurality of nodes, and any one node has an adjacent relationship with other nodes in the field, and the full connectivity graph of the microseismic field, the electric field and the electromagnetic radiation field is obtained respectively; wherein the full connectivity graph attention network adopts the self-attention mechanism of the transformer network; the nodes in the field with adjacent relationship are connected through edges; S22: aggregating neighbors of each node in the fully connected graph in the field to obtain aggregated features, and the aggregated features are represented as: wherein, represents the weight of the edge, i.e., the self-attention coefficient value calculated by the self-attention mechanism in the transformer network; X μ represents the feature before aggregation; represents the feature after aggregation; i = 1,..., h, h represents the number of self-attention heads; μ ∈ (z, d, c), z, d, and c represent the microseismic field, electric field, and electromagnetic radiation field, respectively; d k represents or the dimension of the softmax represents an activation function, all represent weight matrices, and || represents a concatenation operation; S23: updating the state of each node in the fully connected graph in the field to obtain updated features, and the state update is represented as: wherein, denotes the updated feature; and denotes the Hadamard product; is a weight matrix; S24: repeating steps S21 to S23 to perform multi-layer graph learning on the updated features, and finally obtaining the global time sequence features in single-field data by combining adaptive residuals, and the global time sequence features in single-field data are represented as: wherein, represents global timing features within single field data; represents the node feature matrix after L times of neighbor aggregation and state update; represents the weight matrix.

3. The method of claim 2, wherein, In the step S3, the data relationship between fields is learned through a graph attention network, comprising: S31: Construct an inter-field sparse graph attention network for the microseismic field, electric field, and electromagnetic radiation field. This network integrates the data features X of the microseismic field, electric field, and electromagnetic radiation field. μ Each field is divided into several nodes. The inter-field attention between nodes in different fields is calculated. Each node retains C attention coefficient values ​​with the other two fields from largest to smallest, and sets the remaining attention coefficient values ​​to zero. This results in a sparse graph where any node has an adjacency relationship with C nodes in different fields. In this way, three inter-field sparse graphs are obtained: the microseismic field-electric field sparse graph, the microseismic field-electromagnetic radiation field sparse graph, and the electric field-electromagnetic radiation field sparse graph. Among them, the inter-field nodes with adjacency relationships are connected by edges, and the weight of the edge is the attention coefficient value. S32: aggregating neighbors of nodes in the sparse graph between fields to obtain aggregated features, and the aggregated features are represented as: wherein μ j , μ j′ ∈{z,d,c} and μ j ≠ μ j′ , z, d, c represent microseismic field, electric field and electromagnetic radiation field respectively; denotes the feature of the node pair in the μ j′ field after the nodes in the μ j′ field that have an adjacency relationship with the node pair are aggregated; denotes the feature of the node pair in the μ j field after the nodes in the μ j field that have an adjacency relationship with the node pair are aggregated; and both denote the feature before aggregation; and both denote the weight matrix; S33: updating the state of each node in the sparse graph between fields to obtain updated features, and the state update is represented as: wherein, and both represent updated features, both represent weight matrices; S34: Steps S21-S23 are repeated for the updated features to perform multi-layer graph learning. After the multi-layer graph learning is completed, an adaptive residual is introduced to obtain the final local interaction features between multi-field data and is represented as: wherein, and denote features learned through L-layer graph learning, and denote weight matrices.

4. The method of claim 3, wherein, In step S4, the constrained local interaction features between multi-field data are represented as: where μ j ,μ j′ ,μ j” ∈ {z, d, c}, and μ j ≠ μ j′ ≠ μ j” .

5. The method of claim 4, wherein, In step S5, the fusion features X are represented as:

6. A seismic electromagnetic multi-field data fusion system characterized by, The method comprises the following steps: A feature encoder is used to extract data features of microseismic field, electric field and electromagnetic radiation field from sensor data respectively; A field data relationship learning module is used to learn the data relationship in the field to obtain global time sequence features in single-field data respectively; A field data relationship learning module is used to learn the data relationship between fields in pairs to obtain local interaction features between multi-field data respectively; A consistency loss module is used to constrain the local interaction features between multi-field data through consistency loss to obtain constrained local interaction features between multi-field data; A multi-feature fusion module is used to fuse the global time sequence features in single-field data and the constrained local interaction features between multi-field data to obtain fusion features, and realize the fusion of seismo-electromagnetic multi-field data.

7. An apparatus, comprising: The computer readable storage medium stores a computer program, and the processor executes the computer program to realize the seismo-electromagnetic multi-field data fusion method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the processor executes the computer program to realize the seismo-electromagnetic multi-field data fusion method according to any one of claims 1-5.

Images