A meta-learning-based multimodal retrieval method and system for power samples

Abstract

The embodiment of the present invention provides a meta-learning-based multimodal retrieval method and system for power samples, which belongs to the field of Internet and cloud computing technology. The retrieval method includes: obtaining a multimodal keyword group to be retrieved, wherein the multimodal keyword group includes at least one keyword, and the keyword is a text word or an image; processing the keyword group using a source domain meta-learning network to obtain keyword features; inputting the keyword features into a retrieval network to obtain retrieval features; inputting the retrieval features into a target domain meta-learning network to obtain retrieval results corresponding to the keyword group. Compared with the prior art, the retrieval method and system overcome the technical defect that the prior retrieval method is difficult to adapt to the input of diversified keywords, and improves the retrieval efficiency of the retrieval network for multi-dimensional keywords.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet and cloud computing, and specifically relates to a meta-learning-based multimodal retrieval method and system for power samples. Background Art

[0002] As a highly complex and dynamic system, the data characteristics of the power system are massive, high-dimensional, and diverse in spatiotemporal characteristics. These data range from real-time monitoring, equipment maintenance records to operation logs, etc., and are not only diverse, but each type of data is different in format, structure, and update frequency, which poses a great challenge to the effective organization and rapid retrieval of data. Traditional information retrieval techniques, such as keyword search based on Boolean models, although efficient in processing structured text data, often seem to be unable to cope with the complex data types in power systems. These traditional methods usually ignore the correlation and contextual information between data, resulting in retrieval results that often cannot accurately match the actual retrieval needs, especially when accurate information is needed to support decision-making, the problems of retrieval efficiency and accuracy are particularly prominent. Summary of the invention

[0003] The purpose of the embodiments of the present invention is to provide a meta-learning-based multimodal retrieval method and system for power samples, which can improve the keyword retrieval efficiency of the power system.

[0004] In order to achieve the above object, an embodiment of the present invention provides a meta-learning-based multimodal retrieval method for power samples, comprising:

[0005] Acquire a multimodal keyword group to be searched, wherein the multimodal keyword group includes at least one keyword, and the keyword is a text word or an image;

[0006] Processing the keyword group using a source domain meta-learning network to obtain keyword features;

[0007] Inputting the keyword feature into a search network to obtain a search feature;

[0008] The retrieval features are input into a target domain meta-learning network to obtain retrieval results corresponding to the keyword group.

[0009] Optionally, the source domain meta-learning network includes:

[0010] Input layer, used to obtain input keywords;

[0011] A feature extraction layer, used for performing a feature extraction operation on the keyword to obtain a first feature;

[0012] A feature mapping layer, used for performing a multi-dimensional feature mapping operation on the first feature to obtain feature matrices in different dimensions;

[0013] The feature combination layer is used to combine the feature matrices to obtain the keyword features.

[0014] Optionally, the multi-dimensional feature mapping operation includes:

[0015] Different keywords are mapped into real-value space and hash space respectively to obtain the corresponding feature matrix.

[0016] Optionally, the feature combination layer is used to:

[0017] Calculate the distances of the feature matrices in the real-valued space and the hash space respectively to obtain corresponding distance matrices;

[0018] The feature matrix and the distance matrix are combined to obtain the keyword features.

[0019] Optionally, before combining the feature matrix and the distance matrix, the retrieval method further includes:

[0020] Use formula (1) to formula (4) for weighted operation:

[0021] , (1)

[0022] , (2)

[0023] , (3)

[0024] , (4)

[0025] in, is the weighted feature matrix, is the attention weight, is the feature matrix feature vectors, is the weighted distance matrix, is the distance matrix feature vectors, is the normalization parameter, Represents a transpose operation.

[0026] Optionally, the target domain meta-learning network includes:

[0027] An input layer, used for receiving the input retrieval features;

[0028] The embedding layer is used to map the input features into a high-dimensional space to obtain feature embedding;

[0029] A multi-head self-attention mechanism for capturing the global dependencies between the feature embeddings in the high-dimensional space to calculate corresponding attention scores;

[0030] A feed-forward neural network layer for further computing the feature embedding;

[0031] The feature mapping layer is used to map the calculation results of the multi-head self-attention mechanism and the feedforward neural network layer into the real-valued space and hash space to obtain feature matrices of different dimensions;

[0032] Feature classification layer, used to:

[0033] Calculating the distances between the feature matrix and the preset category matrix respectively;

[0034] The category corresponding to the retrieval feature is determined according to the distance.

[0035] Optionally, the category corresponding to the search feature according to the distance includes:

[0036] According to formula (5), the comprehensive distance between the retrieval feature and each category is calculated.

[0037] , (5)

[0038] in, To retrieve features With category The comprehensive distance, To retrieve features With category The distance in real-valued space, To retrieve features With category Distance in hash space;

[0039] The probability that the retrieval feature belongs to each category is calculated according to formula (6):

[0040] , (6)

[0041] in, The search feature Belongs to category probability.

[0042] Optionally, the retrieval method further includes:

[0043] The distance is calculated according to formula (7),

[0044] , (7)

[0045] in, is the distance, is the Hamming distance, is the Euclidean distance.

[0046] On the other hand, the present invention also provides a meta-learning-based multimodal retrieval system for power samples, the retrieval system comprising a processor, and the processor is configured to execute any of the retrieval methods described above.

[0047] In yet another aspect, the present invention further provides a computer-readable storage medium storing instructions, wherein the instructions are used to be read by a machine so that the machine executes any of the above-mentioned retrieval methods.

[0048] Through the above technical solution, the embodiment of the present invention provides a meta-learning-based multimodal retrieval method and system for power samples. The retrieval method and system respectively construct meta-learning networks in the source domain and target domain of the retrieval network, thereby extracting and integrating features of diversified keywords and realizing diversified classification of diversified classification results. Compared with the prior art, the retrieval method and system overcome the technical defect that the existing retrieval method is difficult to adapt to the input of diversified keywords, and improves the retrieval efficiency of the retrieval network for multi-dimensional keywords.

[0049] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific implementations, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:

[0051] Figure 1 is a flowchart of a method for multimodal retrieval of power samples based on meta-learning according to an embodiment of the present invention;

[0052] Figure 2 is a structural block diagram of a source domain meta-learning network according to an embodiment of the present invention;

[0053] Figure 3 is a structural block diagram of a target domain meta-learning network according to an embodiment of the present invention;

[0054] Figure 4 is a schematic diagram of a meta-learning-based multimodal retrieval method for power samples according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The specific implementation of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present invention, and is not used to limit the embodiment of the present invention.

[0056] like Figure 1 FIG. 1 is a flow chart of a method for multimodal retrieval of power samples based on meta-learning according to an embodiment of the present invention. Figure 1 In the retrieval method, the retrieval method may include the following steps:

[0057] In step S10, a multimodal keyword group to be searched is obtained, wherein the multimodal keyword group includes at least one keyword, and the keyword is a text word or an image;

[0058] In step S11, the source domain meta-learning network is used to process the keyword group to obtain keyword features;

[0059] In step S12, the keyword feature is input into the search network to obtain the search feature;

[0060] In step S13, the search features are input into the target domain meta-learning network to obtain the search results corresponding to the keyword group.

[0061] In this Figure 1 In the retrieval method shown, step S10 can be used to obtain a multimodal keyword group to be retrieved. Considering the diversified characteristics of power data samples of power systems in the prior art, the keyword group may include at least one keyword, and the keyword may be a text word or an image.

[0062] Step S11 can be used to process the keyword group using the source domain meta-learning network to obtain keyword features, so as to form a standard keyword feature for the multimodal keyword group input at the input end. The source domain meta-learning network can be in various forms known to those skilled in the art. In one example of the present invention, the source domain meta-learning network can include: Figure 2 The structure shown in Figure 2In the source domain meta-learning network, the source domain meta-learning network may include an input layer, a feature extraction layer, a feature mapping layer, and a feature combination layer. The input layer may be used to obtain input keywords. The feature mapping layer may be used to perform a feature extraction operation on the input keywords to obtain a first feature, thereby realizing feature extraction of different keywords. The feature mapping layer may be used to perform a multi-dimensional feature mapping operation on the first feature to obtain a feature matrix under different dimensions. The feature combination layer may be used to combine the feature matrices to obtain keyword features. The multi-dimensional feature mapping operation performed by the feature mapping layer on the first feature may be in a variety of forms known to those skilled in the art. In one example of the present invention, the multi-dimensional feature mapping operation may be to map different keywords to a real-valued space and a hash space, respectively, to obtain corresponding feature matrices. Accordingly, the feature combination layer may be used to first calculate the distances of each feature matrix in the real-valued space and the hash space, respectively, to obtain a corresponding distance matrix, and then combine the feature matrix and the distance matrix to obtain keyword features. In addition, considering that if the feature aggregation and distance matrix to be combined are directly combined, the keywords in different fields themselves have the problem of weak data migration ability, before combining the feature matrix and the distance matrix, a weight matrix can be used to perform weighted operations on the distance matrix and each feature matrix respectively. As for the specific method of the weighted operation, in an example of the present invention, the following formulas (1) to (4) can be used to perform the weighted operation:

[0063] , (1)

[0064] , (2)

[0065] , (3)

[0066] , (4)

[0067] in, is the weighted feature matrix, is the attention weight, is the feature matrix feature vectors, is the weighted distance matrix, is the distance matrix feature vectors, is a normalization factor used to adjust the similarity measure between the feature matrix and the distance matrix. represents the distance measure or scaling factor to ensure that the result of the weighted operation is within a reasonable range. is the transpose operation.

[0068] Step S12 can be used to input keyword features into a search network to obtain search features. The search network can be in various forms known to those skilled in the art, including but not limited to PRank method, RankNet network, ListNet network, LambdaMART network, etc.

[0069] Step S13 may be used to input the search features into the target domain meta-learning network to obtain the search results corresponding to the keyword phrase. The target domain meta-learning network may be a variety of structures known to those skilled in the art. In one example of the present invention, the target domain meta-learning network may include: Figure 3 The structure shown in Figure 3 In the method, the target meta-learning network may include an input layer, an embedding layer, a multi-head self-attention mechanism, a feedforward neural network layer, a feature mapping layer, and a feature classification layer. Specifically, the input layer may be used to receive input retrieval features. The embedding layer may be used to map the input features to a high-dimensional space to obtain feature embeddings. The multi-head self-attention mechanism may be used to capture the global dependencies between feature embeddings in a high-dimensional space to calculate the corresponding attention scores. The feedforward neural network layer may be used to perform further calculations on the feature embeddings. The feature mapping layer may be used to map the calculation results of the multi-head self-attention mechanism and the feedforward neural network layer to a real-valued space and a hash space to obtain feature matrices of different dimensions. The feature classification layer may be used to respectively calculate the distance between the feature matrix and a preset category matrix, and retrieve the category corresponding to the feature based on the distance. Furthermore, the feature classification layer may be firstly calculated according to formula (5) to obtain the comprehensive distance between the retrieval feature and each category,

[0070] , (5)

[0071] in, To retrieve features With category The comprehensive distance, To retrieve features With category The distance in real-valued space, To retrieve features With category Distance in hash space;

[0072] Then, the probability that the retrieval feature belongs to each category is calculated according to formula (6):

[0073] , (6)

[0074] in, To retrieve features Belongs to category probability.

[0075] In addition, the specific calculation method of the above distance can be various forms known to those skilled in the art, including but not limited to Euclidean distance, cosine similarity distance, etc. In one example of the present invention, the distance can be calculated using the following formula (7):

[0076] , (7)

[0077] in, For this distance, is the Hamming distance, is the Euclidean distance.

[0078] Based on the above Figures 1 to 3 The method shown in the figure can also be as follows: Figure 4 As shown. Figure 4 In the retrieval method, a multimodal knowledge graph is used to construct keyword input, retrieval features, and classification results, namely, operations of sample feature representation, multimodal representation fusion, and multimodal representation alignment. The meta-training task is trained through a meta-learning network located in the source domain, and the meta-test task is trained through a target domain meta-learning network located in the target domain. The retrieval feature classification is then realized in combination with the retrieval network, thereby realizing the retrieval operation of multimodal keywords.

[0079] On the other hand, the present invention also provides a meta-learning-based multimodal retrieval system for power samples, the retrieval system comprising a processor, the processor being configured to execute any of the retrieval methods described above. Specifically, the retrieval method may include the following steps:

[0080] In step S10, a multimodal keyword group to be searched is obtained, wherein the multimodal keyword group includes at least one keyword, and the keyword is a text word or an image;

[0081] In step S11, the source domain meta-learning network is used to process the keyword group to obtain keyword features;

[0082] In step S12, the keyword feature is input into the search network to obtain the search feature;

[0083] In step S13, the search features are input into the target domain meta-learning network to obtain the search results corresponding to the keyword group.

[0084] In this Figure 1 In the retrieval method shown, step S10 can be used to obtain a multimodal keyword group to be retrieved. Considering the diversified characteristics of power data samples of power systems in the prior art, the keyword group may include at least one keyword, and the keyword may be a text word or an image.

[0085] Step S11 can be used to process the keyword group using the source domain meta-learning network to obtain keyword features, so as to form a standard keyword feature for the multimodal keyword group input at the input end. The source domain meta-learning network can be in various forms known to those skilled in the art. In one example of the present invention, the source domain meta-learning network can include: Figure 2 The structure shown in Figure 2 In the source domain meta-learning network, the source domain meta-learning network may include an input layer, a feature extraction layer, a feature mapping layer, and a feature combination layer. The input layer may be used to obtain input keywords. The feature mapping layer may be used to perform a feature extraction operation on the input keywords to obtain a first feature, thereby realizing feature extraction of different keywords. The feature mapping layer may be used to perform a multi-dimensional feature mapping operation on the first feature to obtain a feature matrix under different dimensions. The feature combination layer may be used to combine the feature matrices to obtain keyword features. The multi-dimensional feature mapping operation performed by the feature mapping layer on the first feature may be in a variety of forms known to those skilled in the art. In one example of the present invention, the multi-dimensional feature mapping operation may be to map different keywords to a real-valued space and a hash space, respectively, to obtain corresponding feature matrices. Accordingly, the feature combination layer may be used to first calculate the distances of each feature matrix in the real-valued space and the hash space, respectively, to obtain a corresponding distance matrix, and then combine the feature matrix and the distance matrix to obtain keyword features. In addition, considering that if the feature aggregation and distance matrix to be combined are directly combined, the keywords in different fields themselves have the problem of weak data migration ability, before combining the feature matrix and the distance matrix, a weight matrix can be used to perform weighted operations on the distance matrix and each feature matrix respectively. As for the specific method of the weighted operation, in an example of the present invention, the following formulas (1) to (4) can be used to perform the weighted operation:

[0086] , (1)

[0087] , (2)

[0088] , (3)

[0089] , (4)

[0090] in, is the weighted feature matrix, is the attention weight, is the feature matrix feature vectors, is the weighted distance matrix, is the distance matrix feature vectors, is a normalization factor used to adjust the similarity measure between the feature matrix and the distance matrix. represents the distance measure or scaling factor to ensure that the result of the weighted operation is within a reasonable range. is the transpose operation.

[0091] Step S12 can be used to input keyword features into a search network to obtain search features. The search network can be in various forms known to those skilled in the art, including but not limited to PRank method, RankNet network, ListNet network, LambdaMART network, etc.

[0092] Step S13 may be used to input the search features into the target domain meta-learning network to obtain the search results corresponding to the keyword phrase. The target domain meta-learning network may be a variety of structures known to those skilled in the art. In one example of the present invention, the target domain meta-learning network may include: Figure 3 The structure shown in Figure 3 In the method, the target meta-learning network may include an input layer, an embedding layer, a multi-head self-attention mechanism, a feedforward neural network layer, a feature mapping layer, and a feature classification layer. Specifically, the input layer may be used to receive input retrieval features. The embedding layer may be used to map the input features to a high-dimensional space to obtain feature embeddings. The multi-head self-attention mechanism may be used to capture the global dependencies between feature embeddings in a high-dimensional space to calculate the corresponding attention scores. The feedforward neural network layer may be used to perform further calculations on the feature embeddings. The feature mapping layer may be used to map the calculation results of the multi-head self-attention mechanism and the feedforward neural network layer to a real-valued space and a hash space to obtain feature matrices of different dimensions. The feature classification layer may be used to respectively calculate the distance between the feature matrix and a preset category matrix, and retrieve the category corresponding to the feature based on the distance. Furthermore, the feature classification layer may be firstly calculated according to formula (5) to obtain the comprehensive distance between the retrieval feature and each category,

[0093] , (5)

[0094] in, To retrieve features With category The comprehensive distance, To retrieve features With category The distance in real-valued space, To retrieve features With category Distance in hash space;

[0095] Then, the probability that the retrieval feature belongs to each category is calculated according to formula (6):

[0096] , (6)

[0097] in, To retrieve features Belongs to category probability.

[0098] In addition, the specific calculation method of the above distance can be various forms known to those skilled in the art, including but not limited to Euclidean distance, cosine similarity distance, etc. In an example of the present invention, the distance can be calculated using the following formula (7):

[0099] , (7)

[0100] in, For this distance, is the Hamming distance, is the Euclidean distance.

[0101] Based on the above Figures 1 to 3 The method shown in the figure can also be as follows: Figure 4 As shown. Figure 4 In the retrieval method, a multimodal knowledge graph is used to construct keyword input, retrieval features, and classification results, namely, operations of sample feature representation, multimodal representation fusion, and multimodal representation alignment. The meta-training task is trained through a meta-learning network located in the source domain, and the meta-test task is trained through a target domain meta-learning network located in the target domain. The retrieval feature classification is then realized in combination with the retrieval network, thereby realizing the retrieval operation of multimodal keywords.

[0102] In yet another aspect, the present invention further provides a computer-readable storage medium storing instructions, wherein the instructions are used to be read by a machine so that the machine executes any of the above-mentioned retrieval methods.

[0103] Through the above technical solution, the embodiment of the present invention provides a meta-learning-based multimodal retrieval method and system for power samples. The retrieval method and system respectively construct meta-learning networks in the source domain and target domain of the retrieval network, thereby extracting and integrating features of diversified keywords and realizing diversified classification of diversified classification results. Compared with the prior art, the retrieval method and system overcome the technical defect that the existing retrieval method is difficult to adapt to the input of diversified keywords, and improves the retrieval efficiency of the retrieval network for multi-dimensional keywords.

[0104] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0105] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0106] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0107] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0108] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0109] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0110] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0111] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0112] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A meta-learning-based multimodal retrieval method for power samples, characterized in that: The retrieval method comprises: Acquire a multimodal keyword group to be searched, wherein the multimodal keyword group includes at least one keyword, and the keyword is a text word or an image; Processing the keyword group using a source domain meta-learning network to obtain keyword features; Inputting the keyword feature into a search network to obtain a search feature; Inputting the search features into a target domain meta-learning network to obtain search results corresponding to the keyword group; The source domain meta-learning network includes: Input layer, used to obtain input keywords; A feature extraction layer, used for performing a feature extraction operation on the keyword to obtain a first feature; A feature mapping layer, used for performing a multi-dimensional feature mapping operation on the first feature to obtain feature matrices in different dimensions; A feature combination layer, used for combining the feature matrices to obtain the keyword features; The multi-dimensional feature mapping operation includes: Map different keywords into real-value space and hash space respectively to obtain the corresponding feature matrix; The feature combination layer is used to: Calculate the distances of the feature matrices in the real-valued space and the hash space respectively to obtain corresponding distance matrices; Combining the feature matrix and the distance matrix to obtain the keyword features; Before combining the feature matrix and the distance matrix, the retrieval method further includes: Use formula (1) to formula (4) for weighted operation: ，（1） ，（2） ，（3） ，（4） in, is the weighted feature matrix, is the attention weight, is the feature matrix feature vectors, is the weighted distance matrix, is the distance matrix feature vectors, is the normalization parameter, Represents a transpose operation.

2. The search method according to claim 1, characterized in that: The target domain meta-learning network includes: An input layer, used for receiving the input retrieval features; The embedding layer is used to map the input features into a high-dimensional space to obtain feature embedding; A multi-head self-attention mechanism for capturing the global dependencies between the feature embeddings in the high-dimensional space to calculate corresponding attention scores; A feed-forward neural network layer for further computing the feature embedding; The feature mapping layer is used to map the calculation results of the multi-head self-attention mechanism and the feedforward neural network layer into the real-valued space and hash space to obtain feature matrices of different dimensions; Feature classification layer, used to: Calculating the distances between the feature matrix and the preset category matrix respectively; The category corresponding to the retrieval feature is determined according to the distance.

3. The search method according to claim 2, characterized in that: The categories corresponding to the search features according to the distance include: According to formula (5), the comprehensive distance between the retrieval feature and each category is calculated. ，（5） in, To retrieve features With category The comprehensive distance, To retrieve features With category The distance in real-valued space, To retrieve features With category Distance in hash space; The probability that the retrieval feature belongs to each category is calculated according to formula (6): ，（6） in, The search feature Belongs to category probability.

4. The search method according to claim 3, characterized in that: The retrieval method further comprises: The distance is calculated according to formula (7), ，（7） in, is the distance, is the Hamming distance, is the Euclidean distance, , is the corresponding weight.

5. A meta-learning-based multimodal retrieval system for power samples, characterized in that: The retrieval system comprises a processor, and the processor is configured to execute the retrieval method according to any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and the instructions are used to be read by a machine so that the machine executes the retrieval method according to any one of claims 1 to 4.

Images