Feature enhancement method and system for power multi-service sample
By preprocessing and hierarchical feature extraction of multi-source heterogeneous data of power multi-service samples, combined with the weighted fusion of gated attention units, the data imbalance and insufficient features in the analysis of power multi-service samples in the prior art are solved, and the feature expression ability and service processing effect are significantly improved.
Patent Information
- Application Number
- CN202510275916.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-10
- Publication Date
- 2025-06-27
AI Technical Summary
The existing multi-service sample analysis of power services has problems such as data imbalance, incomplete features, large noise interference and low feature dimensions, which is difficult to accurately reflect the true status and laws of power services, which limits the efficient development of power services and the improvement of intelligence level.
By obtaining real-time multi-source heterogeneous data, the data is preprocessed and outliers are eliminated; a variety of feature extraction algorithms are used for hierarchical feature extraction, including LSTM, Transformer, GNN and conditional GAN algorithms, and a variety of feature sources are generated; the gated attention unit is used to weight and fuse the feature sources to obtain enhanced features and input them into the business sample processing model.
It significantly improves the feature expression ability in complex business scenarios, making sample features more accurate and diverse, and meets the needs of business scenarios such as stable operation of power systems and fault prediction.
Smart Images
Figure CN120217288A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power system data analysis, and in particular to a method and system for enhancing features of power multi-service samples. Background Art
[0002] With the continuous development of the power system, power multi-service scenarios are becoming increasingly complex, posing higher requirements for the analysis and processing of power service sample data. In the analysis of power multi-service samples, the data sources of the samples are different, the data is unbalanced, and at the same time, the quality of the sample features directly affects the accuracy and reliability of subsequent data analysis, model training, and business decision-making. However, existing power multi-service samples often have problems such as complex multi-service data, scarce sample data of some services, incomplete features, large noise interference, and low feature dimensions, making it difficult to accurately reflect the real state and laws of power services, restricting the efficient development of power services and the improvement of the intelligent level. In the prior art, traditional data enhancement methods do not consider physical constraints such as power conservation, and the generated data and features violate the actual operation rules. At the same time, there is a lack of a method for enhancing the sample features of power services, and the enhancement processing of features is insufficient, making it difficult to meet the actual needs and bringing difficulties to power dispatching and resource allocation. Summary of the Invention
[0003] To overcome the above problems, the present invention provides a method for enhancing features of power multi-service samples. The method obtains real-time multi-source heterogeneous data, preprocesses the multi-source heterogeneous data to eliminate outliers, improves the quality and stability of the data, and performs hierarchical feature extraction using multiple feature extraction algorithms according to the data components of the multi-source heterogeneous data to obtain feature sources, making the sample features more accurate and diverse. A gated attention unit is used to perform weighted fusion on the feature sources to obtain enhanced features, realizing the integration of features, significantly improving the feature expression ability in complex service scenarios, and inputting the enhanced features into a service sample processing model to complete service processing, meeting the requirements of service scenarios such as the stable operation and fault prediction of the power system.
[0004] To achieve the above object, on the one hand, the present invention provides a method for enhancing features of power multi-service samples, and the feature enhancement method includes:
[0005] Obtain real-time multi-source heterogeneous data and preprocess the multi-source heterogeneous data;
[0006] Perform hierarchical feature extraction according to the data components of the multi-source heterogeneous data to obtain feature sources;
[0007] Use a gated attention unit to perform weighted fusion on the feature sources to obtain enhanced features;
[0008] Input the enhanced features into the business sample processing model to complete business processing.
[0009] Preferably, the obtaining of real-time multi-source heterogeneous data and the preprocessing of the multi-source heterogeneous data include:
[0010] The multi-source heterogeneous data includes time-series data and multi-modal data;
[0011] Perform cleaning and removal processing on the multi-source heterogeneous data;
[0012] Perform standardization and normalization processing on the processed multi-source heterogeneous data.
[0013] Preferably, hierarchical feature extraction is performed according to the data components of the multi-source heterogeneous data to obtain feature sources, including:
[0014] Adopt LSTM and Transformer algorithms for the time-series data to extract the first feature source of time-series autocorrelation features;
[0015] Adopt the GNN algorithm for the multi-modal data to extract the second feature source of cross-device association features;
[0016] Adopt the conditional GAN algorithm for the time-series data and multi-modal data to extract the third feature source of physical constraint features.
[0017] Preferably, a gated attention unit is used to perform weighted fusion on the feature sources to obtain enhanced features, including:
[0018] Map multiple feature sources to the same-dimensional space for feature alignment, and perform position encoding to retain the time-order information of the time-series autocorrelation features;
[0019] The gated attention unit performs independent linear projection on the feature sources through formula group (1) to obtain generated features,
[0020]
[0021] where, Z q is the generated feature of the first feature source, W q is the learnable weight matrix of the first feature source, F q is the input feature of the first feature source, b q is the bias term of the first feature source, Z m is the generated feature of the second feature source, W m is the learnable weight matrix of the second feature source, F m is the input feature of the second feature source, b m is the bias term of the second feature source, Zp The generated feature for the third feature source, W p The learnable weight matrix for the third feature source, F p The input feature for the third feature source, b p The bias term for the third feature source;
[0022] Process the generated features corresponding to multiple feature sources according to the formula group (2) to obtain the attention scores of multiple feature sources, and normalize the attention scores to obtain the weighted weights of multiple feature sources,
[0023]
[0024] where, α q is the weighted weight of the first feature source, α m is the weighted weight of the second feature source, α p is the weighted weight of the third feature source, W a is the weight matrix, b a is the bias term;
[0025] Perform weighted summation on the feature sources according to the generated weighted weights using formula (3) to obtain the enhanced feature,
[0026] F fusion = α q * F q + α m * F m + α p * F p , (3)
[0027] where, F fusion is the enhanced feature, α q is the weighted weight of the first feature source, α m is the weighted weight of the second feature source, α p is the weighted weight of the third feature source, F q is the input feature of the first feature source, F m is the input feature of the second feature source, F p is the input feature of the third feature source.
[0028] Preferably, the first feature source for extracting the temporal autocorrelation feature from the temporal data using the LSTM and Transformer algorithms includes:
[0029] Process the temporal data using the LSTM algorithm to obtain the first output,
[0030] h t = LSTM(xt , h t-1 ),
[0031] where h t is the first output of the timing data at time t, x t is the input of the timing data at time t, and h t-1 is the output of the timing data at time t - 1;
[0032] The first feature source of the fused temporal autocorrelation feature is obtained by processing the first output using the Transformer algorithm.
[0033]
[0034] s t = ∑(α t * h t )
[0035] where α t is the weight exponent of the attention mechanism layer at time t, w1 is the weight matrix of the attention mechanism layer, b1 is the bias term of the attention mechanism layer, and s t is the first feature source of the temporal autocorrelation feature at time t.
[0036] Preferably, the second feature source of the cross-device association feature is extracted from the multimodal data using the GNN algorithm, including:
[0037] Abstract the elements of the data of the device into graph nodes according to the multimodal data, locate the association relationship between the devices as edges, and independently extract the features of the initialized graph nodes from the multimodal data;
[0038] Concatenate the features of the corresponding graph nodes of the multimodal data according to the channel dimension to obtain a joint feature matrix;
[0039] Use a graph convolutional layer to aggregate the node features of the joint feature matrix, and at the same time fuse them with the features of the initialized graph nodes to obtain a fused feature. Project the fused feature into the subspace between the devices and extract it to obtain the second feature source of the cross-device association feature in the features of the graph nodes.
[0040] Preferably, the third feature source of the physical constraint feature is extracted from the timing data and multimodal data using the conditional GAN algorithm, including:
[0041] Establish a physical constraint discretization mathematical model according to the device type;
[0042] Preprocess the timing data and multimodal data;
[0043] Use the conditional GAN algorithm to process the preprocessed time series data and multi-modal data, generate and verify synthetic data that meets physical constraints;
[0044] Extract the time series autocorrelation features and cross-device association features from the synthetic data, and filter the features that conform to the discretized mathematical model of the physical constraints, and use them as the third feature source of the physical constraint features.
[0045] Preferably, using the conditional GAN algorithm to process the preprocessed time series data and multi-modal data, generating and verifying synthetic data that meets physical constraints, includes:
[0046] Design a constraint condition function;
[0047] Use the conditional GAN algorithm to generate synthetic data according to physical constraint labels;
[0048] Use the constraint condition function to perform constraint verification on the synthetic data;
[0049] Retain the synthetic data that passes the verification.
[0050] The second aspect of the present invention provides a feature enhancement system for power multi-service samples, and the feature enhancement system includes:
[0051] A data processing module for preprocessing multi-source heterogeneous data;
[0052] A feature enhancement module, electrically connected to the data processing module, for executing the feature enhancement method according to any one of the above claims to obtain enhanced features;
[0053] A service adaptation interface, electrically connected to the feature enhancement module, for inputting the enhanced features into a downstream service sample processing model.
[0054] Preferably, the service sample processing model includes at least one of a classifier, a regression model, or an anomaly detection model.
[0055] Through the above technical solution, the method obtains real-time multi-source heterogeneous data, preprocesses the multi-source heterogeneous data to eliminate outliers, improves the quality and stability of the data, and performs hierarchical feature extraction on the data components of the multi-source heterogeneous data using various feature extraction algorithms to obtain feature sources. For time-series data, the first feature source of time-series autocorrelation features is extracted using LSTM and Transformer algorithms. For modal data, the second feature source of cross-device association features is extracted using the GNN algorithm. For time-series data and multi-modal data, the third feature source of physical constraint features is extracted using the conditional GAN algorithm, generating new features that are similar to but slightly different from the original sample features. The new features are merged with the original features to further enrich the feature set, making the sample features more accurate and diverse. The gated attention unit is used to perform weighted fusion on the feature sources to obtain enhanced features, achieving the integration of features, significantly improving the feature expression ability in complex business scenarios, and inputting the enhanced features into the business sample processing model to complete business processing, meeting the requirements of business scenarios such as the stable operation and fault prediction of the power system. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 is a flowchart of a method for feature enhancement of power multi-business samples according to an embodiment of the present invention;
[0057] Figure 2 is a schematic diagram of the extraction process of the third feature source of a method for feature enhancement of power multi-business samples according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0058] The following details the specific embodiments of the embodiments of the present invention with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining and illustrating the embodiments of the present invention, and are not used to limit the embodiments of the present invention.
[0059] Figure 1 Shown is a flowchart of a method for feature enhancement of power multi-business samples according to an embodiment of the present invention. In Figure 1 it, the feature enhancement method may include:
[0060] In step S10, real-time multi-source heterogeneous data is obtained and preprocessed;
[0061] In step S11, hierarchical feature extraction is performed according to the data components of the multi-source heterogeneous data to obtain feature sources;
[0062] In step S12, a gated attention unit is used to perform weighted fusion on the feature sources to obtain enhanced features;
[0063] In step S13, the enhanced features are input into the business sample processing model to complete the business processing.
[0064] In the method as Figure 1 shown, step S10 obtains real-time multi-source heterogeneous data, preprocesses the multi-source heterogeneous data, eliminates outliers, removes noise, and improves the accuracy and quality of the data.
[0065] Step S11 uses different feature extraction algorithms according to different data for feature extraction, ensures the local and global integration of features, improves robustness, and the generated synthetic sample data can achieve feature enhancement of scarce samples and improve the diversity of samples.
[0066] Step S12 uses a gated attention unit to perform weighted fusion on multiple feature sources to obtain enhanced features, and performs weighted distribution according to different feature sources, realizing the integration of features and significantly improving the feature expression ability in complex business scenarios.
[0067] Step S13 completes the business processing by inputting the enhanced features into the business sample processing model, meets the requirements of business scenarios such as the stable operation and fault prediction of the power system, and improves adaptability.
[0068] Through the above technical solution, the method obtains real-time multi-source heterogeneous data, preprocesses the multi-source heterogeneous data, eliminates outliers, improves the quality and stability of the data, uses multiple feature extraction algorithms for hierarchical feature extraction according to the data components of the multi-source heterogeneous data to obtain feature sources, makes the sample features more accurate and diverse, uses a gated attention unit to perform weighted fusion on the feature sources to obtain enhanced features, realizes the integration of features, significantly improves the feature expression ability in complex business scenarios, and inputs the enhanced features into the business sample processing model to complete the business processing, meeting the requirements of business scenarios such as the stable operation and fault prediction of the power system.
[0069] Considering the preprocessing of multi-source heterogeneous data to ensure the quality and accuracy of the data, in an embodiment of the present invention, obtaining real-time multi-source heterogeneous data and preprocessing the multi-source heterogeneous data may include the following steps: The multi-source heterogeneous data includes time-series data and multi-modal data. Clean and remove the multi-source heterogeneous data, perform standardization and normalization processing on the processed multi-source heterogeneous data, classify the data according to labels, remove invalid data and outliers, ensure the consistency of the data, and facilitate subsequent data analysis on the same scale.
[0070] Considering that different methods are used for feature extraction of different data, and at the same time, the features of scarce sample data can be supplemented and enhanced. In an embodiment of the present invention, the steps for obtaining feature sources by performing hierarchical feature extraction according to the data components of multi-source heterogeneous data may include the following: extracting a first feature source of temporal autocorrelation features from temporal data by using LSTM and Transformer algorithms; extracting a second feature source of cross-device association features from multi-modal data by using GNN algorithms; extracting a third feature source of physical constraint features from temporal data and multi-modal data by using conditional GAN algorithms. Feature extraction from local to global is performed by using feature extraction methods that conform to different data, mining the effective information in the data, providing a rich, accurate, and practical feature basis for feature enhancement of power multi-service samples, and greatly improving the accuracy and reliability of power service analysis and processing.
[0071] In order to fuse the features of multiple feature sources to obtain enhanced features, ensure the accuracy of feature integration, and guarantee the effectiveness of feature information. In an embodiment of the present invention, the steps for obtaining enhanced features by performing weighted fusion of feature sources by using a gated attention unit may include the following:
[0072] Mapping multiple feature sources to the same-dimensional space for feature alignment, and performing positional encoding to retain the time-order information of temporal autocorrelation features;
[0073] The gated attention unit performs independent linear projection on the feature sources through formula group (1) to obtain generated features,
[0074]
[0075] where Z q is the generated feature of the first feature source, W q is the learnable weight matrix of the first feature source, F q is the input feature of the first feature source, b q is the bias term of the first feature source, Z m is the generated feature of the second feature source, W m is the learnable weight matrix of the second feature source, F m is the input feature of the second feature source, b m is the bias term of the second feature source, Z p is the generated feature of the third feature source, W p is the learnable weight matrix of the third feature source, F p is the input feature of the third feature source, b p is the bias term of the third feature source;
[0076] Process the generated features corresponding to multiple feature sources according to the formula set (2) to obtain the attention scores of multiple feature sources, and normalize the attention scores to obtain the weighted weights of multiple feature sources.
[0077]
[0078] Among them, α q is the weighted weight of the first feature source, α m is the weighted weight of the second feature source, α p is the weighted weight of the third feature source, W a is the weight matrix, b a is the bias term;
[0079] Perform weighted summation on the feature sources according to the generated weighted weights using formula (3) to obtain the enhanced feature.
[0080] F fusion =α q *F q +α m *F m +α p *F p , (3)
[0081] Among them, F fusion is the enhanced feature, α q is the weighted weight of the first feature source, α m is the weighted weight of the second feature source, α p is the weighted weight of the third feature source, F q is the input feature of the first feature source, F m is the input feature of the second feature source, F p is the input feature of the third feature source.
[0082] Map multiple feature sources to the same dimensional space for feature alignment, then perform independent linear projection on the feature sources to obtain the generated features. Then, obtain the attention scores corresponding to multiple feature sources according to the generated features, and normalize them to obtain the corresponding weighted weights of multiple feature sources. Perform feature fusion according to the weighted weights of multiple feature sources to obtain the enhanced feature.
[0083] In order to effectively extract the features of time series data, in an embodiment of the present invention, the steps for the first feature source to extract the time series autocorrelation features from the time series data using the LSTM and Transformer algorithms may include: processing the time series data using the LSTM algorithm to obtain the first output.
[0084] h t =LSTM(x t , h t-1 ),
[0085] Among them, h t is the first output of the time-series data at time t, x t is the input of the time-series data at time t, and h t-1 is the output of the time-series data at time t-1;
[0086] The first output is processed by the Transformer algorithm to obtain the first feature source of the fused temporal autocorrelation feature,
[0087]
[0088] s t = Σ(α t * h t )
[0089] Among them, α t is the weight exponent of the attention mechanism layer at time t, w1 is the weight matrix of the attention mechanism layer, b1 is the bias term of the attention mechanism layer, and s t is the first feature source of the temporal autocorrelation feature at time t; The combined use of the LSTM and Transformer algorithms can improve the extraction accuracy of the temporal autocorrelation feature of the time-series data.
[0090] In order to effectively extract the features of multimodal data, in an embodiment of the present invention, the steps of using the GNN algorithm to extract the second feature source of the cross-device association feature from the multimodal data may include: abstracting the elements of the data of the device into graph nodes according to the multimodal data, positioning the association relationship between devices as edges, and independently extracting the features of the initialized graph nodes from the multimodal data; concatenating the features of the corresponding graph nodes of the multimodal data according to the channel dimension to obtain a joint feature matrix; using a graph convolutional layer to perform node feature aggregation on the joint feature matrix, and at the same time fusing with the features of the initialized graph nodes to obtain a fused feature, projecting the fused feature into the subspace between devices and extracting it to obtain the second feature source of the cross-device association feature in the features of the graph nodes. Processing graph-like data with devices as nodes, capturing complex connection relationships in multimodal data, and at the same time fusing association information between different modalities to extract cross-device association features in multimodal data.
[0091] As Figure 2 shown is a schematic diagram of the extraction process of the third feature source of a feature enhancement method for power multi-service samples according to an embodiment of the present invention; In Figure 2 , in order to be able to extract physical constraint features from time-series data and multimodal data, in an embodiment of the present invention, the steps of using the conditional GAN algorithm to extract the third feature source of physical constraint features from time-series data and multimodal data may include:
[0092] In step S20, a physical constraint discretization mathematical model is established according to the device type;
[0093] In step S21, the time-series data and multi-modal data are preprocessed;
[0094] In step S22, the conditional GAN algorithm is used to process the preprocessed time-series data and multi-modal data to generate and verify synthetic data that conforms to the physical constraints;
[0095] In step S23, the time-series autocorrelation features and cross-device association features of the synthetic data are extracted, and the features that conform to the physical constraint discretization mathematical model are screened and used as the third feature source of the physical constraint features.
[0096] Considering that the data differences of different devices are large, different physical constraint discretization mathematical models are established according to different devices. At the same time, the conditional GAN algorithm is used to process the preprocessed time-series data and multi-modal data to generate and verify synthetic data that conforms to the physical constraints. The time-series autocorrelation features and cross-device association features of the generated data are extracted, and the features that conform to the physical constraint discretization mathematical model are screened and used as the third feature source of the physical constraint features to enrich the feature set of scarce data.
[0097] Considering the effectiveness of the generated data, in an embodiment of the present invention, using the conditional GAN algorithm to process the preprocessed time-series data and multi-modal data to generate and verify synthetic data that conforms to the physical constraints may include the following steps: designing a constraint condition function, using the conditional GAN algorithm to generate synthetic data according to the physical constraint labels, using the constraint condition function to perform constraint verification on the synthetic data, and retaining the synthetic data that passes the verification; the synthetic data generated by the GAN algorithm according to the physical constraint labels can be effectively used after passing the verification of the constraint condition function to fill the scarce data and facilitate subsequent extraction and analysis of the physical constraint features.
[0098] On the other hand, the present invention provides a feature enhancement system for power multi-service samples. The feature enhancement system includes a data processing module, a feature enhancement module, and a service adaptation interface. Specifically, the data processing module can be used to preprocess multi-source heterogeneous data. The feature enhancement module is electrically connected to the data processing module and can be used to execute the feature enhancement method described in any one of the above claims to obtain enhanced features. The service adaptation interface can be electrically connected to the feature enhancement module and is used to input the enhanced features into the downstream service sample processing model. The three modules are electrically connected in sequence, which can realize the enhancement of data feature extraction and input the obtained enhanced features into the downstream service sample processing model to facilitate subsequent processing.
[0099] Considering the analysis and processing of samples and realizing the effective use of enhanced features, in one embodiment of the present invention, the service sample processing model includes at least one of a classifier, a regression model, or an anomaly detection model, so that the utilization efficiency of data can be improved through enhanced features to meet the subsequent processing of complex power service scenarios.
[0100] Through the above technical solution, the method obtains real-time multi-source heterogeneous data, preprocesses the multi-source heterogeneous data to eliminate outliers, improves the quality and stability of the data, and performs hierarchical feature extraction using multiple feature extraction algorithms according to the data components of the multi-source heterogeneous data to obtain feature sources. For time series data, the LSTM and Transformer algorithms are used to extract the first feature source of time series autocorrelation features. For modal data, the GNN algorithm is used to extract the second feature source of cross-device association features. For time series data and multi-modal data, the conditional GAN algorithm is used to extract the third feature source of physical constraint features, generating new features that are similar to but also have certain differences from the original sample features. The new features are merged with the original features to further enrich the feature set, making the sample features more accurate and diverse. The gated attention unit is used to perform weighted fusion on the feature sources to obtain enhanced features, realizing the integration of features, significantly enhancing the feature expression ability in complex service scenarios, and inputting the enhanced features into the service sample processing model to complete service processing, meeting the requirements of service scenarios such as the stable operation and fault prediction of the power system.
[0101] It should also be noted that the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, commodity or device including a series of elements not only includes those elements but also other elements not expressly listed, or elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, method, commodity or device including the element.
[0102] The above are only embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.
Claims
1. A feature enhancement method for power multi-service samples, characterized in that: The feature enhancement method comprises: Acquire real-time multi-source heterogeneous data, and pre-process the multi-source heterogeneous data; Performing hierarchical feature extraction according to data components of the multi-source heterogeneous data to obtain feature sources; Using a gated attention unit to perform weighted fusion on the feature sources to obtain enhanced features; The enhanced features are input into the business sample processing model to complete the business processing.
2. The feature enhancement method according to claim 1, characterized in that: The acquiring of real-time multi-source heterogeneous data and preprocessing of the multi-source heterogeneous data include: The multi-source heterogeneous data includes time series data and multimodal data; Cleaning and removing the multi-source heterogeneous data; The processed multi-source heterogeneous data are standardized and normalized.
3. The feature enhancement method according to claim 2, characterized in that: Performing hierarchical feature extraction according to the data components of the multi-source heterogeneous data to obtain feature sources includes: Extracting a first feature source of time series autocorrelation features from the time series data using LSTM and Transformer algorithms; Extracting a second feature source of cross-device correlation features from the multimodal data using a GNN algorithm; A conditional GAN algorithm is used to extract the third feature source of the physical constraint feature from the time series data and the multimodal data.
4. The feature enhancement method according to claim 3, characterized in that: The gated attention unit is used to perform weighted fusion on the feature sources to obtain enhanced features, including: Mapping the multiple feature sources to the same dimensional space for feature alignment, and performing position encoding to retain the time sequence information of the temporal autocorrelation feature; The gated attention unit performs independent linear projection on the feature source through formula group (1) to obtain the generated features. Among them, Z q is the generated feature of the first feature source, W q is the learnable weight matrix of the first feature source, F q is the input feature of the first feature source, b q is the bias term of the first characteristic source, Z m is the generated feature of the second feature source, W m is the learnable weight matrix of the second feature source, F m is the input feature of the second feature source, b m is the bias term of the second characteristic source, Z p is the generated feature of the third feature source, W p is the learnable weight matrix of the third feature source, F p is the input feature of the third feature source, b p is the bias term of the third characteristic source; Processing the generated features corresponding to the multiple feature sources according to formula group (2) to obtain the attention scores of the multiple feature sources, and normalizing the attention scores to obtain the weighted weights of the multiple feature sources, Among them, α q is the weight of the first feature source, α m is the weight of the second feature source, α p is the weighted weight of the third characteristic source, W a is the weight matrix, b a is the bias term; According to the generated weighted weight, the feature source is weightedly summed using formula (3) to obtain the enhanced feature. F fusion =a q *F q +a m *F m +a p *F p , (3) Among them, F fusion is the enhancement feature, α q is the weight of the first feature source, α m is the weight of the second feature source, α p is the weighted weight of the third characteristic source, F q is the input feature of the first feature source, F m is the input feature of the second feature source, F p is the input feature of the third feature source.
5. The feature enhancement method according to claim 3, characterized in that: The first feature source of extracting the time series autocorrelation feature by using LSTM and Transformer algorithms on the time series data includes: The time series data is processed using an LSTM algorithm to obtain a first output, h t =LSTM(x t ,h t-1 ), Among them, h t is the first output of the time series data at time t, x t is the input of the time series data at time t, h t-1 The output of the time series data at time t-1; The first output is processed by using a Transformer algorithm to obtain a first feature source of the fused time series autocorrelation feature, s t =∑(α t *h t ) Among them, α t is the weight index of the attention mechanism layer at time t, w1 is the weight matrix of the attention mechanism layer, b1 is the bias term of the attention mechanism layer, and s t It is the first characteristic source of the time series autocorrelation feature at time t.
6. The feature enhancement method according to claim 3, characterized in that: The second feature source of cross-device correlation features is extracted from the multimodal data using a GNN algorithm, including: Abstracting the elements of the data of the device as graph nodes according to the multimodal data, locating the association relationship between the devices as edges, and independently extracting the multimodal data to obtain the features of the initialized graph nodes; Concatenate the features of the graph nodes corresponding to the multimodal data according to the channel dimension to obtain a joint feature matrix; A graph convolution layer is used to aggregate node features of the joint feature matrix, and the features are fused with the initialized features of the graph nodes to obtain fused features. The fused features are projected into the subspace between the devices and extracted to obtain a second feature source of cross-device correlation features in the features of the graph nodes.
7. The feature enhancement method according to claim 3, characterized in that: The conditional GAN algorithm is used to extract the third feature source of the physical constraint feature from the time series data and the multimodal data, including: Establish a physical constraint discretization mathematical model based on the equipment type; Preprocessing the time series data and multimodal data; The conditional GAN algorithm is used to process the preprocessed time series data and multimodal data to generate and verify synthetic data that meets physical constraints; The temporal autocorrelation features and cross-device association features are extracted from the synthetic data, and the features that conform to the physical constraint discretization mathematical model are screened and used as the third feature source of the physical constraint features.
8. The feature enhancement method according to claim 7, characterized in that: The conditional GAN algorithm is used to process the preprocessed time series data and multimodal data to generate and verify synthetic data that meets physical constraints, including: Design constraint function; Using the conditional GAN algorithm to generate synthetic data based on physical constraint labels; Using the constraint condition function to perform constraint checking on the synthetic data; The synthetic data that pass the inspection are retained.
9. A feature enhancement system for power multi-service samples, characterized in that: The feature enhancement system comprises: Data processing module, used to pre-process multi-source heterogeneous data; A feature enhancement module, electrically connected to the data processing module, and configured to execute the feature enhancement method according to any one of claims 1 to 8 to obtain enhanced features; The service adaptation interface is electrically connected to the feature enhancement module and is used to input the enhanced features into a downstream service sample processing model.
10. The feature enhancement system according to claim 9, characterized in that: The business sample processing model includes at least one of a classifier, a regression model or an anomaly detection model.