A multi-dimensional electric power perception information feature fusion method

Abstract

A multi-dimensional electric power perception information feature fusion method, which first performs eigenvalue decomposition on the covariance matrix of data samples from different data sources based on principal component analysis (PCA), selects the first k main features with cumulative contribution rate exceeding a threshold value as principal components, and takes the feature vectors corresponding to the k main features as the column vectors of a second-order statistical feature vector matrix, then determines a high-order statistical feature vector matrix through the remaining m - k eigenvalues, performs feature pre-fusion of statistical data of different orders based on the statistical feature vector matrix, and determines a feature data matrix, and finally inputs the feature data matrix into a space-time characteristic fusion convolutional neural network (CNN) model for feature fusion. The present application adopts a unique PCA-ICA joint processing method for feature extraction, which can effectively save storage space and fusion calculation time.

Description

Technical Field

[0001] This invention belongs to the field of power information fusion technology, specifically relating to a method for fusing multi-dimensional power sensing information features. Background Technology

[0002] With the development of IoT and sensor network technologies, the entire power operation process, including power generation, transmission, distribution, and consumption, generates massive amounts of multi-dimensional indicator data. This data poses significant challenges to storage space and computation time. Furthermore, the heterogeneity of data from different devices often makes direct fusion difficult. Therefore, feature extraction and fusion of different types of data from various data sources are of great significance for saving storage space and computation time, improving the accuracy of data fusion and the relevance of perceptual information across different dimensions, and driving the transformation of data from resource utilization to value creation.

[0003] Most existing data fusion methods rely on the homogeneity of data. Data from different types of devices are usually difficult to fuse directly. In addition, the sensing devices of the power Internet of Things generate high-dimensional data, which contains a lot of data noise and redundancy. Existing solutions for low-dimensional data cannot be effectively applied to the processing of high-dimensional data. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems in the existing technology by providing a multi-dimensional power sensing information feature fusion method that can effectively extract and fuse feature data of multi-dimensional power sensing information.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A multi-dimensional power sensing information feature fusion method includes the following steps:

[0007] Step A: Based on Principal Component Analysis (PCA), perform eigenvalue decomposition on the covariance matrix of data samples from different data sources, select the j principal features whose cumulative contribution rate exceeds a threshold as principal components, and use the eigenvectors corresponding to these j principal features as the second-order statistical eigenvector matrix Q. PCA The column vectors;

[0008] Step B: Determine the higher-order statistical eigenvector matrix Q using the remaining mk eigenvalues. ICA ;

[0009] Step C: Perform feature pre-fusion on statistical data of different orders according to the following formula:

[0010] Q = [w ICA Q ICA ;w PCA QPCA ]

[0011] In the above formula, Q is the pre-fused feature vector matrix, and w PCA w ICA Q PCA Q ICA The corresponding weights;

[0012] Step D: Determine the feature data matrix T based on the pre-fused feature vector matrix Q:

[0013] T = [c1, c2, ..., c k d1, d2, ..., d m-k ]

[0014]

[0015]

[0016] In the above formula, c i Let d be the data vector of the i-th second-order statistical characteristic. j Let α be the j-th higher-order feature data vector. i α j+k Let y be the vectors in the i-th and (j+j)-th columns of matrix Q. i Let n be the column vector of data samples from the i-th data source, and n be the total number of data sources.

[0017] Step E: Input matrix T into the spatiotemporal characteristic fusion convolutional neural network (CNN) model for feature fusion to obtain fused data.

[0018] In step B, Q ICA The following formula is used to calculate:

[0019]

[0020]

[0021] In the above formula, Z is the ICA preprocessing matrix. The projection matrix is ​​∑, and E is the diagonal matrix and eigenvector matrix corresponding to the remaining mk eigenvalues, respectively.

[0022] In step A, the cumulative contribution rate G is calculated using the following formula:

[0023]

[0024] In the above formula, λ represents the cumulative contribution value of the first j principal features. i , λ j These are the contribution values ​​of the i-th and j-th features, respectively;

[0025] In step C, the w PCA w ICA It is calculated using the following formula:

[0026]

[0027]

[0028] Step E includes the following steps in sequence:

[0029] Step E1: Divide matrix T into multiple data streams according to features, and perform feature alignment and noise filtering through the multi-feature linking module set in the CNN model to obtain the processed feature data;

[0030] Step E2: The global feature fusion module in the CNN model performs global fusion of the processed feature data and the data in matrix T to obtain the feature fusion data x. (i) The whole-disk feature fusion module includes multiple deconvolutional layers with a stride of 2.

[0031] Step E3: Fuse the feature data x (i) Fusion data with reality y( i) To make a comparison, if If x is less than or equal to the set threshold ε1, then x will be... (i) Otherwise, after updating the weights of each convolutional layer according to the following formula, return to step E1:

[0032]

[0033]

[0034] In the above formula, w i w′ i ... Let ε be the loss function, ε be the trainable parameters, and s be the number of data types in the fused output.

[0035] In step E1, the multi-path feature linking module set includes an upsampling module and multiple cascaded multi-path feature linking modules. The signal input terminal of the upsampling module is connected to multiple data streams, and the signal output terminal of the upsampling module is connected to the signal output terminal of the full-disk feature fusion module through each multi-path feature linking module in sequence.

[0036] The upsampling module uses a recombination operation with a magnification of 4 to filter out some noise in the feature data of the data stream;

[0037] Multiple multi-path feature linking modules are cascaded to gradually establish alignment relationships between data streams, thereby achieving feature alignment.

[0038] Each of the multi-path feature linking modules includes a residual module and multiple attention modules, each connected to the signal output terminal of the residual module. In the first-level multi-path feature linking module, the signal input terminal of the residual module is connected to the signal output terminal of the upsampling module. The signal output terminals of the multiple attention modules in the first-level multi-path feature linking module are connected to the signal input terminals of the residual modules in the second-level and third-level multi-path feature linking modules. The signal output terminals of the multiple attention modules in the second-level multi-path feature linking module are connected to the signal input terminals of the residual modules in the third-level multi-path feature linking module. The signal output terminals of the multiple attention modules in the third-level multi-path feature linking module are connected to the signal input terminals of the residual modules in the fourth-level and fifth-level multi-path feature linking modules. The signal output terminals of the multiple attention modules in the fourth-level multi-path feature linking module are connected to the signal input terminals of the residual modules in the fifth-level multi-path feature linking module, and so on.

[0039] The attention module is used to extract effective masks for cross features, achieve feature alignment, and suppress noise interference.

[0040] The residual module includes a first average pooling layer, multiple cascaded convolutional layers, a second average pooling layer, and a fully connected layer. In the first-level multi-path feature linking module, the signal input of the first average pooling layer is connected to the signal output of the upsampling module. In other levels of multi-path feature linking modules, the signal input of the first average pooling layer is connected to the signal output of multiple attention modules in the previous level of multi-path feature linking module. The signal output of the first average pooling layer is connected to the signal input of the second average pooling layer through each convolutional layer. The signal output of the second average pooling layer is connected to the corresponding attention module through a fully connected layer. The first and second average pooling layers use a 2×2 pooling window and the average value of the elements within the window is used as the output.

[0041] The convolutional layer uses the following activation function:

[0042]

[0043] In the above formula, is the activation function, and x is the input parameter of the residual module;

[0044] The output of the convolutional layer 212 is:

[0045]

[0046]

[0047]

[0048]

[0049]

[0050]

[0051] In the above formula, f1(x) and f5(x) are the outputs of the first and fifth convolutional layers, respectively; w1 and w5 are the weights of the first and fifth convolutional layers, respectively; and b1 and b5 are the biases of the first and fifth convolutional layers, respectively.

[0052] In step A, before performing eigenvalue decomposition on the covariance matrix of data samples from different data sources, the data samples from different data sources are first standardized as follows to obtain the column vectors of data samples from each data source:

[0053] y i =[y i1 y i2 , ..., y ij , ...y in ] T

[0054]

[0055] In the above formula, y i Let y be the column vector of data samples from the i-th data source. ij For the j-th sample x from the i-th data source ij The results after standardization S i These are the sample mean and sample variance of the data source, respectively.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] 1. The present invention provides a multi-dimensional power sensing information feature fusion method. First, based on Principal Component Analysis (PCA), the covariance matrix of data samples from different data sources is decomposed into eigenvalues. The top j principal features whose cumulative contribution rate exceeds a threshold are selected as principal components, and the eigenvectors corresponding to these j principal features are used as the second-order statistical feature vector matrix Q. PCA The column vectors are then used to determine the higher-order statistical eigenvector matrix Q using the remaining mk eigenvalues. ICA Then based on Q PCA and Q ICAThis method involves pre-fusion of features from different orders of statistical data and determining the feature data matrix T. Finally, matrix T is input into a spatiotemporal characteristic fusion convolutional neural network (CNN) model for feature fusion to obtain fused data. This method employs a unique PCA-ICA joint processing approach for feature extraction. PCA processing yields second-order statistical properties, while ICA processing yields higher-order statistical properties. The combination of the two can effectively preserve the different orders of data properties, thereby retaining as much useful information as possible from the original perceptual data and reducing information loss during data processing. Based on this, it enables the extraction and fusion of different types of data features, effectively saving storage space and computation time.

[0058] 2. This invention provides a multi-dimensional power sensing information feature fusion method that introduces an upsampling module, a residual module, and an attention module into a spatiotemporal characteristic fusion convolutional neural network. On the one hand, the upsampling module can effectively reduce the amount of data that needs to be processed. On the other hand, to avoid the training difficulties caused by a large number of network layers, a residual neural network module is added after the upsampling module. The output of the shallow network is added to the input of the deep network and cascaded, achieving a balance between nonlinear and linear transformations. This results in a clear hierarchy in the neural network. The expressive power of the output features is ensured by progressively increasing the number of features passed with each convolutional layer. At the same time, regularization is performed using average pooling, which accelerates the training speed. In addition, the attention module adjusts the attention of each position of the passed features and the resources on each convolutional channel, allowing the model to focus on more specified regions that need attention. Attached Figure Description

[0059] Figure 1 This is the overall flowchart of the present invention.

[0060] Figure 2 for Figure 1 A block diagram of the overall composition of a convolutional neural network model that integrates spatiotemporal characteristics.

[0061] Figure 3 for Figure 2 Block diagram of the multi-path feature linking module.

[0062] Figure 4 for Figure 3 Block diagram of the residual module.

[0063] Figure 5 This is a schematic diagram of the cascaded convolutional layers in the residual module.

[0064] In the figure, there are upsampling module 1, multi-path feature linking module 2, residual module 21, first average pooling layer 211, convolutional layer 212, second average pooling layer 213, fully connected layer 214, and attention module 22. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0066] This invention provides a multi-dimensional power sensing information feature fusion method. The method first processes the data collected by sensing devices such as sensors in the power Internet of Things through PCA-ICA joint processing to obtain the second-order and higher-order statistical features of the data. Then, the weights of the obtained second-order and higher-order statistical data are adjusted to achieve feature pre-fusion of statistical data of different orders. Finally, the pre-fused data is feature fused by combining spatiotemporal characteristics with a convolutional neural network model.

[0067] The spatiotemporal characteristic fusion convolutional neural network model used in this invention first reduces the amount of data to be processed through an upsampling module, and then establishes effective alignment of the features obtained by the multi-path attention module through a cascaded multi-path feature linking module. At the same time, the cascaded method effectively avoids the problem of shallow data loss during feature alignment. Finally, the feature data that has been extracted and aligned through the whole feature fusion module is fused to reconstruct the feature fusion data.

[0068] Full-disk feature fusion module: In this invention, the data input to the full-disk feature fusion module includes not only the feature data obtained by the multi-path feature link module set, but also the feature data after PCA-ICA joint processing. This is because although the multi-path feature link module can achieve feature alignment and suppress noise interference, some data features may be lost due to the effects of operations such as pooling operation taking window averaging, attention module focusing on acquiring main features and losing secondary features, and downsampling in the residual module.

[0069] Loss function: The loss function used in this invention is cross-entropy error, which can truly reflect the error between the actual result and the predicted result. At the same time, because it is a convex function, there exists a global optimal solution.

[0070] Example 1:

[0071] In this embodiment, the data source consists of data from M substations and N power-consuming sides. The substation data includes: transmission line temperature, oil-immersed transformer temperature (°C), transformer node output load (W), voltage level (kV), capacity (MVA), maximum useful power transmitted (kW), peak useful power time, high-voltage side wiring method, medium-voltage side wiring method, low-voltage side wiring method, and three-phase waveforms. The power-consuming side data includes: power consumption (W), number of faults and time periods, voltage (V), current (A), minimum power (W), and temperature, humidity, and wind speed information collected by sensors.

[0072] See Figure 1A multi-dimensional power sensing information feature fusion method is performed in the following steps:

[0073] 1. The data samples from different data sources are standardized as follows to obtain a standardized data matrix Y composed of column vectors of data samples from each data source:

[0074] y i =[y i1 y i2 , ..., y ij , ...y in ] T

[0075]

[0076] In the above formula, y i Let y be the column vector of data samples from the i-th data source. ij For the j-th sample x from the i-th data source ij The results after standardization S i These are the sample mean and sample variance of the data source, respectively.

[0077] 2. Based on principal component analysis (PCA), eigenvalue decomposition is performed on the covariance matrix of the standardized data matrix Y. The top j principal features with a cumulative contribution rate exceeding 85% are selected as principal components. The cumulative contribution rate G is calculated using the following formula:

[0078]

[0079] In the above formula, λ represents the cumulative contribution value of the first j principal features. i , λ j These are the contribution values ​​of the i-th and j-th features, respectively.

[0080] 3. Construct a second-order statistical eigenvector matrix Q using the eigenvectors corresponding to the first j principal features as column vectors. PCA .

[0081] 4. For the remaining mk eigenvalues, the higher-order statistical eigenvector matrix Q is determined using the following formula. ICA :

[0082]

[0083]

[0084] In the above formula, Z is the ICA preprocessing matrix. The projection matrix, derived from the matrix Obtained using the InfoMax method, ∑ and E are the diagonal matrix and eigenvector matrix corresponding to the remaining mk eigenvalues, respectively. i Let be the column vector of data samples from the i-th data source.

[0085] 5. Perform feature pre-fusion on statistical data of different orders according to the following formula:

[0086] Q = [w ICA Q ICA ;w PCA Q PCA ]

[0087]

[0088]

[0089] In the above formula, Q is the pre-fused feature vector matrix, and w PCA w ICA Q PCA Q ICA The corresponding weights, defined by the contribution values ​​of the features, characterize the importance of second-order and higher-order data, respectively. λ represents the cumulative contribution of the first k principal features. i , λ j These are the contribution values ​​of the i-th and j-th features, respectively.

[0090] 6. Determine the feature data matrix T using the following formula:

[0091] T = [c1, c2, ..., c k d1, d2, ..., d m-k ]

[0092]

[0093]

[0094] In the above formula, c i Let d be the data vector of the i-th second-order statistical characteristic. j Let α be the j-th higher-order feature data vector. i α j+k Let y be the vectors in the i-th and (j+j)-th columns of matrix Q. i Let n be the column vector of data samples from the i-th data source, and n be the total number of data sources.

[0095] 7. See also Figures 2-5The matrix T is divided into multiple data streams according to its features. These streams are then aligned and noise filtered using a set of multi-path feature linking modules in a CNN model to obtain processed feature data. The multi-path feature linking module set includes an upsampling module 1 and multiple cascaded multi-path feature linking modules 2. The upsampling module 1 uses a recombination operation with a multiplier of 4 to filter out some noise in the feature data streams. The multiple multi-path feature linking modules are cascaded to gradually establish alignment relationships between the data streams, achieving feature alignment. Each multi-path feature linking module 2 includes a residual module 21 and multiple attention modules 22 connected to the signal output of the residual module 21. The attention modules 22 are used for effective cross-feature extraction. The mask is used to achieve feature alignment and suppress noise interference. The residual module 21 includes a first average pooling layer 211, multiple cascaded convolutional layers 212, a second average pooling layer 213, and a fully connected layer 214. In the first-level multi-path feature linking module 2, the signal input of the first average pooling layer 211 is connected to multiple data streams through the upsampling module 1. The signal output of the first average pooling layer 211 is sequentially connected to the signal input of the second average pooling layer 213 through each convolutional layer 212. The signal output of the second average pooling layer 213 is connected to the signal input of multiple attention modules 22 through the fully connected layer 214. The signal output of the multiple attention modules 22 is connected to the signal input of the second-level multi-path feature linking module 2. The signal input terminal of the first average pooling layer 211 is connected to the signal input terminal of the first average pooling layer 211 in the third-level multi-path feature linking module 2. In the second-level multi-path feature linking module 2, the signal output terminal of the first average pooling layer 211 is connected to the signal input terminal of the second average pooling layer 213 through each convolutional layer 212. The signal output terminal of the second average pooling layer 213 is connected to the signal input terminals of multiple attention modules 22 through a fully connected layer 214. The signal output terminals of the multiple attention modules 22 are connected to the signal input terminal of the first average pooling layer 211 in the third-level multi-path feature linking module 2. In the third-level multi-path feature linking module 2, the signal output terminals of the multiple attention modules 22 are connected to the signal input terminals of the fourth-level multi-path feature linking module 211. The signal input terminals of the first average pooling layer 211 in the multi-path feature linking module 2 of the fifth level are connected to the signal input terminals of the first average pooling layer 211 in the multi-path feature linking module 2 of the fourth level. The signal output terminals of multiple attention modules 22 in the multi-path feature linking module 2 of the fourth level are connected to the signal input terminals of the first average pooling layer 211 in the multi-path feature linking module 2 of the fifth level, and so on. The signal output terminals of multiple attention modules 22 in the last level multi-path feature linking module 2 are connected to the signal output terminals of the full-disk feature fusion module. The first average pooling layer 211 and the second average pooling layer 213 select a 2×2 pooling window, and use the average value of the elements within the window as the output. The convolutional layer 212 uses the following activation function:

[0096]

[0097] In the above formula, is the activation function, and x is the input parameter of the residual module;

[0098] The output of the convolutional layer 212 is:

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105] In the above formula, f1(x) and f5(x) are the outputs of the first and fifth convolutional layers, respectively; w1 and w5 are the weights of the first and fifth convolutional layers, respectively; and b1 and b5 are the biases of the first and fifth convolutional layers, respectively.

[0106] 8. The global feature fusion module in the CNN model performs global fusion of the processed feature data and the data in matrix T to obtain the feature fusion data x. (i) The whole-disk feature fusion module includes multiple deconvolutional layers with a stride of 2.

[0107] 9. Fuse features into data x (i) Integrated with real data y (i) To make a comparison, if If x is less than or equal to the set threshold ε1, then x will be... (i) Otherwise, after updating the weights of each convolutional layer according to the following formula, return to step E1:

[0108]

[0109]

[0110] In the above formula, w i , ... Let ε be the loss function, ε be the trainable parameters, and s be the number of data types in the fused output.

[0111] The data types output in this embodiment include: total power consumption (kW), load level (a constant between 0 and 1, with a larger value indicating a larger power load), system capacity (MVA), operating safety level (a constant between 0 and 1, with a larger value indicating a higher level of safety), and failure rate.

Claims

1. A multi-dimensional power-aware information feature fusion method, characterized in that: The method comprises the following steps in sequence: Step A, eigenvalue decomposition of the covariance matrix of the data samples from different data sources based on principal component analysis PCA, select cumulative contribution rate The first k main features exceeding the threshold value are taken as principal components, and the eigenvectors corresponding to the k main features are taken as the second-order statistical feature vector matrix Each column vector of the second-order statistical feature vector matrix, wherein the data samples are data collected by sensing devices of the power internet of things, including transmission line temperature, oil-immersed transformer temperature, transformer node output load, voltage level, capacity, maximum useful power transmission, useful power peak time, high-voltage side wiring mode, medium-voltage side wiring mode, low-voltage side wiring mode, three-phase waveform, power consumption, fault frequency and time period, voltage, current, minimum power, temperature, humidity and wind speed information collected by sensors; Step B. Determine the higher order statistical feature vector matrix by the remaining m-k eigenvalues ; Step C: Feature pre-fusion is performed on the statistical data of different orders according to the following formula: ； In the above formula, is the feature vector matrix after pre-fusion, , are respectively , corresponding weights; Step D, based on the pre-fusion feature vector matrix Determining a feature data matrix : ； ； ； In the above formula, is the i-th second order statistical characteristic data vector, is the j-th higher order characteristic data vector, is the number of second order and higher order statistical characteristics, , is the i-th column vector of the matrix is the j+k-th column vector of the matrix is the data sample column vector of the i-th data source, is the total number of data sources; Step E, matrixing The spatio-temporal characteristics fusion convolutional neural network CNN model is used for feature fusion, and fusion data is obtained, including: Step E1, matrix According to the characteristics, the multi-path data stream is divided, the feature alignment and noise filtering are performed through a multi-path feature linking module set in the CNN model, and processed feature data is obtained, wherein the multi-path feature linking module set comprises an up-sampling module (1) and a plurality of cascaded multi-path feature linking modules (2), a signal input end of the up-sampling module (1) is in communication with the multi-path data stream, and signal output ends of the up-sampling module (1) are connected with a signal output end of a full-plate feature fusion module through each multi-path feature linking module (2) in sequence. The up-sampling module (1) uses a recombination operation with a magnification of 4 to filter out part of the noise of the feature data in the data stream; A plurality of the multi-path feature linking modules gradually establish the alignment relationship of each data stream in a cascading manner to realize feature alignment; Each multi-path feature linking module (2) comprises a residual module (21) and a plurality of attention modules (22) connected to the signal output end of the residual module (21), the signal input end of the residual module (21) in the first-level multi-path feature linking module (2) is connected to the signal output end of the up-sampling module (1), the signal output ends of the plurality of attention modules (22) in the first-level multi-path feature linking module (2) are connected to the signal input ends of the residual modules (21) in the second-level and third-level multi-path feature linking modules (2), the signal output ends of the plurality of attention modules (22) in the second-level multi-path feature linking module (2) are connected to the signal input end of the residual module (21) in the third-level multi-path feature linking module (2), the signal output ends of the plurality of attention modules (22) in the third-level multi-path feature linking module (2) are connected to the signal input ends of the residual modules (21) in the fourth-level and fifth-level multi-path feature linking modules (2), the signal output ends of the plurality of attention modules (22) in the fourth-level multi-path feature linking module (2) are connected to the signal input end of the residual module (21) in the fifth-level multi-path feature linking module (2), and so on. The attention module (22) is used to extract effective masks for cross features, realize feature alignment, and suppress noise interference. Step E2, the global feature fusion module in the CNN model globally fuses the processed feature data and the data in the matrix to obtain feature fusion data , wherein the global feature fusion module comprises a plurality of deconvolution layers with a step size of 2. Step E3, fusing data of features with real fused data Compare, if is less than or equal to a set threshold then output as fused data, otherwise update the weights of each convolutional layer according to the following formula and return to step E1: ； ； In the above formula, , are the weights of the i-th convolutional layer before and after updating, respectively, is the learning rate, is the loss function, is the trainable parameter, is the number of data categories of the fusion output. 2.The multi-dimensional power-aware information feature fusion method according to claim 1, characterized in that: In step B, the The result was calculated according to the following formula: ； ； In the above formula, is a projection matrix of the ICA pre-processing matrix , , are diagonal matrices and eigenvector matrices corresponding to the remaining m-k eigenvalues, respectively. 3.The multi-dimensional power-aware information feature fusion method according to claim 1 or 2, characterized in that: In step A, the cumulative contribution rate is calculated from the following equation: ； In the above formula, is the cumulative contribution value of the first k main features, , is the contribution value of the i, jth feature, respectively. In Step C, the , is calculated from the following equation: ； 。 4.The multi-dimensional power-aware information feature fusion method according to claim 1, characterized in that: The residual module (21) comprises a first average pooling layer (211), a plurality of convolution layers (212) in cascade, a second average pooling layer (213), and a full connection layer (214). The signal input end of the first average pooling layer (211) in the first multi-path feature linking module (2) is connected with the signal output end of the up-sampling module (1), the signal input end of the first average pooling layer (211) in the multi-path feature linking module (2) of other levels is connected with the signal output end of the plurality of attention modules (22) in the multi-path feature linking module (2) of the previous level, the signal output end of the first average pooling layer (211) is connected with the signal input end of the second average pooling layer (213) through the convolution layers (212) in sequence, the signal output end of the second average pooling layer (213) is connected with the corresponding attention module (22) through the full connection layer (214), and the first average pooling layer (211) and the second average pooling layer (213) select a 2x2 pooling window and take the average value of the elements in the window as the output.

5. The multi-dimensional power perception information feature fusion method according to claim 4, wherein: The convolution layer (212) adopts the following activation function: ； In the above formula, is an activation function, is an input parameter of the residual module; The output of the convolution layer (212) is: ； In the above formula, , are outputs of the 1st and 5th convolution layers, respectively, , are weights of the 1st and 5th convolution layers, respectively, , are biases of the 1st and 5th convolution layers, respectively.

6. The multi-dimensional power perception information feature fusion method according to claim 1 or 2, wherein: In step A, before performing eigenvalue decomposition on the covariance matrix of the data samples from different data sources, the data samples from different data sources are subjected to the following standardization processing to obtain the data sample column vectors of each data source: ； ； In the above formula, is the data sample column vector of the i-th data source, is the j-th sample of the i-th data source is the result after standardization, 、 is the sample mean and sample variance of the data source, respectively, is the sample quantity of each data source.

Images