Power system signal classification method based on hybrid expert system and visual processing

Abstract

The invention discloses an electric power system signal classification method based on a hybrid expert system and visual processing, and the method comprises the steps: converting a to-be-processed electric power signal of an electric power system into an electric power signal image; processing the electric power signal image by using the plurality of image coding models to obtain a plurality of image feature vectors of the electric power signal image, the model parameters of the plurality of image coding models being different from each other, and each image feature vector being obtained by processing the corresponding image coding model; and determining a classification result of the to-be-processed power signal according to the plurality of image feature vectors. According to the scheme, a plurality of different image coding models are combined to process the electric power signal image, and the classification result is determined based on a plurality of image feature vectors output by the plurality of image coding models, so that the characteristics of different aspects of the electric power signal to be processed can be mined by using the plurality of image coding models; therefore, a more accurate classification result can be obtained.

Description

Technical Field

[0001] The present application relates to the technical field of signal classification, and in particular to a method for classifying power system signals based on a hybrid expert system and visual processing. Background Art

[0002] The signal data generated during the operation of the power system contains a wealth of information about the system's operating status. Studies have shown that converting these signal data into images and then using computer vision processing methods to analyze them is an effective method. However, traditional visual processing methods often have problems such as high computational complexity and limited generalization ability when processing power signals. Summary of the invention

[0003] To this end, this application discloses the following technical solutions:

[0004] The first aspect of the present application provides a method for classifying power system signals based on a hybrid expert system and visual processing, comprising:

[0005] Obtaining a power signal to be processed from a power system;

[0006] Converting the power signal to be processed into a power signal image;

[0007] Processing the power signal image using multiple image coding models to obtain multiple image feature vectors of the power signal image, wherein the model parameters of the multiple image coding models are different from each other, and each of the image feature vectors is obtained by processing the corresponding image coding model;

[0008] A classification result of the power signal to be processed is determined according to the plurality of image feature vectors.

[0009] Optionally, converting the to-be-processed power signal into a power signal image includes:

[0010] The power signal to be processed is processed based on the Gram angular field method to obtain a power signal image.

[0011] Optionally, each of the image coding models includes a central differential convolution module, a horizontal differential convolution module, a vertical differential convolution module, an adaptive differential convolution module and a standard convolution module;

[0012] The process of processing the power signal image using any of the image coding models to obtain a corresponding image feature vector includes:

[0013] Using the central difference convolution module, the horizontal difference convolution module, the vertical difference convolution module, the adaptive difference convolution module and the standard convolution module of the image coding model to process the power signal image respectively, and obtain multiple convolution features of the image coding model;

[0014] Multiple convolutional features of the image coding model are fused to obtain an image feature vector corresponding to the image coding model.

[0015] Optionally, determining the classification result of the power signal to be processed according to the plurality of image feature vectors includes:

[0016] fusing the plurality of image feature vectors according to weight values ​​of the plurality of image coding models to obtain a first fused vector of the power signal image, wherein the weight values ​​of the plurality of image coding models are determined according to the power signal to be processed;

[0017] A classification result of the power signal to be processed is determined according to the first fusion vector.

[0018] Optionally, the method for determining weight values ​​of the plurality of image coding models according to the power signal to be processed includes:

[0019] Extracting a time series noise signal of the power signal to be processed;

[0020] A dynamic weight vector of the power signal to be processed is determined according to the timing noise signal, wherein the dynamic weight vector includes components corresponding to a plurality of the image coding models one by one, and each of the components is a weight value of the corresponding image coding model.

[0021] Optionally, determining the dynamic weight vector of the power signal to be processed according to the timing noise signal includes:

[0022] Processing the time series noise signal according to the long short-term memory network to obtain noise characteristics;

[0023] Normalizing the noise feature to obtain a normalized noise feature;

[0024] The normalized noise feature is processed based on a pre-built weight calculator to obtain a dynamic weight vector of the power signal to be processed.

[0025] Optionally, the method for obtaining a plurality of the image coding models includes:

[0026] Randomly generate multiple different random seeds;

[0027] For each of the random seeds, initialization is performed according to the random seed to obtain an initialization model corresponding to the random seed, and the initialization model is trained using sample data to obtain an image coding model corresponding to the random seed.

[0028] Optionally, the method for obtaining a plurality of the image coding models includes:

[0029] Get an initialization model and multiple different sample data sets;

[0030] For each of the sample data sets, the initialization model is trained using the sample data set to obtain an image coding model corresponding to the sample data set.

[0031] Optionally, determining the classification result of the power signal to be processed according to the plurality of image feature vectors includes:

[0032] fusing a plurality of the image feature vectors and the power signal to be processed to obtain a second fused vector;

[0033] A classification result of the power signal to be processed is determined according to the second fusion vector.

[0034] Optionally, before converting the power signal to be processed into a power signal image, the method further includes:

[0035] Performing signal preprocessing on the power signal to be processed to obtain a preprocessed power signal;

[0036] The converting the to-be-processed power signal into a power signal image comprises:

[0037] The preprocessed power signal is converted into a power signal image.

[0038] The beneficial effects of this solution are:

[0039] This scheme combines multiple different image coding models to process power signal images, and determines the classification results based on multiple image feature vectors output by the multiple image coding models. Therefore, this scheme can use multiple image coding models to mine the characteristics of different aspects of the power signal to be processed, thereby obtaining more accurate classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0041] Figure 1 is a flow chart of a method for classifying power system signals based on a hybrid expert system and visual processing provided in an embodiment of the present application;

[0042] Figure 2 is a schematic diagram of fusing multiple image feature vectors to determine a classification result provided by an embodiment of the present application;

[0043] Figure 3 is another schematic diagram of fusing multiple image feature vectors to determine a classification result provided by an embodiment of the present application;

[0044] Figure 4 It is a structural diagram of an image coding model provided in an embodiment of the present application;

[0045] Figure 5 It is a structural diagram of a power system signal classification system based on a hybrid expert system and visual processing provided in an embodiment of the present application. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0047] The present application also provides a method for classifying power system signals based on a hybrid expert system and visual processing. Figure 1 , the method may include the following steps.

[0048] S101, obtaining a power signal to be processed from a power system.

[0049] S102, converting the power signal to be processed into a power signal image.

[0050] S103, using multiple image coding models to process the power signal image to obtain multiple image feature vectors of the power signal image, the model parameters of the multiple image coding models are different from each other, and each image feature vector is obtained by processing the corresponding image coding model.

[0051] S104: Determine a classification result of the power signal to be processed according to the multiple image feature vectors.

[0052] The beneficial effects of this solution are:

[0053] This scheme combines multiple different image coding models to process power signal images, and determines the classification results based on multiple image feature vectors output by the multiple image coding models. Therefore, this scheme can use multiple image coding models to mine the characteristics of different aspects of the power signal to be processed, thereby obtaining more accurate classification results.

[0054] In step S101, a power signal generated by any power system within any period of time may be obtained as a power signal to be processed.

[0055] See also Figure 2 In step S102, the power signal to be processed may be converted into a power signal image as follows:

[0056] The power signal to be processed is processed based on the Gram angular field method to obtain a power signal image.

[0057] The Gram angular field method is an existing method for converting a time series signal into a corresponding image. The specific implementation principle thereof can be found in the relevant prior art and will not be described in detail.

[0058] In this embodiment, the existing Python signal processing library for implementing the Gram angular field method can be used to process the power signal to be processed to obtain the above-mentioned power signal image.

[0059] Converting the power signal to be processed into a power signal image through the Gram angle field method is beneficial to retaining the time-frequency characteristics of the signal.

[0060] In some optional embodiments, the following steps may also be performed before converting into a power signal image:

[0061] Performing signal preprocessing on the power signal to be processed to obtain a preprocessed power signal;

[0062] In the case of performing the above preprocessing steps, the power signal to be processed is converted into a power signal image, which can be replaced by:

[0063] The preprocessed power signal is converted into a power signal image.

[0064] Signal preprocessing methods include but are not limited to noise reduction, filtering, signal enhancement, etc. The specific processing methods can be found in the relevant prior art and will not be described in detail.

[0065] like Figure 2 As shown, in step S103, multiple pre-built image coding models can be used to process the power signal image, for example, Figure 2 The image coding models 1 to n process the power signal image respectively, and obtain the image feature vectors 1 to n output by the image coding models 1 to n. Each image coding model is equivalent to an expert module for processing the power signal image. Therefore, the present solution can realize multi-expert hybrid processing of the power signal image through the above-mentioned multiple image coding models.

[0066] The image coding model used in this embodiment may be a detail-enhanced lightweight vision encoder (Detail-Enhanced Lightweight Vision Transformer, DEFormer) or other existing encoders without limitation.

[0067] See also Figure 2In step S104, the classification result of the power signal to be processed may be determined according to the plurality of image feature vectors by:

[0068] fusing a plurality of image feature vectors and the power signal to be processed to obtain a second fused vector;

[0069] A classification result of the power signal to be processed is determined according to the second fusion vector.

[0070] In this embodiment, a pre-built neural network model for extracting signal features can be used to process the power signal to be processed to obtain a power signal feature vector, and then the power signal feature vector and multiple image feature vectors are added to obtain the result as the second fusion vector.

[0071] After the second fusion vector is obtained, the second fusion vector can be processed by a pre-built classification model to obtain a classification result of the power signal to be processed.

[0072] Depending on the application scenario, the classification results of the power signal to be processed may be different. For example, when it is necessary to detect whether the power system has a fault, the classification result may be a fault signal or a non-fault signal. When it is necessary to detect the working condition of the power system, the classification result may be different working conditions of the power system.

[0073] Among them, the neural network model and classification model used to extract signal features are not Figure 2 Shown in.

[0074] When the image encoding model is DEFormer, see Figure 4 , which is a structural diagram of the image coding model.

[0075] The image encoding model consists of an input layer, a local representation module, an efficient additive attention module, a multilayer perceptron (MLP), and an output layer.

[0076] In the input layer, the image coding model first uses a 7*7 convolution layer to perform preliminary feature extraction on the input image, and then processes the features extracted by the 7*7 convolution layer through a normalization layer and an activation function. The normalization layer can perform feature normalization, and the activation function can increase the nonlinear expression capability. The normalization layer of the input layer can be a batch normalization layer (BatchNorm), and the activation function can be a rectified linear unit (ReLU). The above input layer structure enables the image coding model to effectively capture the basic visual features of the input image.

[0077] The output of the input layer enters the local representation module, which may include a detail enhancement convolution module. The detail enhancement convolution module is composed of modules corresponding to five complementary convolution operations, including a center difference convolution module, a horizontal difference convolution module, a vertical difference convolution module, an adaptive difference convolution module and a standard convolution module.

[0078] Among them, the central difference convolution module can capture the relative relationship between pixels and their neighborhoods, the horizontal difference convolution module and the vertical difference convolution module can extract gradient information in the horizontal and vertical directions respectively, the adaptive difference convolution module dynamically adjusts the feature extraction strategy through learnable parameters, and the standard convolution module maintains the ability to extract basic features.

[0079] The outputs of the above five convolution modules will be feature fused to obtain the fused features, and the fused features will be processed by the normalization layer, convolution layer, activation function and convolution layer of the local representation module to obtain the output of the local representation module. The normalization layer can be a batch normalization layer (BatchNorm), the two convolution layers can be pointwise convolution layers (i.e., PointWise convolution layers), and the activation function used can be a Gaussian error linear unit (GELU). The structure of the above local representation module can ensure smooth propagation of gradients while maintaining nonlinearity.

[0080] Each convolution module introduced in the local representation module focuses on capturing image details in a specific direction and scale. The synergy of these convolution modules enables the image coding model to enhance the perception of detail features in multiple dimensions. At the same time, through the parameter sharing mechanism composed of normalization layer, convolution layer and activation function, the image coding model of this scheme improves performance while maintaining the lightweight characteristics of the model, so that the increase in computational overhead is controlled within an acceptable range.

[0081] The output of the local representation module will enter the efficient additive attention module. The attention module generates query features (also called query vectors, or Query vectors) and key features (also called key vectors, or Key vectors) based on the output of the local representation module, and then calculates the attention weights of the query features and key features. The calculation results are aggregated together with the fused features provided by the local representation module in the global information integration module as the output of the efficient additive attention module.

[0082] Compared with the traditional self-attention mechanism, the attention mechanism of the above efficient additive attention module significantly reduces the computational complexity while maintaining the ability to effectively model the global context.

[0083] The output of the efficient additive attention module will enter the multilayer perceptron for further processing. Specifically, the multilayer perceptron includes a series of normalization layers, convolution layers, activation functions and convolution layers, where the normalization layer can be BatchNorm, the activation function can be GELU, and both convolution layers are 1*1 convolution layers. In the multilayer perceptron, the output of the efficient additive attention module is processed in sequence by the series of normalization layers, convolution layers, activation functions and convolution layers, and in the second convolution layer of the multilayer perceptron, the output of the efficient additive attention module and the output of the activation function of the multilayer perceptron are fused, and finally the output of the multilayer perceptron is calculated by the second convolution layer of the multilayer perceptron.

[0084] Through multi-layer perceptrons, the output of the efficient additive attention module can be subjected to nonlinear transformation and dimensionality reduction.

[0085] The output of the multi-layer perceptron will enter the output layer, which can use global average pooling to compress the output of the multi-layer perceptron, and then process it through the fully connected layer and the normalization layer in turn, and finally obtain the image feature vector output by the image coding model. The fully connected layer can adjust the feature dimension, and the normalization layer standardizes the output of the fully connected layer.

[0086] Among them, the normalization layer of the output layer can be a layer normalization (LayerNorm) layer.

[0087] Optionally, in some embodiments, the image coding model can also be Figure 4 A multi-level residual connection design is adopted based on the structure of , which not only facilitates the back propagation of gradients, but also enables the shallow detail features to directly affect the deep feature representation.

[0088] The above-mentioned hierarchical structural design enables the image coding model of this embodiment to focus on local details and global semantic information at the same time, achieving powerful feature extraction capabilities while maintaining lightweight. Especially in the processing task of power signal images after power signal conversion, this structure can effectively capture various subtle changes and abnormal patterns in the signal, providing reliable feature support for subsequent classification tasks.

[0089] Based on the above structure, in this embodiment, the process of using any image coding model to process the power signal image to obtain the corresponding image feature vector is equivalent to:

[0090] The power signal image is processed respectively by using the central difference convolution module, the horizontal difference convolution module, the vertical difference convolution module, the adaptive difference convolution module and the standard convolution module of the image coding model to obtain multiple convolution features of the image coding model;

[0091] Multiple convolutional features of the image coding model are fused to obtain the image feature vector corresponding to the image coding model.

[0092] The above convolutional features are equivalent to Figure 4 The output of each convolution module within the detail enhancement convolution module.

[0093] See also Figure 3 In some optional embodiments, the classification result of the power signal to be processed may be determined according to multiple image feature vectors in the following manner:

[0094] fusing multiple image feature vectors according to weight values ​​of multiple image coding models to obtain a first fused vector of the power signal image, wherein the weight values ​​of the multiple image coding models are determined according to the power signal to be processed;

[0095] A classification result of the power signal to be processed is determined according to the first fusion vector.

[0096] Combination Figure 3 Taking the example of, assuming that there are n image coding models, n image feature vectors can be obtained, which are recorded as P1 to Pn in sequence, and the weight value of each image coding model can be recorded as S1 to Sn in sequence. In this embodiment, multiple image feature vectors can be fused as follows to obtain a first fused vector Ps: Ps=P1*S1+P2*S2+…Pn*Sn, and then, the first fused vector is processed by the classification model to obtain the classification result of the power signal to be processed.

[0097] In this embodiment, before fusing multiple image feature vectors, weight values ​​of multiple image coding models can be determined according to the power signal to be processed. Specifically, they can be determined according to the frequency domain characteristics of the power signal to be processed, according to the time domain characteristics of the power signal to be processed, according to the characteristics of the effective signal in the power signal to be processed, or according to the noise characteristics of the power signal to be processed, without limitation.

[0098] The advantage of determining the classification results in the above manner is that multiple image coding models can constitute a multi-expert system, and each image coding model is equivalent to an expert in processing power signal images. Therefore, this scheme can utilize differentiated multiple image coding models to implement a unique multi-expert collaborative learning strategy, thereby improving the accuracy of the classification results.

[0099] The multiple image coding models of this embodiment can be obtained through multiple methods.

[0100] One method of obtaining multiple image coding models may be:

[0101] Randomly generate multiple different random seeds;

[0102] For each random seed, initialization is performed according to the random seed to obtain an initialization model corresponding to the random seed, and the initialization model is trained using sample data to obtain an image coding model corresponding to the random seed.

[0103] In this embodiment, any existing random number generation algorithm can be used to generate a random string of a specific length, and the random string is used as a random seed. Repeating the random string several times can obtain multiple different random seeds. The number of random seeds is consistent with the number of image coding models. For example, if n image coding models need to be constructed, n different random seeds can be generated.

[0104] For each random seed, the existing model initialization method can be used to process the random seed and generate a Figure 4 The initialization model of the structure shown, the model parameters contained in the initialization model are determined by the model initialization method and random seed used.

[0105] The model initialization method used to generate the initialization model includes but is not limited to Xavier initialization.

[0106] Moreover, for each initialization model, the initialization model can be trained using sample data after the initialization model is obtained, and after the training is completed, the initialization model can be trained into an image coding model. The method of training the model can refer to the relevant prior art and will not be described in detail.

[0107] The advantage of obtaining multiple image coding models according to the above method is that it not only ensures a certain degree of complementarity between different image coding models, but also avoids the additional complexity brought by using different model structures. Experiments show that this strategy enables different image coding models to focus on different feature dimensions of the power signal to be processed. For example, one image coding model may tend to extract high-frequency features of the power signal image, another image coding model may tend to extract low-frequency features, another image coding model may tend to extract transient features, and another image coding model may tend to extract other features, thereby forming a comprehensive feature understanding system.

[0108] Moreover, using different random seeds when obtaining each image coding model can ensure that multiple image coding models can analyze the processed signals in different dimensions based on their own different model parameters. Therefore, in the process of training the image coding models, even if they share the same network structure, different feature extraction preferences can be evolved during the training process.

[0109] One method of obtaining multiple image coding models may be:

[0110] Get an initialization model and multiple different sample data sets;

[0111] For each sample data set, the sample data set is used to train the initialization model to obtain an image coding model corresponding to the sample data set.

[0112] In this embodiment, only one random seed may be generated, and based on the random seed and any model initialization method, a model having Figure 4 Initialization model for the structure shown.

[0113] On the other hand, multiple different sample data sets can be prepared in advance, each sample data set includes multiple sample data that can be used to train the initialization model. The number of sample data sets can be consistent with the number of image coding models to be constructed. For example, if n image coding models are to be constructed, n different sample data sets can be prepared.

[0114] Then, for each sample data set, an initialization model obtained by training the sample data contained in the sample data set can be used to obtain an image coding model corresponding to the sample data set, thereby obtaining n image coding models corresponding to n sample data sets.

[0115] The weight values ​​of multiple image coding models can be determined according to the noise characteristics of the power signal to be processed. For details, see Figure 3 , the weight values ​​of multiple image coding models can be determined as follows:

[0116] extracting a time series noise signal of a power signal to be processed;

[0117] A dynamic weight vector of the power signal to be processed is determined according to the time series noise signal, wherein the dynamic weight vector includes components corresponding to a plurality of image coding models one by one, and each component is a weight value of a corresponding image coding model.

[0118] Optionally, the method of determining the dynamic weight vector of the power signal to be processed according to the time series noise signal may be:

[0119] Process the time series noise signal according to the long short-term memory network to obtain the noise characteristics;

[0120] Normalizing the noise feature to obtain a normalized noise feature;

[0121] The normalized noise characteristics are processed based on a pre-built weight calculator to obtain a dynamic weight vector of the power signal to be processed.

[0122] Combined with the previous example, assuming that there are n image coding models, the obtained dynamic weight vector can be expressed as S (S1, S2, S3, ... Sn), where S1 to Sn are the n components of the dynamic weight vector, and are also the weight values ​​corresponding to the n image coding models.

[0123] The method of extracting the timing noise signal can be to use any existing signal noise reduction algorithm to process the power signal to be processed, obtain the noise reduced signal, and then subtract the power signal to be processed from the noise reduced signal, and the difference obtained can be used as the timing noise signal of the power signal to be processed.

[0124] like Figure 3 As shown, after obtaining the time series noise signal, the time series noise signal can be input into a long short-term memory network (Long Short-Term Memory, LSTM), and the noise feature corresponding to the time series noise signal can be obtained by using LSTM, and then the noise feature can be normalized to obtain a normalized noise feature. In this embodiment, the noise feature can be normalized based on layer normalization (LayerNorm).

[0125] The normalized noise features and noise features can be input into the weight calculator together. The weight calculator can calculate the input normalized noise features and noise features through a normalized exponential function (such as a softmax function), and update the preset initial weight vector based on the calculation results. The update process can be iterated multiple times until the number of iterations reaches a preset upper limit of the number of iterations. The weight calculator can output the vector obtained at the end of the last iteration as a dynamic weight vector.

[0126] The dimension of the initial weight vector may be consistent with the dimension of the dynamic weight vector, and the value of each component in the initial weight vector may be a preset value, for example, each component of the initial weight vector may be equal to 0.01.

[0127] The operation data of the power system often contains various types of noise, which may come from multiple sources such as equipment operation, environmental interference or measurement errors. Existing signal processing methods often regard noise as an interference factor that needs to be eliminated. In this embodiment, the time series noise signal is analyzed through the LSTM network, and the weight distribution of different image coding models is dynamically adjusted according to the analysis results, so as to adaptively combine multiple image coding models that tend to extract different dimensional features according to different types of noise characteristics to obtain more accurate classification results.

[0128] For example, when the power signal to be processed contains high-frequency noise, the above method can make the image coding model that tends to extract high-frequency features have a higher weight value; when the power signal to be processed contains burst noise, the above method can make the image coding model that tends to extract transient features have a higher weight value.

[0129] Therefore, this dynamic adaptation mechanism can significantly improve the robustness and generalization ability of this solution, achieve optimal feature fusion for various types of processed signals, and effectively improve the accuracy of classification results.

[0130] In summary, when the method of this embodiment processes the power signal to be processed, it implements a "divide and conquer" strategy based on multiple image coding models. The multiple image coding models as experts process different features of the power signal to be processed respectively, and then the noise features of the power signal to be processed are analyzed by LSTM to optimize the combination of multiple image coding models according to dynamic weight values. This design can not only effectively process various complex patterns in power signals, but also maintain a low computational overhead.

[0131] Moreover, when the image enhancement model is the DEFormer encoder, the detail enhancement design of the encoder is particularly suitable for processing image features after power signal conversion, which can capture subtle changes in the signal and further improve the accuracy of the classification results.

[0132] The experimental results on the power quality signal classification task (as shown in Table 1) show that the method of this embodiment has achieved significant performance improvement compared with the existing methods. Wherein PowerVisionExpert represents the classification method of this embodiment.

[0133] Specifically, PowerVisionExpert has achieved a level of 0.9957 in terms of accuracy, recall and F1 score, far exceeding existing models based on 1D convolutional neural network (CNN), 2D convolutional neural network (CNN), encoder (Transformer) and convolutional bidirectional long short-term memory neural network (CNN-BILSTM). This result fully demonstrates the superiority of the method of this embodiment in the task of classifying power system operation data. Compared with the prior art, the method of this embodiment not only improves the classification accuracy, but also has better generalization ability and robustness.

[0134] Table 1

[0135] Model Precision Recall F1-Score CNN (1D) 0.8585 0.8347 0.8347 CNN (2D) 0.9795 0.9772 0.9776 Transformer 0.6849 0.6027 0.5651 CNN-BILSTM 0.7577 0.7696 0.7406 CNN-TCN-ATTENTION 0.9500 0.9503 0.9498 PowerVisionExpert 0.9957 0.9957 0.9957

[0136] Among them, CNN (1D) regards the power signal as a time series signal, uses a one-dimensional convolutional neural network model, and extracts features from the time series signal through convolution kernels of different sizes. This method focuses on capturing local features in the time dimension and is particularly suitable for processing continuous and ordered data.

[0137] CNN (2D) first converts the time series power signal into a two-dimensional format, and then uses a deep convolutional neural network architecture (such as VGG16) to process the signal in the two-dimensional format. This method is able to capture more complex patterns and is particularly suitable for processing structured two-dimensional data.

[0138] Transformer is a model based on the self-attention mechanism that can effectively handle long-distance dependencies and is particularly suitable for complex sequence prediction tasks. Transformer's parallel processing capability gives it a significant advantage in processing longer time series data.

[0139] CNN-BILSTM can be seen as a combination of convolutional neural network (CNN) and bidirectional long short-term memory network (BILSTM). The model first uses CNN to extract the features of the power signal, and then captures the contextual relationship of the power signal as a time series through BILSTM. This combined design can fully understand the depth and breadth of the power signal.

[0140] CNN-TCN-ATTENTION is a model that combines CNN, TCN and attention mechanism. The model first uses CNN for spatial feature extraction, then uses TCN to capture temporal dependencies, and finally enhances the expression of key features through the attention mechanism. This design integrates the efficiency of convolutional neural networks and the ability to process long sequences, and is particularly suitable for processing tasks with complex temporal dependencies. TCN stands for Temporal Convolutional Networks (TCN).

[0141] The results in Table 1 show that the method of this embodiment (PowerVisionExpert) is significantly better than the existing methods in all evaluation indicators. Specifically, compared with the basic CNN (1D) and CNN (2D), the various indicators of the method of this embodiment are improved by about 3 to 13 percentage points; compared with models such as Transformer and CNN-BILSTM, the various indicators of the method of this embodiment are improved by about 25 to 43 percentage points; compared with the CNN-TCN-ATTENTION model, the various indicators of the method of this embodiment are improved by about 4.5 percentage points.

[0142] This embodiment also provides a power system signal classification system based on a hybrid expert system and visual processing, see Figure 5 , the system may include the following units.

[0143] An obtaining unit 501 obtains a power signal to be processed from a power system;

[0144] A conversion unit 502 converts the power signal to be processed into a power signal image;

[0145] The processing unit 503 processes the power signal image using multiple image coding models to obtain multiple image feature vectors of the power signal image, wherein the model parameters of the multiple image coding models are different from each other, and each image feature vector is obtained by processing the corresponding image coding model;

[0146] The classification unit 504 determines a classification result of the power signal to be processed according to the multiple image feature vectors.

[0147] Optionally, when the conversion unit 502 converts the power signal to be processed into a power signal image, it is used to:

[0148] The power signal to be processed is processed based on the Gram angular field method to obtain a power signal image.

[0149] Optionally, each image coding model includes a central differential convolution module, a horizontal differential convolution module, a vertical differential convolution module, an adaptive differential convolution module and a standard convolution module;

[0150] When the processing unit 503 processes the power signal image using any image coding model to obtain the corresponding image feature vector, it is used to:

[0151] The power signal image is processed respectively by using the central difference convolution module, the horizontal difference convolution module, the vertical difference convolution module, the adaptive difference convolution module and the standard convolution module of the image coding model to obtain multiple convolution features of the image coding model;

[0152] Multiple convolutional features of the image coding model are fused to obtain the image feature vector corresponding to the image coding model.

[0153] Optionally, when the classification unit 504 determines the classification result of the power signal to be processed according to the multiple image feature vectors, it is used to:

[0154] fusing multiple image feature vectors according to weight values ​​of multiple image coding models to obtain a first fused vector of the power signal image, wherein the weight values ​​of the multiple image coding models are determined according to the power signal to be processed;

[0155] A classification result of the power signal to be processed is determined according to the first fusion vector.

[0156] Optionally, the method in which the classification unit 504 determines the weight values ​​of multiple image coding models according to the power signal to be processed includes:

[0157] extracting a time series noise signal of a power signal to be processed;

[0158] A dynamic weight vector of the power signal to be processed is determined according to the time series noise signal, wherein the dynamic weight vector includes components corresponding to a plurality of image coding models one by one, and each component is a weight value of a corresponding image coding model.

[0159] Optionally, when the classification unit 504 determines the dynamic weight vector of the power signal to be processed according to the time series noise signal, it is used to:

[0160] Process the time series noise signal according to the long short-term memory network to obtain the noise characteristics;

[0161] Normalizing the noise feature to obtain a normalized noise feature;

[0162] The normalized noise characteristics are processed based on a pre-built weight calculator to obtain a dynamic weight vector of the power signal to be processed.

[0163] Optionally, the method for the processing unit 503 to obtain multiple image coding models includes:

[0164] Randomly generate multiple different random seeds;

[0165] For each random seed, initialization is performed according to the random seed to obtain an initialization model corresponding to the random seed, and the initialization model is trained using sample data to obtain an image coding model corresponding to the random seed.

[0166] Optionally, the method for the processing unit 503 to obtain multiple image coding models includes:

[0167] Get an initialization model and multiple different sample data sets;

[0168] For each sample data set, the sample data set is used to train the initialization model to obtain an image coding model corresponding to the sample data set.

[0169] Optionally, when the classification unit 504 determines the classification result of the power signal to be processed according to the multiple image feature vectors, it is used to:

[0170] fusing a plurality of image feature vectors and the power signal to be processed to obtain a second fused vector;

[0171] A classification result of the power signal to be processed is determined according to the second fusion vector.

[0172] Optionally, before the conversion unit 502 converts the power signal to be processed into a power signal image, it is further used to:

[0173] Performing signal preprocessing on the power signal to be processed to obtain a preprocessed power signal;

[0174] Convert the power signal to be processed into a power signal image, including:

[0175] The preprocessed power signal is converted into a power signal image.

[0176] This embodiment provides a power system signal classification system based on a hybrid expert system and visual processing. Its working principle can be found in the relevant steps of the power system signal classification method based on a hybrid expert system and visual processing in the aforementioned embodiment, and will not be repeated here.

[0177] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0178] For the convenience of description, the above system or device is described by dividing it into various modules or units according to its functions. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0179] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0180] Finally, it should be noted that, in this article, relational terms such as first, second, third and fourth are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article 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, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0181] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for classifying power system signals based on hybrid expert system and visual processing, characterized in that: include: Obtaining a power signal to be processed from a power system; Converting the power signal to be processed into a power signal image; Processing the power signal image using multiple image coding models to obtain multiple image feature vectors of the power signal image, wherein the model parameters of the multiple image coding models are different from each other, and each of the image feature vectors is obtained by processing the corresponding image coding model; A classification result of the power signal to be processed is determined according to the plurality of image feature vectors.

2. The method according to claim 1, characterized in that The converting the to-be-processed power signal into a power signal image comprises: The power signal to be processed is processed based on the Gram angular field method to obtain a power signal image.

3. The method according to claim 1, characterized in that Each of the image coding models includes a central differential convolution module, a horizontal differential convolution module, a vertical differential convolution module, an adaptive differential convolution module and a standard convolution module; The process of processing the power signal image using any of the image coding models to obtain a corresponding image feature vector includes: Using the central difference convolution module, the horizontal difference convolution module, the vertical difference convolution module, the adaptive difference convolution module and the standard convolution module of the image coding model to process the power signal image respectively, and obtain multiple convolution features of the image coding model; Multiple convolutional features of the image coding model are fused to obtain an image feature vector corresponding to the image coding model.

4. The method according to claim 1, characterized in that: Determining the classification result of the power signal to be processed according to the plurality of image feature vectors includes: fusing the plurality of image feature vectors according to weight values ​​of the plurality of image coding models to obtain a first fused vector of the power signal image, wherein the weight values ​​of the plurality of image coding models are determined according to the power signal to be processed; A classification result of the power signal to be processed is determined according to the first fusion vector.

5. The method according to claim 4, characterized in that The method for determining the weight values ​​of the plurality of image coding models according to the power signal to be processed comprises: Extracting a time series noise signal of the power signal to be processed; A dynamic weight vector of the power signal to be processed is determined according to the timing noise signal, wherein the dynamic weight vector includes components corresponding to a plurality of the image coding models one by one, and each of the components is a weight value of the corresponding image coding model.

6. The method according to claim 5, characterized in that The step of determining the dynamic weight vector of the power signal to be processed according to the timing noise signal comprises: Processing the time series noise signal according to the long short-term memory network to obtain noise characteristics; Normalizing the noise feature to obtain a normalized noise feature; The normalized noise feature is processed based on a pre-built weight calculator to obtain a dynamic weight vector of the power signal to be processed.

7. The method according to claim 1, characterized in that The method for obtaining a plurality of the image coding models comprises: Randomly generate multiple different random seeds; For each of the random seeds, initialization is performed according to the random seed to obtain an initialization model corresponding to the random seed, and the initialization model is trained using sample data to obtain an image coding model corresponding to the random seed.

8. The method according to claim 1, characterized in that The method for obtaining a plurality of the image coding models comprises: Get an initialization model and multiple different sample data sets; For each of the sample data sets, the initialization model is trained using the sample data set to obtain an image coding model corresponding to the sample data set.

9. The method according to claim 1, characterized in that: Determining the classification result of the power signal to be processed according to the plurality of image feature vectors includes: fusing a plurality of the image feature vectors and the power signal to be processed to obtain a second fused vector; A classification result of the power signal to be processed is determined according to the second fusion vector.

10. The method according to claim 1, characterized in that Before converting the to-be-processed power signal into a power signal image, the method further includes: Performing signal preprocessing on the power signal to be processed to obtain a preprocessed power signal; The converting the to-be-processed power signal into a power signal image comprises: The preprocessed power signal is converted into a power signal image.

Images