Training device, training method and application device of power fault identification network model

Abstract

The invention provides a training device, a training method and an application device of a power failure recognition network model. The training device of the power fault identification network model comprises a conversion module used for converting a plurality of one-dimensional current time sequence data in a training sample into a plurality of current two-dimensional data and converting a plurality of one-dimensional voltage time sequence data in the training sample into a plurality of voltage two-dimensional data; the dual-channel multi-layer perceptron is used for carrying out transverse dimension and longitudinal dimension feature fusion on the plurality of current feature vectors and carrying out transverse dimension and longitudinal dimension feature fusion on the plurality of voltage feature vectors; the multi-head cross attention module is used for performing multi-head cross attention processing on output vectors of the dual-channel multi-layer perceptron module, and performing self-attention processing on combined feature data of current attention feature data and voltage attention feature data obtained through cross attention processing; and training the power fault identification network model based on the output characteristic quantity obtained by the self-attention processing.

Description

Technical Field

[0001] The present application relates to the technical field of power grid fault identification, and in particular to a training device, a training method and an application device for a power fault identification network model. Background Art

[0002] With the rapid development of social economy and the growth of energy demand, the distribution network is becoming increasingly important for the stable supply of electricity. However, the distribution network often faces many potential fault risks during the power supply process, such as short circuit and overload. At present, the power grid fault detection methods based on deep learning directly process channel data such as voltage and current waveforms, or superimpose channel data to extract features, or introduce multi-modal enhanced feature representation, which has a weak effect on suppressing noise interference, affecting the accuracy of system identification and increasing the complexity of the power grid fault detection system. Summary of the invention

[0003] The present application provides a training device, a training method and an application device for a power fault identification network model, which is used to solve the problem that the current deep learning-based power grid fault detection method has a weak suppression effect on noise interference. The following introduces various aspects involved in the present application.

[0004] In a first aspect, the present application provides a training device for a power fault identification network model, comprising: a conversion module, used to convert multiple one-dimensional current time series data in a training sample into multiple current two-dimensional data, and convert multiple one-dimensional voltage time series data in a training sample into multiple voltage two-dimensional data; a feature extraction module, used to perform feature extraction on multiple current two-dimensional data to obtain multiple two-dimensional current feature vectors, and perform feature extraction on multiple voltage two-dimensional data to obtain multiple two-dimensional voltage feature vectors; a dual-channel multi-layer perceptron module, used to perform feature fusion of horizontal and vertical dimensions on multiple current feature vectors, and perform feature fusion of horizontal and vertical dimensions on multiple voltage feature vectors, the horizontal dimension is used for continuous time data features, and the vertical dimension is used for cross-time data features; a multi-head cross-attention module, used to perform multi-head cross-attention fusion processing on the output vectors of the dual-channel multi-layer perceptron module respectively to obtain current attention feature data and voltage attention feature data, perform self-attention processing on the joint feature data of the current attention feature data and the voltage attention feature data, and train the power fault identification network model based on the output feature quantity obtained by the self-attention processing.

[0005] In the second aspect, the present application provides an application device of a power fault identification network model, comprising: a recording module for acquiring current time series data and voltage time series data to be identified; a model processing module for inputting the current time series data and voltage time series data to be identified into a pre-trained power fault identification network model for processing to obtain corresponding identification results; wherein the power fault identification network model is a model trained by the training device described in the first aspect.

[0006] In a third aspect, the present application provides a training method for a power fault identification network model, including: converting multiple one-dimensional current time series data in a training sample into multiple current two-dimensional data, and converting multiple one-dimensional voltage time series data in a training sample into multiple voltage two-dimensional data; performing feature extraction on multiple current two-dimensional data to obtain multiple two-dimensional current feature vectors, and performing feature extraction on multiple voltage two-dimensional data to obtain multiple two-dimensional voltage feature vectors; using a dual-channel multi-layer perceptron module to perform feature fusion in horizontal and vertical dimensions on multiple current feature vectors, and perform feature fusion in horizontal and vertical dimensions on multiple voltage feature vectors, the horizontal dimension is used for continuous time data features, and the vertical dimension is used for cross-time data features; performing multi-head cross-attention fusion processing on the output vectors of the dual-channel multi-layer perceptron module to obtain current attention feature data and voltage attention feature data, performing self-attention fusion on the joint feature data of the current attention feature data and the voltage attention feature data, and training the power fault identification network model based on the output feature quantity obtained by the self-attention processing.

[0007] In a fourth aspect, the present application provides an application method of a power fault identification network model, comprising: obtaining current time series data and voltage time series data to be identified; inputting the current time series data and voltage time series data to be identified into a pre-trained power fault identification network model for processing to obtain a corresponding identification result; wherein the power fault identification network model is a network model trained using the training method described in the third aspect.

[0008] In a fifth aspect, the present application provides an electronic device, comprising: a memory for storing code; and a processor connected to the memory, for executing the code stored in the memory, so that the electronic device executes the method described in the third aspect or the fourth aspect.

[0009] In a sixth aspect, the present application provides a non-volatile computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement the method described in the third aspect or the fourth aspect.

[0010] In a seventh aspect, the present application provides a computer program product, which, when executed on an electronic device, enables the electronic device to implement the method described in the third aspect or the fourth aspect.

[0011] In the embodiment of the present application, one-dimensional long time series data of current and voltage are converted into two-dimensional data. The cross-feature fusion of the horizontal and vertical dimensions can be realized through the improved dual-channel multi-layer perceptron, and the feature fusion of long time series data can be realized. By strengthening the time series features, it helps to weaken noise interference. The embodiment of the present application combines the dual-path neural network structure and the hybrid multi-layer perceptron to construct a multi-scale fault recognition network based on the improved cross-attention mechanism. The joint feature data of the current attention feature data and the voltage attention feature data is obtained by using the cross-attention unit, and the joint feature data is processed based on the self-attention unit to complete the multi-dimensional feature fusion of the time series feature data. The power fault recognition network is trained and applied, which helps to strengthen the time series correlation between the current and voltage measurement data, weaken noise interference, and improve the accuracy of model fault recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in describing the embodiments of the present application are briefly introduced below.

[0013] Figure 1 This is a structural diagram of a typical deep learning power grid fault detection system.

[0014] Figure 2 It is a schematic diagram of a training device for a power fault identification network model provided in an embodiment of the present application.

[0015] Figure 3 yes Figure 2 A schematic diagram of a possible implementation of a power fault identification network model is shown.

[0016] Figure 4 yes Figure 3 A schematic diagram of a possible implementation of the conversion module 210 is shown.

[0017] Figure 5 It is a schematic diagram of a multi-layer perceptron module.

[0018] Figure 6 yes Figure 3 A schematic diagram of a possible implementation of a dual-channel multi-layer perceptron module 230 is shown.

[0019] Figure 7 yes Figure 3 A schematic diagram of a possible implementation of the multi-head cross-attention module 240 is shown.

[0020] Figure 8 yes Figure 3 A schematic diagram of another possible implementation of the multi-head cross-attention module 240 is shown.

[0021] Figure 9a-9b yes Figure 8 A possible schematic diagram of the multi-head cross-attention module 240 is shown.

[0022] Fig.10 It is a schematic diagram of an application device of the power fault identification network model provided in an embodiment of the present application.

[0023] Fig.11 yes Fig.10 A possible structural schematic diagram of an application device is shown.

[0024] Fig.12 It is a flowchart of a training method for a power fault identification network model provided in an embodiment of the present application.

[0025] Fig.13 It is a flow chart of the application method of the power fault identification network model provided in the embodiment of the present application.

[0026] Fig.14 yes Fig.12 A flowchart of a possible implementation of the training method shown.

[0027] Fig.15 yes Fig.13 A flow chart of a possible implementation of the application method shown.

[0028] Fig.16 It is a schematic diagram of a component unit / partial component unit of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The same or similar reference numerals are used in the drawings to represent the same or similar modules. It should be understood that the drawings are only schematic, and the protection scope of the present application is not limited thereto.

[0030] With the rapid development of social economy and the growth of energy demand, the distribution network is becoming increasingly important for the stable supply of electricity. Distribution networks often face many potential fault risks during the power supply process, such as short circuits and overloads. Traditional distribution network fault detection methods achieve systematic protection of distribution networks by setting protection settings and cascade communications, which usually rely on rule-making or algorithms based on specific models. Since rules or parameters need to be manually formulated, it is difficult to adapt to rapidly changing situations, cannot handle complex and nonlinear system situations, and also requires personnel with professional knowledge for design and maintenance.

[0031] With the development of artificial intelligence technology, digital relay protection systems have increasingly introduced the ideas and methods of expert systems to make up for the shortcomings of traditional systems. The current distribution network is developing in the direction of intelligence and technology, which enhances the security and stability of the system, as well as its adaptability after the access of new energy sources (such as wind power and photovoltaic power). As an important technology in the field of artificial intelligence, deep learning has powerful data processing and feature extraction capabilities. It can learn and analyze laws from a large amount of data to achieve accurate detection and rapid response to distribution network faults. The fault detection method based on deep learning is highly adaptive, can automatically learn features and laws from data, and can handle large-scale data, nonlinear relationships, and high-dimensional features.

[0032] Figure 1 This is a typical structural diagram of a deep learning power grid fault detection system. Figure 1 As shown, the power grid fault detection system is a digital automation system, which mainly relies on the measurement of physical quantities, internal digital logic and threshold triggering to determine whether it is a fault signal, and then take protective measures. The power grid fault detection system has the problems of many internal signals and complex structure.

[0033] At present, most of the power grid fault detection methods based on deep learning directly process waveform data such as voltage and current channels, or superimpose channel data to extract features, or introduce multimodal enhanced feature representation. The suppression effect of noise interference is weak, which affects the accuracy of system identification and increases the complexity of the system.

[0034] Therefore, it is necessary to design a technical solution for power fault type identification with good noise interference suppression effect and high accuracy.

[0035] Based on this, the present application embodiment proposes a training device for a power fault identification network model. Figure 2 As shown, the training device 200 for the power fault identification network model may include a conversion module 210 , a feature extraction module 220 , a dual-channel multi-layer perceptron module 230 , and a multi-head cross attention module 240 .

[0036] Figure 3 yes Figure 2A schematic diagram of a possible implementation of the power fault identification network model is shown in FIG. A-Data and V-Data represent data of the current analog channel and data of the voltage analog channel, respectively, including real-time current data and real-time voltage data.

[0037] The conversion module 210 is used to convert a plurality of one-dimensional current time series data in the training sample into a plurality of two-dimensional current data, and to convert a plurality of one-dimensional voltage time series data in the training sample into a plurality of two-dimensional voltage data.

[0038] In order to realize the modeling of long time series waveform data, it is necessary to transform the data through a coding layer, that is, to perform segmentation operation and reorganize it into two-dimensional data.

[0039] Exemplarily, the structure of the conversion module 210 may be as follows: Figure 4 As shown in the figure, if the length of the original long time series waveform data (such as one-dimensional current time series data) is L and the width is C, the length, width and height after block division are 2P, S and C respectively. 2P is the length of the block. For the convenience of subsequent slicing, 2P and S can be set to be integer multiples of the block size (Patch_size), and 2P and S are equal. The model output after block division is X s The expression is as follows:

[0040] X S =Segmention(X W ,Patch_size,S,C)

[0041] Among them, X W Represents the output volume of the encoding layer.

[0042] After the dimension transformation of the block operation, the one-dimensional multiple current time series data in the training sample can be converted into multiple current two-dimensional data, and the one-dimensional multiple voltage time series data in the training sample can be converted into multiple voltage two-dimensional data. In this way, the feature information extracted in the horizontal and vertical directions is cross-time series, which helps to improve the feature correlation of long time series waveform data.

[0043] The feature extraction module 220 is used to perform feature extraction on a plurality of two-dimensional current data to obtain a plurality of two-dimensional current feature vectors, and to perform feature extraction on a plurality of two-dimensional voltage data to obtain a plurality of two-dimensional voltage feature vectors.

[0044] The feature extraction module 220 can be implemented in a variety of ways, such as Gaussian feature extraction and multi-scale feature extraction. Gaussian features are used for residual connection, which helps to accelerate model convergence and avoid gradient disappearance. Multi-scale feature extraction can be used for multi-dimensional representation of analog data, which helps to ensure the depth of feature extraction. Gaussian feature extraction and multi-scale feature extraction are exemplarily described below.

[0045] Gaussian transform is a commonly used feature extraction method. It transforms the data by using the Gaussian kernel function to map the data from the original space to a higher-dimensional feature space, which facilitates better extraction of data features. The Gaussian kernel function is expressed as follows:

[0046]

[0047] Among them, x and y represent two input vectors respectively, and sigma() represents the bandwidth parameter of the Gaussian kernel function, which is used to control the distribution range of data in the feature space.

[0048] For example, A and V can be used to represent the data of the current channel and the voltage channel in a certain period of time, respectively, and vector a and vector v can be used to represent the two data vectors. Then for data points A and V, the Gaussian transformation expression in the feature space is:

[0049] phi(a)=[K(a,a),K(a,v)]

[0050] phi(v)=[K(v,v),K(v,a)]

[0051] After Gaussian transformation, the original data points A and V are mapped to new feature vectors phi(a) and phi(v) in the feature space respectively. The new feature vectors fuse the data of two channels and can be used in subsequent classification tasks.

[0052] Since the window of a single convolutional layer is relatively fixed, it is not possible to flexibly display the temporal feature relationship in the face of a longer time series. In some implementations, multi-scale convolution feature extraction can be used. For example and not limitation, three convolutions with different window sizes can be used to collect information of different time spans respectively, and the feature data of the three dimensions can be fused together to achieve the effect of strengthening the temporal correlation.

[0053] First, define three convolution kernels of different scales as f 1 、f 2 、f 3 , the corresponding convolution kernel size is k 1 , k 2 , k 3 Perform one-dimensional convolution operations on the current data vector a (such as current two-dimensional data) to obtain three new feature vectors n 1 、n 2 、n 3 , then the corresponding mathematical expression is:

[0054] n 1 =a*f 1

[0055] n2 =a*f 2

[0056] n 3 =a*f 3

[0057] Among them, * represents a one-dimensional convolution operation.

[0058] The lengths of feature data generated by different convolution kernels are different, which may cause errors in data weights. In some implementations, the feature extraction module 220 is used to use multiple convolution kernels of different scales to perform one-dimensional convolution operations on the first current two-dimensional data to obtain a first current initial vector set, and use multiple convolution kernels of different scales to perform one-dimensional convolution operations on the first voltage two-dimensional data to obtain a first voltage initial vector set; uniformly fill any current initial vector shorter than the longest current initial vector in the first current initial vector set with zeros, and superimpose the vectors in the first current initial vector set after data expansion to obtain the corresponding current feature vector; uniformly fill any voltage initial vector shorter than the longest voltage initial vector in the first voltage initial vector set with zeros, and superimpose the vectors in the first voltage initial vector set after data expansion to obtain the corresponding voltage feature vector; wherein the first current two-dimensional data is any current two-dimensional data among multiple current two-dimensional data, and the first voltage two-dimensional data is any voltage two-dimensional data among multiple voltage two-dimensional data.

[0059] For example, the longest data frame (such as the current initial vector) can be used as a reference, short frame data (shorter than the longest data frame) can be uniformly padded with zeros, and each data frame can be expanded to a uniform length and then superimposed. Where m represents three convolution channels of different scales, n m That is, any eigenvector (such as the initial current vector). The corresponding mathematical expression is as follows:

[0060] n=padding(n m )

[0061] The three feature vectors are added together in a simple superposition method to obtain the data feature vector n of the feature collection. _merged .

[0062] n _merged =add(n 1 ,n 2 ,n 3 )

[0063] You can use convolutions with multiple different window sizes to collect information of different time spans, and fuse feature data of multiple dimensions together to help strengthen the correlation of time series. Setting the length of feature data generated by different convolution kernels to the same length helps avoid errors in data weights.

[0064] The dual-channel multi-layer perceptron module 230 is used to perform horizontal and vertical dimension feature fusion on multiple current feature vectors and to perform horizontal and vertical dimension feature fusion on multiple voltage feature vectors. The horizontal dimension is used for continuous time data features, and the vertical dimension is used for cross-time data features.

[0065] In some implementations, the dual-channel multi-layer perceptron module may be a multi-layer perceptron module based on residual connection.

[0066] In some other implementations, the horizontal dimension may also be used for data features across time, and the vertical dimension may be used for data features over continuous time.

[0067] Figure 5 This is a schematic diagram of an improved multi-layer perceptron module. Figure 5 As shown in the figure, by stacking multiple network structures to build a multi-layer perceptron model of long time series features, the two-dimensional time series feature representation of one-dimensional data can better achieve pixel-level feature fusion and reduce the interference of input noise. Multi-layer perceptrons have strong nonlinear modeling capabilities and usually perform well in multi-classification tasks.

[0068] Figure 6 yes Figure 3 FIG. 2 is a schematic diagram of a possible implementation of a dual-channel multi-layer perceptron module 230. Figure 6 As shown in the figure, a two-layer double-path multi-layer perceptron (DPMLP) structure is introduced, which can introduce the most original channel data into the feature representation of long time series, and perform horizontal and vertical feature fusion on the feature map respectively. The horizontal direction is the continuous time series, and the vertical direction is the cross-time data features, which reduces the complexity of model calculation while ensuring the deep features.

[0069] The DPMLP module uses a different network structure from the traditional one. It adds convolution operations to two different types of multi-layer perceptrons (MLPs), so that the deep-dimensional feature influence is added to the fully connected weight calculation. It can extract and fuse the features of the MLP within the channel, and extract and fuse the features between channels. Dimension switching can be achieved by array transposition.

[0070] Assume that the feature scale of the input data vector X is [w, d, n], i represents the number of input feature maps, i∈(1, S). Then the output M of the MLP between channels is *，j It can be expressed as:

[0071]

[0072] Among them, W 1 , W 2 They represent the weights of the activation function σ() and the normalization function LN() between channels. Then the output N of the MLP in the channel *，j It can be expressed as:

[0073]

[0074] Among them, W 3 , W 4 They represent the weights of the activation function σ() and the normalization function LN() within the channel respectively.

[0075] The embodiment of the present application introduces convolution into the multi-layer perceptron structure, and combined with the reorganized two-dimensional features, it can focus on the amplitude features within a larger time scale, and use the feature reuse structure of the residual connection to model data within multiple time scales.

[0076] In the embodiment of the present application, one-dimensional long time series data of current and voltage are converted into two-dimensional data. Through the improved dual-channel multi-layer perceptron, cross-feature fusion in the horizontal and vertical dimensions can be achieved, and long time series data feature fusion can be realized. By strengthening the time series features, it helps to weaken noise interference.

[0077] The multi-head cross-attention module 240 is used to perform multi-head cross-attention fusion processing on the output vectors of the dual-channel multi-layer perceptron module to obtain current attention feature data and voltage attention feature data, perform self-attention processing on the joint feature data of the current attention feature data and the voltage attention feature data, and train the power fault identification network model based on the output feature quantity obtained by the self-attention processing.

[0078] In a single self-attention mechanism, when encoding the information of the current position, it will over-focus on its own position, while multi-head attention will make more full use of the multi-dimensional features of the data. Compared with the traditional single attention mechanism, the multi-head attention mechanism introduces a multi-head structure, which divides the input sequence into multiple subsets of independent attention points for processing in parallel, and performs well in multi-target tasks. However, its multi-head structure usually uses the same input features and cannot focus on features of deeper dimensions.

[0079] Therefore, the embodiment of the present application proposes to utilize the feature extraction capability of convolution to perform multi-layer feature collection before inputting the multi-head structure. In some implementations, the multi-head cross attention module can be used to perform multi-layer feature collection on the output vector of the dual-channel multi-layer perceptron module to obtain feature data at different scales; the feature data at different scales are subjected to multi-head cross attention fusion processing to obtain current attention feature data and voltage attention feature data.

[0080] Figure 7 yes Figure 3 A schematic diagram of a possible implementation of the multi-head cross attention module 240 is shown. Figure 7 As shown, C 1 , C 2 , C 3 They represent the feature maps after different convolution kernels. After multi-head attention, multi-dimensional feature fusion is completed. It has been verified that the multi-scale and multi-channel multi-head attention mechanism can further improve the feature representation ability of the model and perform better when the data increases.

[0081] If a given query key Sum W 0 represents the concatenation weight of the multi-head structure. Then the overall representation of the multi-head attention is:

[0082]

[0083] Among them, the learnable parameters include different depth features j And the corresponding attention function f i , and the splicing weight W 0 Therefore, each head may focus on different parts at different feature depths, and can represent more complex feature weights than single window attention.

[0084] In the embodiment of the present application, the voltage and current channels are respectively fused with cross-attention and self-attention mechanisms to realize multi-dimensional feature fusion. Cross-attention is to capture local information by alternating attention within the feature map rather than the entire feature map, and to capture global information by applying attention between image blocks divided from a single-channel feature map. Both of these operations are less computationally intensive than the standard self-attention in the Transformer network. By alternating the application of attention between internal patches and patches, cross-attention can maintain equivalent performance at a lower computational cost, and a hierarchical network called cross-attention transformer (CAT) is constructed for other tasks.

[0085] Figure 8 yes Figure 3 A schematic diagram of a possible implementation of the multi-head cross-attention module 240 is shown. Figure 9a is a schematic diagram of a cross-attention unit of the multi-head cross-attention module 240. Figure 8 , Figure 9a As shown, use X t and Y tThey represent the current feature vector and voltage feature vector of the input model respectively. They are obtained by multi-scale encoders and retain the timing information. Therefore, the feature fusion interaction of the two channels can be realized in the time dimension. The core part of the cross attention network is the attention layer. It can be understood that X t and Y t It is also the output vector of the dual-channel multi-layer perceptron module 230.

[0086] The input current and voltage characteristic vector X t and Y t After being mapped by the linear attention layer (Linear), they become vectors (Q a ,Q v ), key (K a ,K v ), and value (V a ,V v ). Figure 9a In the example, Matmul corresponds to QK T , the scale corresponds to √d, where d represents the dimension of Q, K and V. The output of the module is the current attention feature F a and voltage attention feature F v The detailed expression is shown in the following formula:

[0087]

[0088] In order to allow the network to learn multi-dimensional features, F a The output is X t Generate key K as source sequence a , the attention layer applies Y t Generate Q as source sequence a . F v The value-taking process of Y is the opposite. t Generate key K as source sequence v , the amount of attention layer applied is X t Generate Q as source sequence v This mutual value-taking method allows the current channel and voltage channel information to influence each other and be fully integrated. The attention layer is after the feedback layer, and the residual link and normalization layer are also after the feedback layer, which can be used to generate the entire cross-modal attention network.

[0089] If you use Y av Represents the current attention feature F a and voltage attention feature F v The joint features of , the formula of cross attention can be expressed as follows:

[0090] y av =softmax((W Q S2 )(W K S 1 ) T )W V S 1

[0091] Among them, the sequence S 1 , S 2 Represent the current attention features F a Matrix, voltage attention feature F v The matrix, W Q , W K , W V are the weights corresponding to Q, K, and V respectively.

[0092] Calculate S 1 Key K, value V. Calculate S 2 The query Q is calculated based on K and Q, and the value V is applied to the attention matrix. The length of the output sequence is the same as S 2 Consistent.

[0093] Based on the cross-attention network, the self-attention network is introduced to establish a long-term network model. Figure 9b yes Figure 8 Schematic diagram of the self-attention unit of the multi-head cross attention module 240 shown. Figure 9b As shown, query Q av , key K av Sum value V av In the attention layer, all come from the joint feature Y av , the self-attention network is similar to the cross-attention network in structure. The main task of the self-attention layer is to establish the dependency relationship between the elements in the input sequence and to establish the feature model of the context in the input sequence. Therefore, the formula of self-attention is as follows:

[0094]

[0095] Based on the improved multi-head attention mechanism, dual-channel data can be fused at different feature scales. Therefore, the network model of the embodiment of the present application can also be called a multi-channel cross-attention network (MCCANet) model. The embodiment of the present application encodes the frame-based input current and voltage into a time series of current and voltage embedding representing the time context. The back-end classifier may include a cross-mode cross-attention mechanism for dynamically aligning current and voltage data, and implementing a self-attention mechanism between channels.

[0096] In some implementations, the training device 200 may further include an output module 250. The output module 250 may normalize the output result of the model, for example, by normalizing the output result of the model through a Softmax function.

[0097] The Softmax function, also known as the normalized exponential function, can compress a K-dimensional vector containing any real number into another K-dimensional real vector so that the range of each element is between (0,1) and the sum of all elements is 1.

[0098] In multi-classification tasks, the output of the model is often multiple probability parameters. For example, if there are α data categories, the classification model needs to have P α probability outputs to cover the entire classification task, where P i Represents the probability that the data belongs to the i-th category. Therefore, the output M of the network model can be:

[0099] M=[P 1 ,P 2 ...P α ]

[0100] Among them, the output result of the model is the fault category corresponding to the maximum probability, that is:

[0101] P Max =Max(P 1 ,P 2 …P α )

[0102] In some implementations, in the fault prediction stage, data from current and voltage channels can be collected in real time, fed into the model and calculated to determine whether the current state is faulty. The specific fault type is given and internal signals such as alarms are transmitted to drive external devices for protection. Compared with historical data, if the current channel data is abnormal but does not reach the fault action threshold, the current channel state can be analyzed to achieve fault prediction. Compared with conventional deep learning networks, the cross-attention mechanism can better capture the correlation between features, better capture the global information of long time series data, and reduce information loss.

[0103] In some implementations, the training device 200 may further include a deployment module. The deployment module is used to perform the following operations: deploy an initial network model, which is a simplified model of a power fault identification network model. Using the knowledge distillation (KD) and model relearning (Fine-tuning, FN) methods, new current time series data and new voltage time series data are collected to update the training samples, and the parameters of the initial network model are trained and optimized through the updated training samples, so that the recognition rate of the initial network model reaches a preset accuracy rate, that is, it gradually evolves into the desired target power fault identification network model.

[0104] In the embodiment of the present application, the one-dimensional long time series data of current and voltage are converted into two-dimensional data. The cross-feature fusion of the horizontal and vertical dimensions can be realized through the improved dual-channel multi-layer perceptron, and the feature fusion of long time series data can be realized. By strengthening the time series features, it helps to weaken noise interference. The embodiment of the present application combines the dual-path neural network (DPRNN) structure and the hybrid multi-layer perceptron (MLP-Mixer) to construct a multi-scale fault recognition network based on an improved cross-attention mechanism. The joint feature data of the current attention feature data and the voltage attention feature data are obtained by using the cross-attention unit, and the joint feature data is processed based on the self-attention unit to complete the multi-dimensional feature fusion of the time series feature data. The power fault recognition network is trained and applied, which helps to strengthen the time series correlation between the current and voltage measurement data, weaken noise interference, and improve the accuracy of model fault recognition. It helps to improve the data mining ability of the model and improve the dynamic modeling ability of the model.

[0105] Fig.10 Schematic diagram of an application device of the power fault identification network model provided in an embodiment of the present application. Fig.10 As shown, the application device 1000 of the power fault identification network model may include a first wave recording module 1010 and a model processing module 1030 .

[0106] The wave recording module 1010 is used to obtain the current time series data and voltage time series data to be identified.

[0107] The recording module 1010 can be a relay protection recording component, which usually plays the role or function of recording, monitoring, analysis, debugging and recording. For example, the recording function can record abnormal current and voltage waveforms occurring in the power grid for analyzing the cause of the fault and event replay. Fault analysis can provide detailed waveform data when the fault occurs to help engineers analyze the type and location of the fault in the power grid. The debugging function can help confirm the correct operation and response of the protection equipment during the debugging stage of the relay protection device. Event recording can record important events in the power grid and provide data support for subsequent power grid analysis and improvement. The fault location function can combine the recording data and protection action information to help determine the specific location of the fault point and guide the repair work.

[0108] In some implementations, the recording module 1010 may be a continuous recording component. The continuous recording component is used for analog data acquisition of the current channel and analog data acquisition of the voltage channel, and sends the collected data to the subsequent data processing module for model training and testing. In other embodiments, the recording module 1010 may be a fault recording component. The fault recording component is a recording component that is automatically executed when the device detects a fault. It can be used in traditional relay protection process components. The purpose is to record the waveforms before and after the fault for fault analysis and backup. The file can contain data from both analog and digital channels.

[0109] In other implementations, the recording module 1010 may also include a continuous recording component and a fault recording component. In the fault recording component, the fault analysis can provide waveform data before and after the fault occurs, such as taking a 5:9 (or 4:9) ratio mode before and after the fault, which is helpful for analyzing the type and location of the fault in the power grid; the triggering method may be a fault mark. For example, the continuous recording component can set a sampling frequency of 40Hz or 50Hz to collect real-time data on the AC current channel, and the data is processed by the network model for real-time monitoring of the power grid operation status. The functional parameter comparison of the continuous recording component and the fault recording component is shown in Table 1.

[0110] Table 1

[0111]

[0112] The model processing module 1030 is used to input the current time series data and voltage time series data to be identified into a pre-trained power fault identification network model for processing to obtain corresponding identification results. The power fault identification network model is a network model trained by the training device of the power fault identification network model described in any of the above items.

[0113] In some implementations, the application device 1000 of the power fault identification network model may also include a data preprocessing component. The data preprocessing component is used to classify and reorganize the AC channel data collected by the recording and broadcasting component to generate training samples and test samples. The data preprocessing component can perform preliminary processing on the collected waveform data, including filtering, type labeling, data feature extraction, data generation, and data set packaging. For example, using a Gaussian function to perform weighted averaging on the waveform neighborhood can effectively reduce noise interference. The data preprocessing component can read the backup file of the fault recording component and mark the sample data corresponding to the fault type for model training. For another example, wavelet transform can also be used to extract data features of the voltage and current channels.

[0114] Fig.11 yes Fig.10 A possible structural diagram of the application device is shown in FIG. Fig.11 As shown in the figure, the fault identification system based on the multi-channel cross-attention network model (MCCANet) can realize the identification and prediction functions of power grid faults by only monitoring the analog channels.

[0115] It is usually difficult to collect comprehensive multiple fault data in a short period of time, and the small number of fault samples may lead to the problem of biased parameters of the network model. In some implementations, data generation and optimization processing algorithms (DGOP) can be used to expand data training samples. The data generation and optimization processing algorithm can include two parts: regeneration of existing data and simulation of protection event data. Among them, the regeneration of existing data can be achieved through waveform data processing and data generation algorithms based on deep learning. The generated data is classified and packaged into a data set as a training sample for the network model, which helps to avoid the problem of biased model parameters due to the small number of fault samples.

[0116] In the embodiment of the present application, the one-dimensional long time series data of current and voltage are converted into two-dimensional data. The cross-feature fusion of the horizontal and vertical dimensions can be realized through the improved dual-channel multi-layer perceptron, and the feature fusion of long time series data can be realized. By strengthening the time series features, it helps to weaken noise interference. In the embodiment of the present application, the cross-attention unit is used to obtain the joint feature data of the current attention feature data and the voltage attention feature data, and the joint feature data is processed based on the self-attention unit to complete the multi-dimensional feature fusion of the time series feature data, and the power fault identification network is trained and applied, which helps to strengthen the time series correlation between the current and voltage measurement data, weaken noise interference, and improve the accuracy of model fault identification. It helps to improve the data mining ability of the model and improve the dynamic modeling ability of the model.

[0117] The embodiment of the present application builds a fault identification system based on a deep learning method to achieve real-time monitoring of power grid lines. By training a high-precision fault identification network model, the intelligent transformation of distribution network protection can be achieved. It helps to improve the high difficulty of operation and maintenance of traditional distribution network fault detection methods, improve the level of automation and intelligence of distribution networks, and achieve the accuracy and timeliness of fault detection. While improving the accuracy of fault identification, it can also achieve multiple functions such as load forecasting, power grid efficiency detection, fault warning, intelligent operation and maintenance, etc., which helps to improve the high degree of information integration of distribution terminals and improve the high degree of coordination and intelligence of power grid systems.

[0118] Combination of the above Figure 1-Figure 11 The training device and application device embodiments of the present application are described in detail. Figures 12 to 15 The method embodiment of the present application is described in detail. It should be understood that the description of the method embodiment corresponds to the description of the device embodiment, so the parts not described in detail can refer to the previous device embodiment.

[0119] Fig.12 FIG. 1 is a flow chart of a method for training a power fault identification network model provided in an embodiment of the present application. Fig.12 As shown, the training method of the power fault identification network model in the embodiment of the present application may mainly include steps S1210 to S1230, and these steps are described in detail below.

[0120] It should be pointed out that the size of the serial numbers of the steps in the embodiments of the present application does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0121] In step S1210, a plurality of one-dimensional current time series data in the training sample is converted into a plurality of two-dimensional current data, and a plurality of voltage time series data in the training sample is converted into a plurality of two-dimensional voltage data.

[0122] In step S1220, feature extraction is performed on a plurality of two-dimensional current data to obtain a plurality of two-dimensional current feature vectors, and feature extraction is performed on a plurality of two-dimensional voltage data to obtain a plurality of two-dimensional voltage feature vectors.

[0123] In step S1230, a dual-channel multi-layer perceptron module is used to perform feature fusion of the horizontal dimension and the vertical dimension on multiple current feature vectors, and to perform feature fusion of the horizontal dimension and the vertical dimension on multiple voltage feature vectors. The horizontal dimension is used for continuous time data features, and the vertical dimension is used for cross-time data features.

[0124] Optionally, the dual-channel multi-layer perceptron module is a multi-layer perceptron based on residual connection.

[0125] In step S1240, the output vectors of the dual-channel multi-layer perceptron module are subjected to multi-head cross-attention fusion processing to obtain current attention feature data and voltage attention feature data, and self-attention fusion is performed on the joint feature data of the current attention feature data and the voltage attention feature data. The power fault recognition network model is trained based on the output feature quantity obtained by the self-attention processing.

[0126] Optionally, in step S1220, feature extraction of multiple current two-dimensional data to obtain multiple two-dimensional current feature vectors, and feature extraction of multiple voltage two-dimensional data to obtain multiple two-dimensional voltage feature vectors may include: using multiple convolution kernels of different scales to perform one-dimensional convolution operations on the first current two-dimensional data to obtain a first current initial vector set, and using multiple convolution kernels of different scales to perform one-dimensional convolution operations on the first voltage two-dimensional data to obtain a first voltage initial vector set; uniformly padding any current initial vector shorter than the longest current initial vector in the first current initial vector set with zeros, and superimposing the vectors of the first current initial vector set after data expansion to obtain the corresponding current feature vector; uniformly padding any voltage initial vector shorter than the longest voltage initial vector in the first voltage initial vector set with zeros, and superimposing the vectors of the first voltage initial vector set after data expansion to obtain the corresponding voltage feature vector; wherein the first current two-dimensional data is any current two-dimensional data among the multiple current two-dimensional data, and the first voltage two-dimensional data is any voltage two-dimensional data among the multiple voltage two-dimensional data.

[0127] Optionally, the output vectors of the dual-channel multi-layer perceptron module are respectively subjected to multi-head cross-attention fusion processing to obtain current attention feature data and voltage attention feature data, including: performing multi-layer feature collection on the output vectors of the dual-channel multi-layer perceptron module to obtain feature data at different scales; and performing multi-head cross-attention fusion processing on the feature data at different scales to obtain current attention feature data and voltage attention feature data.

[0128] Optionally, the method of the embodiment of the present application further includes: deploying an initial network model, the initial network model being a simplified model of the power fault identification network model; using a knowledge distillation algorithm and a model relearning method to collect new current time series data and new voltage time series data to update the training samples. The parameters of the initial network model are trained and optimized through the updated training samples so that the recognition rate of the initial network model reaches a preset accuracy rate.

[0129] In the embodiment of the present application, the one-dimensional long time series data of current and voltage are converted into two-dimensional data. The cross-feature fusion of the horizontal and vertical dimensions can be realized through the improved dual-channel multi-layer perceptron, and the feature fusion of long time series data can be realized. By strengthening the time series features, it helps to weaken noise interference. The embodiment of the present application combines the dual-path neural network structure and the hybrid multi-layer perceptron to construct a multi-scale fault recognition network based on the improved cross-attention mechanism. The joint feature data of the current attention feature data and the voltage attention feature data are obtained by using the cross-attention unit, and the joint feature data is processed based on the self-attention unit to complete the multi-dimensional feature fusion of the time series feature data. The power fault recognition network is trained and applied, which helps to strengthen the time series correlation between the current and voltage measurement data, weaken noise interference, and improve the accuracy of model fault recognition. It helps to improve the data mining ability of the model and improve the dynamic modeling ability of the model.

[0130] Fig.13 FIG. 1 is a flow chart of an application method of a power fault identification network model provided in an embodiment of the present application. Fig.13 As shown, the application method of the power fault identification network model in the embodiment of the present application may mainly include steps S1310 to S1320, and these steps are described in detail below.

[0131] In step S1310, current time series data and voltage time series data to be identified are obtained.

[0132] In step S1320, the current time series data and voltage time series data to be identified are input into a pre-trained power fault identification network model for processing to obtain corresponding identification results. The power fault identification network model is a network model trained using the training method described above.

[0133] The method of the embodiment of the present application is further explained below in combination with some possible implementation methods of the present application.

[0134] Fig.14 yes Fig.12 A flow chart of a possible implementation of the training method shown in FIG. The embodiment of the present application uses a knowledge distillation algorithm plus a model relearning method to continuously optimize network parameters on the basis of simplifying model parameters, thereby lowering the threshold for system deployment. The initial deployment of the system requires generating an initial network model of nodes based on historical data, and then optimizing the model by continuously monitoring the power grid, ultimately achieving the preset accuracy. Fig.14 As shown, the training method of the embodiment of the present application may mainly include steps S1410 to S1480, and these steps are described in detail below.

[0135] In step S1410, start.

[0136] In step S1420, it is determined whether it is the first deployment. If it is the first deployment, it means that the deployed network is a simplified initial network model, and the process proceeds to step S1430; if it is not the first deployment, the process proceeds to step S1440.

[0137] In step S1430, network model training is performed.

[0138] In step S1440, it is determined whether the training samples need to be updated. If so, the process proceeds to step S1450; if not, the process proceeds to step S1460.

[0139] In step S1450, the parameters of the initial network model are updated and optimized using a knowledge distillation algorithm and a model relearning method.

[0140] In step S1460, a power fault identification network model that meets a preset accuracy is deployed.

[0141] In step S1470, based on the power fault identification network model, grid faults are predicted and identified to protect the power system.

[0142] In step S1480, end.

[0143] Complete the training and deployment process of the power fault identification network model.

[0144] Fig.15 yes Fig.13 A flow chart of a possible implementation of the application method shown. In the embodiment of the present application, before enabling the power fault identification network model, a sufficient amount of fault samples must first be collected, a data set must be generated through the data processing component, and the data must be sent to the initial network model for training. The initial network model can only be put into use when the prediction accuracy reaches the requirement. The data collected by the current AC channel and the voltage AC channel are monitored in real time to identify and predict faults in the power grid distribution lines. In order to adapt to different application scenarios and increasingly complex distribution networks, the power fault identification network model supports the input of new sample data, and incremental learning can be achieved by inputting new sample data. Fig.15 As shown, the application method of the embodiment of the present application may mainly include steps S1510 to S1590, and these steps are described in detail below.

[0145] In step S1510, start.

[0146] In step S1520, it is determined whether the power fault identification network model is enabled. If enabled, the process proceeds to step S1540; if not enabled, the process proceeds to step S1530.

[0147] In step S1530, it is determined whether training is still needed. If so, the process proceeds to step S1540; if not, the process proceeds to step S1550.

[0148] In step S1540, the protection component performs real-time monitoring of the power grid. The protection component can combine the model output to drive the hardware interface to implement the protection action, and then enter step S1560.

[0149] In step S1550, real-time recording is performed using a continuous recording component.

[0150] In step S1552, the power fault identification network model performs fault detection and identification, and if a fault occurs, it can be predicted and identified more accurately.

[0151] In step S1553, after the fault is identified, the data when the fault occurs is recorded and saved.

[0152] In step S1554, the fault type is determined.

[0153] In step S1555, a protection action corresponding to the fault type is executed.

[0154] In step S1560, after a failure occurs, the process proceeds to step S1570 and step S1580.

[0155] In step S1570, fault recording is performed, and the data when the fault occurs is recorded and saved.

[0156] In step S1571, the fault recording data is processed.

[0157] In step S1572, model training is performed, and then the process proceeds to step S1552.

[0158] In step S1580, a protection action corresponding to the fault type is executed, and the exit is usually determined by the protection component.

[0159] In step S1590, end.

[0160] Fig.16 Schematic diagram of component units / partial component units of an electronic device provided in an embodiment of the present application. Fig.16 As shown, the electronic device 1600 may include: a memory 1610 and at least one processor 1620 .

[0161] The memory 1610 is used to store codes or computer programs.

[0162] The processor 1620 is connected to the memory 1610 and is used to execute the code or computer program stored in the memory 1610 to control the electronic device 1600 to execute any of the processing methods described above.

[0163] Exemplarily, the computer program may be divided into one or more modules / units, one or more modules / units are stored in the memory 1610 and executed by the processor 1620 to complete the present application.

[0164] Those skilled in the art will understand that Fig.16 The electronic device 1600 is merely an example and does not constitute a limitation of the electronic device, and may include more or fewer components than shown in the figure, or combine some components, or different components.

[0165] The processor 1620 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor, or the processor may be any conventional processor, etc.

[0166] The electronic device 1600 provided in this embodiment can execute the above method embodiment, and its implementation principle and technical effect are similar, which will not be repeated here.

[0167] An embodiment of the present application further provides a non-volatile computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0168] An embodiment of the present application provides a computer program product. When the computer program product runs on an electronic device, the electronic device can implement the steps in the above-mentioned method embodiments when executing the computer program product.

[0169] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above method embodiments, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above method embodiments can be implemented. Among them, the computer program includes computer program code, which can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device that can carry the computer program code to the camera / electronic device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), read-only compact disc (CD-ROM), magnetic tape, floppy disk and optical data storage device. The computer-readable storage medium mentioned in the present application can be a non-volatile storage medium, in other words, it can be a non-transient storage medium.

[0170] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0171] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0172] In the embodiments provided in the present application, it should be understood that the disclosed devices / equipment and methods can be implemented in other ways. For example, the device / equipment embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0173] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0174] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0175] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0176] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A training device for a power fault identification network model, characterized in that: include: A conversion module, used to convert a plurality of one-dimensional current time series data in a training sample into a plurality of two-dimensional current data, and to convert a plurality of one-dimensional voltage time series data in the training sample into a plurality of two-dimensional voltage data; A feature extraction module, used for performing feature extraction on the plurality of two-dimensional current data to obtain a plurality of two-dimensional current feature vectors, and performing feature extraction on the plurality of two-dimensional voltage data to obtain a plurality of two-dimensional voltage feature vectors; A dual-channel multi-layer perceptron module, used to perform feature fusion of the horizontal dimension and the vertical dimension on the multiple current feature vectors, and perform feature fusion of the horizontal dimension and the vertical dimension on the multiple voltage feature vectors, wherein the horizontal dimension is used for continuous time data features, and the vertical dimension is used for cross-time data features; A multi-head cross-attention module is used to perform multi-head cross-attention fusion processing on the output vectors of the dual-channel multi-layer perceptron module to obtain current attention feature data and voltage attention feature data, perform self-attention processing on the joint feature data of the current attention feature data and the voltage attention feature data, and train the power fault identification network model based on the output feature quantity obtained by the self-attention processing.

2. The training device according to claim 1, characterized in that The training device also includes: A feature extraction module, used to use multiple convolution kernels of different scales to perform one-dimensional convolution operations on the first current two-dimensional data to obtain a first current initial vector set, and use multiple convolution kernels of different scales to perform one-dimensional convolution operations on the first voltage two-dimensional data to obtain a first voltage initial vector set; uniformly fill any current initial vector shorter than the longest current initial vector in the first current initial vector set with zeros, and superimpose the vectors of the first current initial vector set after data expansion to obtain the corresponding current feature vector; uniformly fill any voltage initial vector shorter than the longest voltage initial vector in the first voltage initial vector set with zeros, and superimpose the vectors of the first voltage initial vector set after data expansion to obtain the corresponding voltage feature vector; The first two-dimensional current data is any one of the plurality of two-dimensional current data, and the first two-dimensional voltage data is any one of the plurality of two-dimensional voltage data.

3. The training device according to claim 1, characterized in that: The multi-head cross-attention module is used to perform multi-layer feature collection on the output vector of the dual-channel multi-layer perceptron module to obtain feature data at different scales; and perform multi-head cross-attention fusion processing on the feature data at different scales to obtain the current attention feature data and the voltage attention feature data.

4. The training device according to claim 1, characterized in that: The training device also includes: A deployment module is used to deploy an initial network model, and utilize a knowledge distillation algorithm and a model relearning method to collect new current time series data and new voltage time series data to update the training samples. The parameters of the initial network model are trained and updated through the updated training samples so that the recognition rate of the initial network model reaches a preset accuracy rate. The initial network model is a simplified model of the power fault identification network model.

5. An application device of a power fault identification network model, characterized in that: include: A wave recording module, used to obtain the current time series data and voltage time series data to be identified; A model processing module, used for inputting the current time series data and voltage time series data to be identified into the pre-trained power fault identification network model for processing to obtain a corresponding identification result; Wherein, the power fault identification network model is a model trained by using the training device described in any one of claims 1-4.

6. A training method for a power fault identification network model, characterized in that: include: Convert multiple one-dimensional current time series data in the training sample into multiple two-dimensional current data, and convert multiple one-dimensional voltage time series data in the training sample into multiple two-dimensional voltage data; Performing feature extraction on the plurality of two-dimensional current data to obtain a plurality of two-dimensional current feature vectors, and performing feature extraction on the plurality of two-dimensional voltage data to obtain a plurality of two-dimensional voltage feature vectors; Using a dual-channel multi-layer perceptron module, a plurality of current feature vectors are subjected to feature fusion in a horizontal dimension and a vertical dimension, and a plurality of voltage feature vectors are subjected to feature fusion in a horizontal dimension and a vertical dimension, wherein the horizontal dimension is used for data features of continuous time, and the vertical dimension is used for data features across time; The output vectors of the dual-channel multi-layer perceptron module are respectively subjected to multi-head cross-attention fusion processing to obtain current attention feature data and voltage attention feature data, self-attention fusion is performed on the joint feature data of the current attention feature data and the voltage attention feature data, and the power fault identification network model is trained based on the output feature quantity obtained by the self-attention processing.

7. The training method according to claim 6, characterized in that: The step of extracting features from the plurality of two-dimensional current data to obtain a plurality of two-dimensional current feature vectors, and extracting features from the plurality of two-dimensional voltage data to obtain a plurality of two-dimensional voltage feature vectors, comprises: Using multiple convolution kernels of different scales to perform one-dimensional convolution operations on the first current two-dimensional data to obtain a first current initial vector set, and using the multiple convolution kernels of different scales to perform one-dimensional convolution operations on the first voltage two-dimensional data to obtain a first voltage initial vector set; Performing uniform zero padding on any current initial vector shorter than the longest current initial vector in the first current initial vector set, and superimposing the vectors of the first current initial vector set after data expansion to obtain the corresponding current characteristic vector; performing uniform zero padding on any voltage initial vector shorter than the longest voltage initial vector in the first voltage initial vector set, and superimposing the vectors of the first voltage initial vector set after data expansion to obtain the corresponding voltage characteristic vector; The first two-dimensional current data is any one of the plurality of two-dimensional current data, and the first two-dimensional voltage data is any one of the plurality of two-dimensional voltage data.

8. An application method of a power fault identification network model, characterized in that: include: Acquiring current time series data and voltage time series data to be identified; Inputting the current time series data and voltage time series data to be identified into the pre-trained power fault identification network model for processing to obtain corresponding identification results; Wherein, the power fault identification network model is a network model trained by applying the training method described in any one of claims 6-7.

9. An electronic device, characterized in that: include: A memory for storing instructions; A processor is used to execute instructions stored in the memory so that the electronic device executes the training method described in any one of claims 6-7 or the application method described in claim 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, it is used to implement the training method according to any one of claims 6-7 or the application method according to claim 8.

Images