Dual-channel time-frequency fusion driven mamba's main power distribution network detection classification method

Abstract

The application discloses a kind of main distribution network detection classification methods of double-channel time-frequency fusion driven Mamba, comprising: S1.acquires the fault voltage current signal data under different main distribution network topology, and constructs training set;S2. Model based on double-channel time-frequency fusion driven Mamba is constructed, for calculating and obtaining the fault probability of the kth fault and fault category label;S3. the training set is trained to model using back propagation and gradient descent method, and the trained main grid fault detection network is obtained, for mapping corresponding fault classification label after inputting fault data set;S4. using the model after training, input main distribution network fault voltage current signal data, execute detection classification operation.The application improves the detection robustness in complex noise environment significantly through multi-scale feature reconstruction and cross-modal fusion, realizes the accurate capture to fault starting time, duration and propagation path.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of distribution network fault detection, in particular to a main distribution network detection and classification method based on double-channel time-frequency fusion driving Mamba, equipment and storage medium, which is suitable for high-precision detection and classification of power grid faults in complex noise environment. BACKGROUND

[0002] Main distribution network fault detection and classification is a key link for safe operation of power systems. Traditional methods mainly rely on time-frequency analysis and model-driven technology based on state estimation to identify faults by analyzing time-domain or frequency-domain characteristics of voltage and current signals. However, with the expansion of power grid scale and the access of distributed energy, the topology of main distribution network is becoming increasingly complex, and fault signals exhibit multi-scale and non-stationary characteristics, and are easily disturbed by noise. At the same time, existing technologies lack the ability to capture the starting time, duration and propagation path of faults in complex scenarios, resulting in limited detection accuracy and robustness.

[0003] Current conventional fault detection methods usually use single time-frequency analysis technology to extract local features, or rely on machine learning models for classification. However, such methods have obvious defects:

[0004] (1) Insufficient multi-scale feature fusion: traditional methods are difficult to capture both high-frequency mutations and low-frequency steady-state characteristics of signals, resulting in loss of key information;

[0005] (2) High noise sensitivity: in strong interference environment, signal features are easily submerged, leading to misjudgment or omission;

[0006] (3) Poor adaptability to dynamic topology: the time-varying nature of fault propagation path in complex topology is difficult to model, and traditional models are difficult to accurately locate fault position and path evolution;

[0007] In addition, even if existing optimization methods use improved wavelet basis selection or combine multi-scale transformation or introduce deep learning models to solve the above problems, but in actual use process still cannot effectively solve the problem of multi-scale time-frequency feature cross-modal sufficient fusion, the generalization ability of model in complex noise scene and the dynamic topology modeling ability of model are still insufficient, the noise suppression and feature reconstruction efficiency are low.

[0008] In addition, for the specific scenario of main power distribution network fault detection, the prior art also discloses a Pan-Mamba and a Fusion-Mamba and the like similar model architecture, wherein, for example, the Pan-Mamba model, the core target of which is image pan-sharpening, focuses on improving the spatial resolution and spectral consistency of the image, and the processing object is the static or slowly changing image spatial-spectral features, which does not involve time-frequency signal analysis or dynamic fault detection in the actual use process in the field of main power distribution network fault detection; or for example, the Fusion-Mamba, which mainly aims at feature splicing or simple weighted fusion of general multi-modal data (such as visual-text, infrared-visible light, etc.), which does not design a special mechanism for the time-frequency coupling characteristics specific to the power system (such as the coexistence characteristics of high-frequency mutation and low-frequency steady state of the fault signal) in the actual use process in the field of main power distribution network fault detection

[0009] To this end, the application provides a dual-channel time-frequency fusion driven Mamba main power distribution network detection classification method to solve the above technical problems. SUMMARY

[0010] The main purpose of the application is to provide a dual-channel time-frequency fusion driven Mamba main power distribution network detection classification method, propose a dual-channel time-frequency fusion architecture, combine wavelet transform, cross-modal attention enhancement mechanism and Mamba state space model, and significantly improve the detection robustness in a complex noise environment through multi-scale feature reconstruction and cross-modal fusion, realize accurate capture of the fault starting time, duration and propagation path, and provide an efficient and reliable solution for main power distribution network fault diagnosis in a high dynamic and strong interference scene, to solve the technical problems proposed in the background art, and the Mamba model architecture designed by the application is significantly different from the Pan-Mamba, Fusion-Mamba and the like prior art in architecture target, feature processing mechanism and application adaptability, and can solve the specific technical problems in the field of main power distribution network fault detection.

[0011] The application solves the above technical problems by adopting the following technical solutions:

[0012] A dual-channel time-frequency fusion driven Mamba main power distribution network detection classification method, comprising the following steps:

[0013] S1. Collecting fault voltage and current signal data under different main power distribution network topologies, and constructing a training set FD;

[0014] S2. Construct a model based on a dual-channel time-frequency fusion driving Mamba, which adopts a time-domain feature processing channel and a frequency-domain feature processing channel to realize dynamic cooperation of time-frequency features through a cross-modal attention enhancement mechanism, the model comprising a fault feature reconstruction layer, a fault feature fusion layer and a mamba network layer, for calculating and obtaining a fault probability and a fault category label of the kth fault, the fault feature reconstruction layer comprising an attention weight distribution mechanism, a high-low frequency fusion and inverse transformation mechanism, and the fault feature fusion layer comprising a multi-scale time-frequency graph convolution feature extraction mechanism and a cross-modal attention enhancement mechanism;

[0015] S3. The training set FD is trained by using a back propagation and gradient descent method to obtain a trained main distribution network fault detection network, which is used to map the corresponding fault classification label after inputting the fault data set;

[0016] S4. Using the trained model, input the main distribution network fault voltage and current signal data, and perform detection and classification operations.

[0017] Preferably, the construction process of the training set FD in the S1 step comprises:

[0018] S11. Collect fault voltage and current signal data under different main distribution network topologies, the fault voltage and current signal data comprising sampling values of three-phase voltage and three-phase current independent signals at each sampling time, the fault voltage and current being three-phase voltage and three-phase current independent signals, and constructing a fault detection set , represents the kth fault voltage and current, and , represents the kth fault voltage and current at the mth sampling time, wherein , represents the kth fault voltage and current at the mth sampling time, wherein represents the total number of faults; , represents the total sampling time; , represents the total sampling time;

[0019] S12. Construct a label information set of fault voltage and current starting time and duration , represents the kth fault label value, and , , wherein is used to indicate whether a fault occurs at the mth time;

[0020] S13. Randomly shuffle the fault voltage and current data set with labels as the training set FD.

[0021] ​Preferably, the attention weight allocation mechanism included in the fault feature reconstruction layer in step S2 is used to... Input fault three-phase voltage and current signals Obtain the low-frequency feature matrix of the k-th fault at time step t after enhancement. The calculation formula is as follows:

[0022]

[0023]

[0024] in, , Indicates the number of floors. Indicates the total number of floors. It is the kth fault. Layer The low-frequency components of the time step, It is the first Cross-scale projection weights of low-frequency components of the layer. Indicates the first Fault voltage and current, It is the first Layer scale.

[0025] Preferably, the high- and low-frequency fusion and inverse transformation mechanism included in the fault feature reconstruction layer in step S2 is used to base the low-frequency feature matrix. Calculate and obtain the first Feature vector of a fault The calculation formula is as follows:

[0026]

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] in, It's a vector concatenation operation. , Indicates the total sampling time. It is the first The fault is in the first Reconstructing the time-domain signal value at the time step. It is the enhanced version of the first Fault No. time step high frequency component matrix, is the layer wavelet basis function, is the th fault layer high frequency component normalized attention weight, is the th fault layer time step high frequency component, is the Hadamard product, is the layer learnable adaptive weight matrix, is the learnable temperature coefficient, is the th fault layer high frequency component total energy, is the square of the two-norm.

[0033] Preferably, the multi-scale time-frequency map convolution feature extraction mechanism included in the fault feature fusion layer in the S2 step is used to calculate the th fault voltage and current , obtain the th fault layer wavelet coefficient and the th fault layer component amplitude , and the calculation formula is:

[0034]

[0035]

[0036]

[0037] wherein, , denotes the total sampling time, , is the frequency, denotes the highest frequency in the fault signal, is the discrete wavelet transform, is the layer wavelet basis function, is the Daubechies wavelet, is the symmetrical Daubechies wavelet, is the th fault voltage and current signal-to-noise ratio, is the continuous wavelet transform, is the scale parameter range.

[0038] Preferably, the multi-scale time-frequency map convolutional feature extraction mechanism included in the fault feature fusion layer in step S2 is also used to extract features through the first... Fault No. Layer wavelet coefficients Calculate and obtain the first Temporal direction sensitivity characteristics of faults The calculation formula is as follows:

[0039]

[0040]

[0041]

[0042] in, , Indicates the number of floors. Indicates the total number of floors. It is the Gaussian error linear unit activation function. It is the convolutional feature vector of the k-th fault. It is a topology mask dynamic filtering. It is a two-dimensional weight matrix. It is a two-dimensional convolution. It is a left-propagating feature convolution kernel. It is a right-propagating feature convolution kernel. It is a convolution kernel for the transient oscillation characteristics of the busbar.

[0043] Preferably, the cross-modal attention enhancement mechanism included in the fault feature fusion layer in step S2 is used to... Temporal direction sensitivity characteristics of faults Calculate and obtain the first residual gated output of a fault The calculation formula is as follows:

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050] in, It is a layer of regularization. It is the first Fault fusion output, It is a regularization to prevent overfitting. It is the Sigmoid function. It is a learnable gating weight matrix. It is Hadamaji. For the first Input fault three-phase voltage and current signals , It is a normalized exponential activation function. It is the first Fault time-domain query vector, For the first The frequency domain key vector of the fault, For the first The transpose of the frequency domain key vector of the fault. It is a scaling factor. It is the first Fault query projection matrix, For the first Fault key projection matrix, For the first The frequency domain direction sensitivity characteristics of the fault. It is the first Fault No. Layer energy weight vector, It is the Tollitz matrix. , It is an L1 norm.

[0051] Preferably, in step S2, the Mamba network layer calculates fault classification label data based on the output results of the fault feature reconstruction layer and the fault feature fusion layer, specifically including:

[0052] L1. Construct a Mamba model and calculate the output features of the k-th fault;

[0053] L2. Calculate the unnormalized classification score of the k-th fault of type i in the MLP layer of the Mamba model, and use the score result to input into the softmax layer of the Mamba model to obtain the probability that the k-th fault belongs to type i and the fault category label.

[0054] Preferably, the formula for calculating the output feature of the k-th fault in the L1 step is:

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] wherein, represents the output feature of the kth fault at the tth time, represents the matrix splicing operation, , represents the total sampling time, is the state of the kth fault at the tth time, is the skip connection weight matrix, is the feature vector of the kth fault is the fusion matrix of the residual gate output of the kth fault and the kth fault, is the state evolution matrix of the kth fault dynamically adjusted, is the hidden state matrix of the kth fault at the tth time, is the time resolution of the control state update of the kth fault, is a predefined diagonal matrix, and e represents the exponential function with the natural constant as the base number, is the state transition matrix of the kth fault after discretization, is the state control matrix of the kth fault, is the output space projection matrix of the dynamic input mapping of the kth fault, is a learnable weight matrix, is a bias term, is a hidden layer activation function. Preferably, the specific operation process of the L2 step comprises:

[0066] In the MLP layer calculation, the unnormalized classification score of the kth fault of the ith fault is obtained , and the calculation formula is:

[0067]

[0068] ​​​​​​

[0069] wherein, is the weight matrix of the first layer of the MLP, is the weight matrix of the first class of faults, is the weight matrix of the second layer of the MLP, is the weight matrix of the second class of faults, is the bias matrix of the first layer of the MLP, is the bias matrix of the second layer of the MLP, is a Gaussian error linear unit activation function;

[0070] The score is input into a softmax layer to calculate the probability that the kth fault belongs to the ith class of faults and the fault class label , and the calculation formula is:

[0071]

[0072]

[0073] wherein, , is the total number of fault classes, represents a natural exponential function.

[0074] In yet another aspect, the present application also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the above method.

[0075] In still another aspect, the present application also discloses a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the above method.

[0076] From the above technical solution, the present application provides a dual-channel time-frequency fusion driven Mamba main power distribution network detection and classification method. Compared with the prior art, the present application has the following advantages:

[0077] 1. The present application sets up an attention weight distribution mechanism in the fault feature reconstruction layer, which can dynamically allocate the weights of different time steps, highlight the importance of key features in the fault signal, effectively suppress noise interference, improve fault feature extraction accuracy, and realize accurate identification of fault signals in complex noise environment.

[0078] ​2. The application can perform multi-scale fusion of low-frequency features and high-frequency details through the setting of high-low frequency fusion and inverse transformation mechanism in the fault feature reconstruction layer, and reconstruct the original signal features through inverse transformation, thereby fully mining the time-frequency complementary information of the fault signal, enhancing the feature expression ability, and ultimately improving the capture accuracy of the model for the fault starting time and duration.

[0079] 3. The application can further enhance the feature expression richness and improve the model's ability to distinguish fault types under complex topology by setting a multi-scale time-frequency graph convolution feature extraction mechanism in the fault feature fusion layer, decomposing the multi-layer time-frequency coefficients of the signal through wavelet transform, and combining direction-sensitive feature extraction to realize joint modeling of the multi-scale space-time features of the fault signal.

[0080] 4. The application can enhance the synergistic effect between time-frequency domain features through the setting of a cross-modal attention enhancement mechanism in the fault feature fusion layer, and optimize the feature transmission path through residual gate output, thereby effectively alleviating the information loss problem in multi-modal feature fusion, improving the feature fusion efficiency, and enhancing the model's anti-noise robustness in strong interference scenarios.

[0081] 5. The application can capture the time sequence evolution law of the fault signal in the dynamic feature modeling process of the model by combining state space modeling and classification output in the Mamba network layer, and realize accurate prediction of the fault category probability through joint optimization of MLP and Softmax layer, thereby improving the detection and classification accuracy of high-dynamic main and distribution networks.

[0082] 6. The application can deeply couple the Mamba state space model with the time-frequency characteristics of power signals, design a special dual-channel time-frequency fusion architecture, and realize accurate fault detection under complex noise and dynamic topology.

[0083] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the application, nor is it used to limit the scope of the application. Other features of the application will become apparent through the following description. Of course, any product implementing the application does not necessarily need to achieve all the advantages mentioned above. BRIEF DESCRIPTION OF DRAWINGS

[0084] The drawings accompanying the specification of this application form a part thereof, serve to provide further understanding of the application, and together with the description of the exemplary embodiments of the application and its description serve to explain the application. The drawings in the accompanying specification of the application do not constitute an inappropriate limitation on the application. In the drawings:

[0085] Figure 1 is a schematic diagram of the overall process of the application;

[0086] Figure 2 is a schematic diagram of the model framework structure of the application;

[0087] Figure 3 The model system schematic diagram based on the dual-channel time-frequency fusion driving Mamba network of the application;

[0088] Figure 4 The data processing flow schematic diagram of the fault feature reconstruction layer of the application;

[0089] Figure 5 The data processing flow schematic diagram of the fault feature fusion layer of the application;

[0090] Figure 6 The use performance comparison diagram of the application and other fault detection algorithms of the prior art. DETAILED DESCRIPTION

[0091] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. In the case of no conflict, the embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0092] In the embodiments, refer to Figures 1 to 6 .

[0093] As shown in Figure 1 and Figure 2 , the main power distribution network detection and classification method of the dual-channel time-frequency fusion driving Mamba proposed in the embodiments of the application comprises the following steps:

[0094] Step 1: Data preparation

[0095] Collect fault voltage and current signal data under different main power distribution network topologies, and construct a training set FD.

[0096] The fault voltage and current signal data here contains the sampling values of three-phase voltage and three-phase current independent signals at each sampling time. The fault voltage and current are three-phase voltage and three-phase current independent signals. When processing three-phase voltage and three-phase current, the data set involved in the application strictly retains the independence, that is, three-phase voltage (A phase, B phase, C phase) and three-phase current (A phase, B phase, C phase) are six independent signal sequences, rather than pre-fused data. Specifically, in the data collection and construction stage, each fault data in the fault detection set X contains the sampling values of the above six independent signals at each sampling time, that is, , The actual corresponding six independent signal sampling value combination at the t moment (the t moment value of three-phase voltage and the t moment value of three-phase current respectively).

[0097] Specifically, the construction process of the training set FD includes:

[0098] S11. Data acquisition: collect fault voltage / current signal (three-phase signal) data under different main power distribution network topologies, and construct a fault detection set , represents the t th fault voltage and current, , represents the total number of faults, and , represents the t th fault voltage and current at the t th sampling moment; , represents the total sampling time, i.e. each signal contains T sampling points;

[0099] At this time, it should be noted that the signal should include typical fault types and different noise levels (SNR ≥ 40dB and SNR<40dB);

[0100] S12. Label annotation: annotate the fault start time and duration of each signal, and construct a time sequence label information set of fault voltage and current start time and duration , represents the t th fault label value, and , is used to indicate whether a fault occurs at the t th moment;

[0101] S13. Data cleaning and preprocessing: remove outliers, standardize signals to [-1, 1], and proportionally divide the training set and the validation set to obtain a fault voltage and current dataset with labels After that, randomly shuffle the order to serve as the training set FD.

[0102] Step 2: Model construction

[0103] A model based on a dual-channel time-frequency fusion driven Mamba is constructed, which specifically includes a dual-channel time-frequency fusion architecture and a mamba network layer, for calculating and obtaining the fault probability and fault category label of the kth fault. Specifically:

[0104] (1) Referring to Figure 3 , the model system of the present application adopts a time domain feature processing channel and a frequency domain feature processing channel to realize dynamic cooperation of time-frequency features through a cross-modal attention enhancement mechanism to construct a dual-channel time-frequency fusion architecture.

[0105] It should be noted at this time that the dual-channel architecture of the present application specifically refers to "time-frequency" parallel processing channels: one is a time-domain feature processing channel that enhances low-frequency steady-state features (such as voltage and current trends in the fault background) through the attention weight distribution mechanism of the fault feature reconstruction layer, reconstructs the time-domain signal through the high-low frequency fusion and inverse transform mechanism, and retains the timing details of the fault starting time; the second is a frequency-domain feature processing channel that relies on the multi-scale time-frequency graph convolution feature extraction mechanism of the fault feature fusion layer, captures high-frequency mutation features (such as harmonic components at the fault moment) through discrete wavelet transform (DWT) and continuous wavelet transform (CWT), and realizes dynamic coordination of time-frequency features through cross-modal attention enhancement mechanism. The two-channel features finally converge into the Mamba network layer, which adapts to the timing evolution of the fault signal through the dynamic state evolution matrix, forming a complete time-frequency feature processing link of "separation extraction-targeted enhancement-deep fusion", which is specifically designed to solve the problem of effective utilization of features under the condition of time-frequency coupling and strong noise interference of power grid fault signals.

[0106] The specific architecture content includes:

[0107] A. Reference Figure 4 The fault feature reconstruction layer includes an attention weight distribution mechanism, a high-low frequency fusion and inverse transform mechanism:

[0108] (1) Perform multi-scale wavelet decomposition (number of layers S, recommended 3-5 layers) on the input signal to separate high-frequency components and low-frequency components , and at the same time, the energy weight of each layer of high-frequency components is calculated through the attention mechanism to dynamically fuse high and low frequency features, and the feature vector of the fault , where the calculation steps are as follows:

[0109] Use the attention weight distribution mechanism to input the fault three-phase voltage and current signal , obtain the enhanced low-frequency feature matrix of the kth fault at the tth time step , and the calculation formula is:

[0110]

[0111]

[0112] wherein, , represents the number of layers, represents the total number of layers, is the low-frequency component of the kth fault at the tth time step of the lth layer , and is the low-frequency component of the kth fault at the tth time step of the lth layer , and ​Cross-scale projection weight of layer low-frequency component, representing the first the fault voltage and current, is the first layer scale;

[0113] At this time, the importance of key features in the fault signal can be highlighted by dynamically allocating the weights of different time steps, thereby effectively suppressing noise interference, improving fault feature extraction accuracy, and achieving accurate identification of fault signals in complex noise environments.

[0114] Then, the high-low frequency fusion and inverse transformation mechanism is used to calculate the feature vector of the first fault based on the low-frequency feature matrix , and the calculation formula is as follows:

[0115]

[0116]

[0117]

[0118]

[0119]

[0120]

[0121] wherein, is the vector splicing operation, , represents the total sampling time, is the reconstructed time-domain signal value of the first fault at the first time step, is the enhanced high-frequency component matrix of the first fault at the first time step, is the first layer wavelet basis function, is the normalized attention weight of the first layer high-frequency component of the first fault, is the high-frequency component of the first layer at the first time step of the first fault, is the Hadamard product, is the first layer learnable adaptive weight matrix, is the learnable temperature coefficient, is the first​ layer the total energy of the high-frequency component of the fault, is the square of the two-norm;

[0122] At this time, low-frequency features can be fused with high-frequency details in multiple scales, and the original signal features can be reconstructed through inverse transformation, thereby fully mining the time-frequency complementary information of the fault signal, enhancing the feature expression ability, and ultimately improving the capture accuracy of the fault starting time and duration.

[0123] B. Reference Figure 5 The fault feature fusion layer includes a multi-scale time-frequency graph convolution feature extraction mechanism and a cross-modal attention enhancement mechanism:

[0124] Discrete / continuous wavelet transform (DWT / CWT) is performed on the reconstructed signal to extract time-domain direction-sensitive features (left / right propagation, bus oscillation) and frequency energy distribution, and the time-frequency features are fused through a cross-modal attention mechanism to generate a residual gate output The calculation steps are as follows:

[0125] Using the multi-scale time-frequency graph convolution feature extraction mechanism, the first fault voltage and current , the first fault layer wavelet coefficients and the first fault layer component amplitude are calculated and obtained, and the calculation formula is:

[0126]

[0127]

[0128]

[0129] wherein, , denotes the total sampling time, , is the frequency, denotes the highest frequency in the fault signal, is the discrete wavelet transform, is the first layer wavelet basis function, is the Daubechies wavelet, is the symmetrical Daubechies wavelet, is the signal-to-noise ratio of the first fault voltage and current , and is the continuous wavelet transform, ​​is a scale parameter range;

[0130] Then the first The fault of the first Layer wavelet coefficient , the time domain direction sensitive feature of the first Fault is calculated and obtained , the calculation formula is:

[0131]

[0132]

[0133]

[0134] Among them, , Indicates the number of layers, Indicates the total number of layers, Is a Gaussian error linear unit activation function, Is the convolution feature vector of the first Fault, Is a topological mask dynamic filtering, Is a two-dimensional weight matrix, Is a two-dimensional convolution, Is a left propagation feature convolution kernel, Is a right propagation feature convolution kernel, Is the bus transient oscillation feature convolution kernel;

[0135] At this time, the multi-layer time-frequency coefficients of the signal are decomposed by wavelet transform, and combined with the direction sensitive feature extraction, the joint modeling of the multi-scale space-time features of the fault signal can be realized, so as to further enhance the feature expression richness and improve the model's ability to distinguish the fault type under complex topology;

[0136] Finally, the cross-modal attention enhancement mechanism is used, and the time domain direction sensitive feature of the first Fault is calculated and obtained The residual gate output of the first Fault is calculated and obtained , the calculation formula is:

[0137]

[0138]

[0139]

[0140]

[0141]

[0142]

[0143] wherein, is layer regularization, is the th fault fusion output, is regularization to prevent overfitting, is a Sigmoid function, is a learnable gating weight matrix, is a Hadamard product, is the th input fault three-phase voltage current signal , is a normalized exponential activation function, is the th fault time domain query vector, is the th fault frequency domain key vector, is the th fault frequency domain key vector transpose vector, is a scaling factor, is the th fault query projection matrix, is the th fault key projection matrix, is the th fault frequency domain direction sensitive feature, is the th fault layer energy weight vector, is a Toeplitz matrix, , is an L1 norm;

[0144] At this time, the synergistic effect between the time-frequency domain features can be enhanced, and the feature transmission path is optimized through the residual gating output, thereby effectively alleviating the information loss problem in multi-modal feature fusion, improving the feature fusion efficiency, and enhancing the anti-noise robustness of the model in a strong interference scene.

[0145] C. Constructing a Mamba state space model in the mamba network layer for input fusion features , dynamic modeling is performed through a state transition matrix A, a control matrix B, and an output matrix C to capture the fault propagation path and time step. At this time, the mamba network layer calculates the fault classification label data based on the fault feature reconstruction layer output result and the fault feature fusion layer output result, specifically including:

[0146] L1. Construct a Mamba model and calculate the output feature of the k-th fault, where the formula for calculating the output feature of the k-th fault is:

[0147]

[0148]

[0149]

[0150]

[0151]

[0152]

[0153]

[0154]

[0155]

[0156]

[0157] in, Indicates the first Fault No. Output characteristics at time step This represents a matrix concatenation operation. , Indicates the total sampling time. This is the state of the k-th fault at time t. It is the skip connection weight matrix. It is the first Feature vector of a fault and the residual gated output of a fault The fusion matrix, and This is the state evolution matrix for the k-th fault dynamic adjustment. It is the kth fault. The hidden state matrix at time step 1. This is the time resolution of the k-th fault control state update. It is a predefined diagonal matrix, where e represents an exponential function with base to the natural constant. It is the state transition matrix after the k-th fault is discretized. This is the state control matrix for the k-th fault. It is the output space projection matrix of the k-th fault dynamic input mapping. It is a learnable weight matrix. It is a bias term. It is the activation function of the hidden layer;

[0158] L2. Calculate the unnormalized classification score of the k-th fault of type i in the MLP layer of the Mamba model, and input the score into the softmax layer of the Mamba model to obtain the probability that the k-th fault belongs to type i and the fault category label. The specific operation process at this time includes:

[0159] The unnormalized classification score of the k-th fault and the i-th type of fault is calculated in the MLP layer. The calculation formula is as follows:

[0160]

[0161] in, It is the first layer of MLP. Weight matrix of fault types, It is the second layer of MLP. Fault weight matrix, It is the bias matrix of the first layer of the MLP. It is the weight bias matrix of the second layer of the MLP. It is the activation function of the Gaussian error linear unit;

[0162] Score The input is fed into a softmax layer to calculate the probability that the k-th fault belongs to the i-th type of fault. and fault category labels ;

[0163] At this point, the temporal evolution of fault signals can be captured during the dynamic feature modeling process, and the probability of fault categories can be accurately predicted through joint optimization of MLP and Softmax layers, thereby improving the detection and classification accuracy of highly dynamic main and distribution networks.

[0164] Step 3: Model Training

[0165] The training set FD uses backpropagation and gradient descent to train the model. Training stops when the maximum number of training epochs is reached or the loss function reaches its minimum, resulting in a trained main and distribution network fault detection network. This network is then used to map the corresponding fault classification labels to the input fault data set.

[0166] (1) Parameter configuration

[0167] Learning rate: (Adam optimizer).

[0168] Temperature coefficient : 0.1-0.5 (to control the concentration of attention weight).

[0169] Loss function: Cross-entropy loss (multi-classification scenario).

[0170] (2) Training process

[0171] The training set is input into the model in batches, and the loss of the predicted value and the label is calculated.

[0172] Optimize model parameters (weight matrix W, bias term b, etc.) through backpropagation.

[0173] Verify the performance of the model on the validation set after each round of training, and stop training when the loss converges or reaches the maximum number of rounds.

[0174] Step 4: Fault detection and classification

[0175] Use the trained model to input the main power distribution network fault voltage and current signal data, and perform detection and classification operations, including:

[0176] (1) Real-time signal input includes: standardization and wavelet decomposition of real-time collected voltage / current signals, generating input features in the same format as training data.

[0177] (2) Inference process includes: outputting fault category probability distribution through the trained model , then taking the class with the maximum probability as the final result , the calculation formula is:

[0178]

[0179]

[0180] where, , is the total number of fault categories, denotes the natural exponential function.

[0181] (3) Result analysis: output includes fault type (such as A-phase ground, BC-phase short circuit), starting time and duration, at this time, combined with power grid topology information, the fault propagation path (such as bus number, line direction) can be located.

[0182] Further, it needs to be pointed out that the application scenario of this method is suitable for:

[0183] (a) Real-time monitoring: high dynamic, strong interference main power distribution network environment.

[0184] (b) Fault warning: transient fault analysis of new energy access to power grid.

[0185] (c) Accident backtracking: verify model accuracy through historical data, optimize power grid protection strategy.

[0186] In addition, it should be noted that the method also needs to ensure that the training data covers multiple noise scenarios and complex power grid structures during use to avoid model overfitting and ensure the quality of the data obtained after the method is used.

[0187] In actual use, the method uses GPU accelerated inference (such as NVIDIA A100) to ensure real-time performance, and at the same time, the model is fine-tuned with new data periodically (such as every quarter) to perform model updates to adapt to changes in the power grid operating environment. Compared with the prior art, the performance of the fault detection of the present application is as follows: Figure 6 As shown in the table. In software failure, sensor failure, circuit failure and mechanical failure and other types of detection, the algorithm based on dual-channel time-frequency fusion driving Mamba has outstanding accuracy, reaching 96.8%, higher than the traditional SVM of 95.6% and the traditional transformer of 95.4%. This shows that the patented algorithm can more accurately identify various types of faults through multi-scale feature reconstruction and cross-modal fusion, has stronger detection robustness in complex noise environments, effectively improves the capture accuracy of the fault starting time, duration and propagation path, and thus realizes more efficient and reliable main distribution network fault diagnosis, and has obvious performance superiority in the field of fault detection.

[0188] In summary, through the above steps, the present application can significantly improve the detection robustness in complex noise environments through multi-scale feature reconstruction and cross-modal fusion, accurately capture the fault starting time, duration and propagation path, and facilitate users to quickly realize main distribution network fault detection and classification based on dual-channel time-frequency fusion driving Mamba.

[0189] In addition, in further embodiments, the present application has further differences in actual application compared to similar model architectures in the prior art:

[0190] The present application constructs a three-level architecture of "fault feature reconstruction layer + fault feature fusion layer + Mamba network layer", which has special optimization for power system characteristics compared to existing Mamba variants. The present application is designed specifically for main distribution network fault detection, and the processing object is non-stationary, strong noise voltage and current time-frequency signals. The core goal is to accurately capture the fault starting time, duration and propagation path, and to solve the problems of time-frequency feature cross-scale fusion, noise suppression and dynamic topology adaptability, which are completely different from the application scenarios and goals of the above-mentioned technologies.

[0191] The architecture of the present application deeply combines the physical characteristics of main distribution network fault signals:

[0192] (a) Time-frequency coupling: By combining discrete wavelet transform (DWT) and continuous wavelet transform (CWT) to extract time domain coefficients and frequency domain amplitude, the problem of single transform unable to consider both sudden change and steady state characteristics is solved;

[0193] (b) Noise robustness: By high-frequency component energy weight and residual gate (Dropout + Sigmoid), the feature distortion under strong noise (SNR < 40 dB) is suppressed;

[0194] (c) Topological dynamics: Topological mask is introduced to filter redundant features under different power grid topologies, solving the problem of poor adaptability of traditional models to complex topologies.

[0195] However, the conventional similar model architecture of the prior art is not designed for the above characteristics of power signals, and cannot be directly applied to the power grid fault detection scene.

[0196] In addition, for the core technical problems (insufficient multi-scale feature fusion, high noise sensitivity, and poor dynamic topology adaptability) in the field of main distribution network fault detection, the above architecture innovation is used to solve the problems:

[0197] (1) Solve the problem of "insufficient multi-scale feature fusion"

[0198] The prior art (such as single time-frequency analysis and simple machine learning model) cannot simultaneously capture high-frequency sudden change (fault initiation) and low-frequency steady state (fault duration) features. The present application realizes cross-scale complementarity of time-frequency features through high-low frequency fusion mechanism of fault feature reconstruction layer and multi-scale convolution of fault feature fusion layer, significantly improving the completeness of feature expression.

[0199] (2) Solve the problem of "high noise sensitivity"

[0200] In a strong noise environment (such as SNR < 40 dB), traditional methods are easily disturbed, leading to misjudgment. The present application uses: a. Attention weight distribution to enhance key feature weight; b. High-frequency component energy normalization to suppress noise components; c. Cross-modal attention to strengthen time-frequency feature consistency, so that the model still maintains high detection accuracy (3%-5% higher than the prior art) in low SNR scenarios.

[0201] (3) Solve the problem of "poor dynamic topology adaptability"

[0202] Under complex power grid topology, the fault propagation path is time-varying, and traditional models are difficult to model. The present application uses: a. Direction-sensitive convolution kernel to capture fault propagation direction; b. Topological mask to dynamically filter topology-independent features; c. Mamba layer to dynamically update state to adapt to path evolution, realizing accurate tracking of fault path under complex topology.

[0203] In summary, the application can deeply couple the Mamba state space model with the time-frequency characteristics of the power signal, design a special dual-channel time-frequency fusion architecture, and realize accurate fault detection under complex noise and dynamic topology. Specifically embodied in:

[0204] (1) The fault feature reconstruction layer is created, which solves the problem of time-frequency feature loss through high-low frequency separation, attention enhancement and inverse transform reconstruction;

[0205] (2) The multi-scale time-frequency graph convolution and cross-modal attention fusion mechanism are innovated, which adapts to the time-space evolution characteristics of the power grid fault;

[0206] (3) The dynamic adaptive Mamba network layer is constructed, which dynamically adjusts the model state combined with the power grid topology, and improves the adaptability in complex scenarios.

[0207] In another aspect, the application also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the above method.

[0208] In another aspect, the application also discloses a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the above method.

[0209] In another embodiment provided by the application, a computer program product containing instructions is also provided, which makes a computer execute the power grid detection and classification method of the dual-channel time-frequency fusion driven Mamba in any of the above embodiments when running on the computer.

[0210] It can be understood that the system provided by the embodiments of the application corresponds to the method provided by the embodiments of the application, and the explanation, examples and beneficial effects of the related content can refer to the corresponding parts in the above method.

[0211] The application also provides an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete mutual communication through the communication bus,

[0212] The memory is used to store a computer program;

[0213] The processor is used to execute the program stored on the memory to realize the power grid detection and classification method of the dual-channel time-frequency fusion driven Mamba.

[0214] The communication bus mentioned in the above electronic device can be a peripheral component interconnect standard bus or an extended industry standard architecture bus. The communication bus can be divided into an address bus, a data bus and a control bus.

[0215] The communication interface is configured to communicate between the electronic device and other devices.

[0216] The memory can include a random access memory, and can also include a non-volatile memory, such as at least one disk memory. Optionally, the memory can also be at least one storage device located away from the aforementioned processor.

[0217] The aforementioned processor can be a general-purpose processor, including a central processing unit, a network processor, etc. The processor can also be a digital signal processor, an application-specific integrated circuit, a field programmable gate array, or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0218] It should also be noted that the electronic device also includes a terminal device, which can also be referred to as a terminal, a user equipment, a mobile station, a mobile terminal, etc. The terminal device can be a mobile phone, a smart television, a wearable device, a tablet computer, a computer with wireless transceiver function, a virtual reality terminal device, an augmented reality terminal device, a wireless terminal in industrial control, a wireless terminal in unmanned driving, a wireless terminal in remote surgery, a wireless terminal in smart power grid, a wireless terminal in transportation safety, a wireless terminal in smart city, a wireless terminal in smart home, etc. The embodiments of the present application do not limit the specific technology and specific device form of the terminal device.

[0219] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (such as floppy disk, hard disk, magnetic tape), optical media, or semiconductor media (such as solid state disk), etc.

[0220] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0221] In addition, it should be noted that if the embodiments of the present application involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative position relationship, movement condition, etc. between components in a certain posture, and if the certain posture changes, the directional indications also change accordingly.

[0222] In addition, if the embodiments of the present application involve descriptions of "first", "second", etc., the descriptions of "first", "second", etc. are only for description purposes and cannot be understood as indicating or implying the relative importance of the technical features or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include at least one of the features. In addition, the meaning of "and / or" appearing throughout the text includes three parallel solutions. Taking "A and / or B" as an example, it includes A solution, or B solution, or A and B solutions. In addition, in the embodiments of the present application, "a plurality of" means two or more. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the realization of ordinary skilled in the art, when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, nor is it within the protection scope required by the present application.

Claims

1. A dual-channel time-frequency fusion driven Mamba main distribution network detection and classification method, characterized in that, Includes the following steps: S1. Collect fault voltage and current signal data under different main distribution network topologies and construct a training set FD; S2. Construct a model based on dual-channel time-frequency fusion driven Mamba. The model uses a cross-modal attention enhancement mechanism to achieve dynamic collaboration of time-frequency features through a time-domain feature processing channel and a frequency-domain feature processing channel. The model includes a fault feature reconstruction layer, a fault feature fusion layer, and a Mamba network layer, which are used to calculate and obtain the fault probability and fault category label of the k-th fault. The fault feature reconstruction layer includes an attention weight allocation mechanism, a high-low frequency fusion and inverse transformation mechanism, and the fault feature fusion layer includes a multi-scale time-frequency map convolutional feature extraction mechanism and a cross-modal attention enhancement mechanism. S3. The training set FD uses backpropagation and gradient descent to train the model to obtain a trained main distribution network fault detection network, which is used to map the corresponding fault classification label after inputting the fault data set; S4. Using the trained model, input the fault voltage and current signal data of the main distribution network and perform detection and classification operations; The dual-channel architecture includes: The time-domain feature processing channel separates high-frequency and low-frequency components by performing multi-scale wavelet decomposition on the input signal. It calculates the energy weight of the high-frequency components at each scale through an attention mechanism, performs multi-scale fusion of the high-frequency components based on the energy weight, and performs multi-scale fusion of the low-frequency components based on the cross-scale projection weight of the low-frequency components. Then, it performs inverse transform on the fused high-frequency and low-frequency components respectively and then fuses them to reconstruct the original signal features. The frequency domain feature processing channel extracts time-domain direction-sensitive features after performing discrete wavelet transform on the input signal, obtains the frequency domain energy distribution through continuous wavelet transform, and fuses time-frequency features through a cross-modal attention mechanism to generate residual gated output. The dual-channel feature is ultimately incorporated into the Mamba network layer.

2. The main distribution network detection and classification method using dual-channel time-frequency fusion-driven Mamba as described in claim 1, characterized in that, The process of constructing the training set FD in step S1 includes: S11. Collect fault voltage and current signal data under different main distribution network topologies. The fault voltage and current signal data includes the sampled values ​​of three-phase voltage and three-phase current independent signals at each sampling time. The fault voltage and current are three-phase voltage and three-phase current independent signals. Construct a fault detection set. , Indicates the first Fault voltage and current, and , Indicates the first Fault voltage and current, number Sampling time, among which , Indicates the total number of faults; , Indicates the total sampling time; S12. Construct a set of tag information for the start time and duration of fault voltage and current. , Indicates the first The fault label value, and ,in Used to indicate in the Whether a malfunction occurs at any given time; S13. Generate a labeled fault voltage and current dataset. The randomized order is used as the training set FD.

3. The main distribution network detection and classification method using dual-channel time-frequency fusion-driven Mamba as described in claim 1, characterized in that... The attention weight allocation mechanism of the fault feature reconstruction layer in step S2 is used to... Input fault three-phase voltage and current signals Obtain the low-frequency feature matrix of the k-th fault at time step t after enhancement. The calculation formula is as follows: in, , Indicates the number of floors. Indicates the total number of floors. It is the kth fault. Layer The low-frequency components of the time step, It is the first Cross-scale projection weights of low-frequency components of the layer. Indicates the first Fault voltage and current, It is the first Layer scale.

4. The main distribution network detection and classification method using dual-channel time-frequency fusion-driven Mamba as described in claim 3, characterized in that, The high- and low-frequency fusion and inverse transformation mechanism of the fault feature reconstruction layer in step S2 is used to base the low-frequency feature matrix. Calculate and obtain the first Feature vector of a fault The calculation formula is as follows: in, It's a vector concatenation operation. , Indicates the total sampling time. It is the first The fault is in the first Reconstructing the time-domain signal value at the time step. It is the enhanced version of the first Fault No. Time step high frequency component matrix, It is the first Layer wavelet basis functions, It is the first Fault No. Normalized attention weights for high-frequency components of the layer. It is the first Fault No. Layer High-frequency components of the time step, It is Hadamaji. It is the first The layer can learn an adaptive weight matrix. It is a learnable temperature coefficient. It is the first Layer The total energy of the high-frequency components of the fault. It is the square of the L2 norm.

5. The main distribution network detection and classification method using dual-channel time-frequency fusion-driven Mamba as described in claim 1, characterized in that, The multi-scale time-frequency graph convolutional feature extraction mechanism in the fault feature fusion layer of step S2 is used to extract features through the first... Fault voltage and current Calculate and obtain the first Fault No. Layer wavelet coefficients and the Fault No. Layer component amplitude The calculation formula is as follows: in, , Indicates the total sampling time. , For frequency, This indicates the highest frequency in the fault signal. For discrete wavelet transform, It is the first Layer wavelet basis functions, For the Dobercy wavelet, It is a symmetrical Dobercy wavelet. It is the first Fault voltage and current The signal-to-noise ratio, For continuous wavelet transform, It refers to the range of scale parameters.

6. The main distribution network detection and classification method using dual-channel time-frequency fusion-driven Mamba as described in claim 5, characterized in that, The multi-scale time-frequency graph convolutional feature extraction mechanism of the fault feature fusion layer in step S2 is also used to extract features through the first... Fault No. Layer wavelet coefficients Calculate and obtain the first Temporal direction sensitivity characteristics of faults The calculation formula is as follows: in, , Indicates the number of floors. Indicates the total number of floors. It is the Gaussian error linear unit activation function. It is the first The convolutional feature vector of the fault, It is a topological mask SSS code dynamic filtering. It is a two-dimensional weight matrix. It is a two-dimensional convolution. It is a left-propagating feature convolution kernel. It is a right-propagating feature convolution kernel. It is a convolution kernel for the transient oscillation characteristics of the busbar.

7. The main distribution network detection and classification method using dual-channel time-frequency fusion-driven Mamba as described in claim 6, characterized in that, The cross-modal attention enhancement mechanism of the fault feature fusion layer in step S2 is used to enhance the attention of the fault feature fusion layer through the first... Temporal direction sensitivity characteristics of faults Calculate and obtain the first residual gated output of a fault The calculation formula is as follows: in, It is a layer of regularization. It is the first Fault fusion output, It is a regularization to prevent overfitting. It is the Sigmoid function. It is a learnable gating weight matrix. It is Hadamaji. For the first Input fault three-phase voltage and current signals , It is a normalized exponential activation function. It is the first Fault time-domain query vector, For the first The frequency domain key vector of the fault, For the first The transpose of the frequency domain key vector of the fault. It is a scaling factor. It is the first Fault query projection matrix, For the first Fault key projection matrix, For the first The frequency domain direction sensitivity characteristics of the fault. It is the first Fault No. Layer energy weight vector, It is the Tollitz matrix. It is an L1 norm.

8. The main distribution network detection and classification method using dual-channel time-frequency fusion-driven Mamba as described in claim 1, characterized in that, In step S2, the Mamba network layer calculates fault classification label data based on the output results of the fault feature reconstruction layer and the fault feature fusion layer, specifically including: L1. Construct a Mamba model and calculate the output features of the k-th fault; L2. Calculate the unnormalized classification score of the k-th fault of type i in the MLP layer of the Mamba model, and use the score result to input into the softmax layer of the Mamba model to obtain the probability that the k-th fault belongs to type i and the fault category label.

9. The main distribution network detection and classification method using dual-channel time-frequency fusion-driven Mamba as described in claim 8, characterized in that, The formula for calculating the output feature of the k-th fault in the L1 step is: in, Indicates the first Fault No. Output characteristics at time step This represents a matrix concatenation operation. , Indicates the total sampling time. This is the state of the k-th fault at time t. It is the skip connection weight matrix. It is the first Feature vector of a fault and the residual gated output of a fault The fusion matrix, and This is the state evolution matrix for the k-th fault dynamic adjustment. It is the kth fault. The hidden state matrix at time step 1. It is the time resolution of the k-th fault control state update. It is a predefined diagonal matrix, where e represents an exponential function with base to the natural constant. It is the state transition matrix after the k-th fault is discretized. This is the state control matrix for the k-th fault. It is the output space projection matrix of the k-th fault dynamic input mapping. It is a learnable weight matrix. It is a bias term. It is the activation function of the hidden layer.

10. The main distribution network detection and classification method using dual-channel time-frequency fusion-driven Mamba as described in claim 9, characterized in that... The specific operation process of the L2 step includes: The unnormalized classification score of the k-th fault and the i-th type of fault is calculated in the MLP layer. The calculation formula is as follows: in, It is the first layer of MLP. Weight matrix of fault types, It is the second layer of MLP. Fault weight matrix, It is the bias matrix of the first layer of the MLP. It is the weight bias matrix of the second layer of the MLP. It is the activation function of the Gaussian error linear unit; Score The input is fed into a softmax layer to calculate the probability that the k-th fault belongs to the i-th type of fault. and fault category labels The calculation formula is as follows: in, , This represents the total number of fault categories. This represents the natural exponential function.

Images