A Method and Equipment for Fault Selection in Distribution Networks Based on Multimodal Augmentation and Hybrid Deep Learning Models

Abstract

This invention relates to the field of distribution network fault diagnosis technology, specifically to a method and device for fault line selection in distribution networks based on a multimodal enhancement and hybrid deep learning model. The method includes the following steps: acquiring zero-sequence current and voltage transient waveform data of each line in the distribution network at the time of a fault; performing Gram angle field transformation on the one-dimensional waveform data using a penalized inner product to convert it into two-dimensional GASF and GADF images, respectively, and stitching them together to generate a four-channel E-GAF image; inputting the E-GAF image into a Trans-DenseNet model for fault discrimination, and fusing local-global features by embedding a Transformer encoder layer in the DenseNet to output the final fault line selection result. Compared with existing technologies, this invention can adapt to complex faults such as high resistance and arcing, achieving high-precision fault location and providing reliable support for intelligent fault self-healing in new distribution networks.

Description

Technical Field

[0001] This invention relates to the field of distribution network fault diagnosis technology, and in particular to a distribution network fault selection method that integrates multimodal data augmentation, improved Gram angle field transform and hybrid deep learning model. Background Technology

[0002] Existing distribution network fault detection methods based on Gram angle field (GAF), such as Chinese patent applications CN202411303011.7 and CN202411302666.2, although able to improve diagnostic accuracy by utilizing two-dimensional image features, still have the following shortcomings:

[0003] Noise sensitivity: Traditional GAFs have limited ability to suppress high-frequency noise, and are easily interfered with, especially in weak transient signal scenarios.

[0004] Single data source: Relies solely on zero-sequence current data, ignoring the potential value of multi-modal information such as voltage.

[0005] Insufficient model generalization: Fixed normalization methods are difficult to adapt to complex working conditions, and classification models have limited ability to distinguish between multiple types of faults.

[0006] Real-time bottleneck: The high computational complexity of deep models makes it difficult to meet the real-time processing needs of large-scale power distribution networks. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology, such as noise sensitivity, single data source and insufficient model generalization, and to provide a distribution network fault selection method that integrates multimodal data augmentation, improved Gram angle field transformation and hybrid deep learning model.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] As a first aspect of the present invention, a fault location method for distribution networks based on a multimodal enhancement and hybrid deep learning model is provided, comprising the following steps:

[0010] Collect zero-sequence current and voltage transient waveform data of each line in the distribution network when a fault occurs;

[0011] Gram angular field transformation was performed on the zero-sequence current and voltage transient waveform data using a penalized inner product, which was converted into two-dimensional GASF and GADF images of the zero-sequence current and voltage transient waveforms, respectively, and then stitched together to generate a four-channel E-GAF image.

[0012] Four-channel E-GAF images are input into the Trans-DenseNet model for fault identification. The Trans-DenseNet model integrates local and global features by embedding a Transformer encoder layer in DenseNet-121 and outputs the final fault line selection result.

[0013] As a preferred technical feature, the method preprocesses the zero-sequence current and voltage transient waveform time series data of each group of samples, specifically including: normalization scaling using sliding window dynamic normalization, and wavelet thresholding for noise reduction.

[0014] As a preferred technical feature, the Gram angle field transformation based on the inner product with a penalty term is performed on the zero-sequence current and voltage transient waveform data, specifically as follows:

[0015] Polar coordinate transformation is performed on the time series signals of zero-sequence current and voltage transient waveforms to convert the time series data into vectors:

[0016]

[0017] In the formula, Indicates the phase angle; For time series signals after standardization and scaling; The radius representing polar coordinates; For timestamps; The constant coefficients representing the generated space of the regularized polar coordinates;

[0018] Based on the inner product between vectors with a penalty term, the Gram angle of the zero-sequence current and voltage transient waveforms, and the field GASF and Gram angle difference field GADF are calculated respectively:

[0019]

[0020]

[0021] As a preferred technical feature, the current data, when differential-mode interference occurs, makes... When common-mode interference occurs, let ;

[0022] Substitute the two different interferences mentioned above into the following inner product formula to calculate GASF and GADF respectively:

[0023] ,

[0024] .

[0025] In the formula, It represents the difference in angle between adjacent time points.

[0026] As a preferred technical feature, the Trans-DenseNet model uses DenseNet-121 as the backbone network, retains the four dense blocks of the original DenseNet for extracting local texture features of the image; a Transformer encoder layer is inserted between the third and fourth dense blocks to capture global temporal dependencies using a self-attention mechanism; and line, type, and location prediction results are generated through a multi-task output head.

[0027] As a preferred technical feature, the specific workflow of the Trans-DenseNet model is as follows:

[0028] Four-channel E-GAF images are input into the Trans-DenseNet model;

[0029] Local features are extracted from the first three dense blocks of DenseNet;

[0030] The Transformer encoder flattens the feature map output from the third dense block of DenseNet into a sequence, adds positional encoding, and performs multi-head attention computation.

[0031] The Transformer output sequence is reshaped into a feature map and residually connected to the features of subsequent layers of DenseNet, while multi-scale features are preserved through skip connections.

[0032] The features are further refined using the fourth dense block of DenseNet;

[0033] The final feature map is processed by global average pooling and then input into the multi-task output head, where three fully connected layers generate prediction results for the line, type, and location, respectively.

[0034] As a preferred technical feature, the prediction result of the multi-task output head is specifically as follows:

[0035] Faulty circuit classification: The faulty circuit is classified through a fully connected layer, the number of output neurons of the fully connected layer is equal to the number of circuits, and then the faulty probability of each circuit is output through the Softmax activation function.

[0036] Fault type classification: The output is then converted into a probability distribution of high resistance / arc / intermittent faults through a fully connected layer with the number of neurons equal to the number of fault types; the output is then converted into a probability distribution of high resistance / arc / intermittent faults through a Softmax activation function.

[0037] Fault location regression: Predict the fault distance through a fully connected layer and output a normalized fault distance.

[0038] As a preferred technical feature, the Trans-DenseNet model is deployed in a lightweight manner using knowledge distillation technology. Using a teacher-student model architecture, the Trans-DenseNet model is used as the teacher model, and knowledge distillation technology is used to transfer the knowledge of the teacher model to the lightweight student model for deployment.

[0039] As a second aspect of the present invention, a distribution network fault location device based on a multimodal enhancement and hybrid deep learning model is provided, comprising:

[0040] One or more processors;

[0041] Memory, used to store one or more programs;

[0042] When the one or more programs are executed by the one or more processors, the one or more processors implement the distribution network fault selection method based on multimodal enhancement and hybrid deep learning model as described above.

[0043] As a third aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the distribution network fault selection method based on a multimodal enhancement and hybrid deep learning model as described above.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1) This invention employs a multi-modal fusion method, simultaneously performing Gram-angle field transformation (GASF) on both the zero-sequence current and voltage transient waveforms at the time of a fault. The zero-sequence current reflects the amplitude and waveform distortion of the fault current, while the voltage transient waveform captures the voltage drop or resonance characteristics at the moment of the fault, enabling in-depth mining of potential information about distribution network faults. Furthermore, the one-dimensional waveform data is transformed into two-dimensional GASF and GADF images, making the information contained in the data more easily utilized by deep learning.

[0046] 2) This invention proposes a hybrid deep learning model (Trans-DenseNet), embedding a Transformer encoder layer into DenseNet-121 to construct a local-global feature fusion module. The dense blocks of DenseNet can focus on local fault modes (such as current abrupt changes), while the Transformer encoder can capture global waveform periodicity (such as high-frequency oscillations in arc faults). The two are fused through residual connections to improve feature richness. For the feature map obtained by the final local-global feature fusion, a multi-task output head can predict the fault probability, fault type classification, and fault location for each line. This improves the model's adaptability to complex faults such as high resistance and arcing, while significantly improving line selection efficiency in the end-to-end process.

[0047] 3) This invention also utilizes knowledge distillation technology to transfer the knowledge of the hybrid deep learning model Trans-DenseNet to the lightweight MobileNetV3, and implements hardware acceleration based on FPGA, which improves the inference speed by 3 times and reduces the latency to less than 10ms, meeting the real-time diagnosis needs of large-scale power distribution networks. Attached Figure Description

[0048] Figure 1 This is a flowchart of the fault location method for distribution networks based on multimodal enhanced Gram angle field and hybrid deep learning model of the present invention.

[0049] Figure 2 The confusion matrix of the route selection results for traditional Gram angle field (GAF) based distribution network fault detection methods;

[0050] Figure 3 This is the confusion matrix of the fault location result of the distribution network fault location method based on multimodal enhanced Gram angle field and hybrid deep learning model of the present invention. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0052] Example 1

[0053] This invention proposes a fault location method for distribution networks based on multimodal enhanced Gram angle field and a hybrid deep learning model. Multimodal data fusion is employed, simultaneously inputting signals for both zero-sequence current and voltage dimensions to improve fault identification capabilities. Gram angle field transform is used to convert one-dimensional waveform data into two-dimensional GASF and GADF images, which are then stitched together into a four-channel input to preserve multi-dimensional temporal correlation features. A local-global feature fusion module is constructed by embedding a Transformer encoder layer in DenseNet-121 to achieve accurate fault line identification. Knowledge distillation compresses the model into a lightweight MobileNetV3, and hardware acceleration is implemented on FPGA, achieving a single-sample processing latency of ≤8ms, meeting the real-time diagnostic requirements of large-scale distribution networks. Figure 1 As shown, the specific steps of the method of the present invention are as follows:

[0054] Step S1: Collect zero-sequence current and voltage waveform data of each line when a fault occurs;

[0055] Step S2: Perform dynamic normalization and wavelet denoising on the waveform data. Specifically, the zero-sequence current and voltage time series data are dynamically normalized through a sliding window to suppress local noise, and wavelet threshold denoising is used to enhance signal characteristics.

[0056] Step S3: Perform Gram angle field transformation on the one-dimensional zero-sequence current and voltage waveform data to convert them into two-dimensional GASF and GADF images of current and voltage, and stitch them together to generate a four-channel E-GAF image input.

[0057] Step S4: Input the four-channel E-GAF image (GASF+GADF of current / voltage) into the Trans-DenseNet model for fault identification. By embedding the Transformer encoder layer in DenseNet-121, a local-global feature fusion module is constructed to obtain the final fault line selection result.

[0058] In step S3, the one-dimensional time series signal is transformed into a two-dimensional image using Gram angle fields. This method preserves the characteristics of the original signal to the greatest extent possible without losing any features. The theoretical basis of Gram angle fields is to transform one-dimensional data from the Cartesian coordinate system to the polar coordinate system. Then, based on its unique inner product definition, it uses trigonometric functions and methods such as sums and differences to obtain information representing time correlation. This information is then tiled onto the image from the top left to the bottom right, thereby generating two types of images: GASF and GADF.

[0059] For time series signals , here Represents the total number of points in time: For a point in time, .Will The GAF process is as follows:

[0060] Time series signal of each group of samples Local mean is calculated using a sliding window. ) and standard deviation ( Scaling is achieved through dynamic normalization, as shown in equation (1).

[0061] (1)

[0062] In the formula, The length of the sliding window. To prevent the use of tiny constants in division by zero, dynamic normalization is employed compared to fixed scaling of the data. Significantly improved noise resistance across intervals: Local statistics suppress noise interference, avoiding the impact of global extreme values ​​on normalization. For example, if high-frequency noise exists within the window, It amplifies local fluctuations in noise and dynamically adjusts the scaling ratio. Adaptability is significantly improved: window length... It can be adjusted according to signal characteristics (e.g., a short window captures transient changes, while a long window smooths periodic fluctuations). The values ​​are more stable: a small constant is introduced. (like ( ), to prevent the denominator from being zero.

[0063] Formula (2) represents the Gram matrix. The inner product between different vectors can show the vector correlation, while the angle between vectors shows the degree of correlation between the same vectors.

[0064] (2)

[0065] In the formula, It is a Gram matrix; This is for inner product operations.

[0066] Since one-dimensional time series data is not a vector, it is necessary to perform polar coordinate transformation on the time series to convert the time series current data into a vector, as shown in equation (3):

[0067] (3)

[0068] In the formula, For timestamps; The constant coefficients representing the generated space of the regularized polar coordinates; Indicates the phase angle; The radius representing polar coordinates; For standardized and scaled time series signals .

[0069] If the Gram matrix is ​​used to define the inner product, it is difficult to distinguish the Gaussian noise in the output data. Therefore, in order to eliminate the influence of Gaussian white noise, the Gram corner field defines two unique inner product forms with penalty terms, as shown in formulas (4) and (5):

[0070] (4)

[0071] (5)

[0072] Based on the two different definitions of inner product, two different Gram angle fields can be obtained: Gram angle summation fields (GASF) and Gram angle difference fields (GADF). The formulas are shown in (6) and (7):

[0073] (6)

[0074] (7)

[0075] For current data that are close in time, the GAF transform has the ability to suppress signal disturbances. When differential-mode interference occurs: When common-mode interference occurs: Substituting the two different types of interference into equations (8) and (9) respectively, the interference can be automatically removed. The process of GAF generating a two-dimensional image is as follows: Figure 1 As shown.

[0076] (8)

[0077] (9)

[0078] In the formula, This represents the difference in angle between adjacent time points. This difference is used to capture the dynamic changes in the time series and is a key parameter for constructing the Gram angle field matrix.

[0079] This invention employs a multimodal fusion method to expand the data source, and the multimodal data specifically includes:

[0080] Zero-sequence current: reflects the amplitude and waveform distortion of the fault current (such as the high-frequency oscillation of an arc fault).

[0081] Voltage transient waveform: Captures the voltage drop or resonance characteristics at the moment of a fault (such as the low amplitude transient of a high impedance fault).

[0082] Fusion method: Wavelet denoising and dynamic normalization are performed on the current and voltage signals respectively (as shown above). Then, GASF (Equation 6) and GADF (Equation 7) images are generated from the current and voltage signals respectively, forming a four-channel input.

[0083] Voltage signals are sensitive to weak transient faults (such as high-resistance grounding), compensating for the shortcomings of a single current signal.

[0084] In step 4, this invention proposes a hybrid deep learning model (Trans-DenseNet) to achieve accurate identification of faulty lines.

[0085] 1. Trans-DenseNet Model Design:

[0086] The core idea of ​​the hybrid deep learning model is to embed a Transformer encoder layer into DenseNet-121 to build a local-global feature fusion module and achieve end-to-end multi-task learning.

[0087] DenseNet-121 backbone network: retains the four dense blocks of the original DenseNet for extracting local texture features of images (such as abrupt changes and periodic patterns in GAF images).

[0088] Each dense block contains multiple convolutional layers, which enhance gradient flow and feature representation through dense connections (feature reuse).

[0089] Transformer encoder insertion location: Insert a Transformer encoder layer between the third and fourth dense blocks to capture global temporal dependencies using a self-attention mechanism.

[0090] Input processing: Flatten the feature map output by DenseNet (size H×W×C) into a sequence (length H×W, dimension C), and add positional encoding:

[0091] (10)

[0092] In the formula, This represents the augmented sequence input to the Transformer encoder, which incorporates feature information. and location information ; Indicates the flattening operation; This represents the 3D feature map output by the DenseNet backbone network (DenseNet-121), which is usually generated from the first few dense blocks and has a size of H×W×C (height × width × number of channels). This refers to positional encoding, which is used to inject positional information into a sequence.

[0093] Multi-head attention calculation:

[0094] (11)

[0095] In the formula, Obtained by linear transformation of the input sequence. For attention head dimension.

[0096] 2. Local-Global Feature Fusion:

[0097] The Transformer output sequence is reshaped into feature maps and residually concatenated with features from subsequent layers of DenseNet.

[0098] (12)

[0099] In the formula, The fused feature map combines local spatial features (Dense Block) and global temporal features (Transformer). To be The input is then fed into the subsequent Dense Block to further extract higher-level local features; The 3D feature map output by the DenseNet backbone network (DenseNet-121) is usually generated from the first few dense blocks and has a size of H×W×C (height × width × number of channels). To generate the sequence output by the Transformer Reconstructed into a 3D feature map, restored to its original state. The same dimensions H×W×C; The output sequence of the Transformer encoder captures global temporal dependencies through a self-attention mechanism.

[0100] Skip connections preserve multi-scale features and enhance model robustness. The location and function of skip connections.

[0101] Connection module: DenseNet's intermediate layer features (such as the feature map output by the third Dense Block) With global features of the Transformer encoder ( (Reconstructed feature map).

[0102] Dense Block: Continue to extract local spatial features (such as abrupt changes in fault waveforms and periodic textures).

[0103] Reshape(Z_Transformer): Converts the global temporal dependency features (long-range associations modeled by self-attention) output by the Transformer into a 3D tensor with the same dimensions as the DenseNet feature map.

[0104] Jump connection method:

[0105] Residual Addition: The output of the Dense Block is added element-wise to the reconstructed feature map output by the Transformer to form a fused feature map. .

[0106] While preserving local details (DenseNet), a global context (Transformer) is introduced to enhance the model's ability to model complex fault modes (such as high resistance and arcing).

[0107] Fusion advantages: By using skip connections, the two types of features complement each other, improving the model's robustness to noise interference and its sensitivity to weak transient signals.

[0108] 3. Multi-task output head:

[0109] Fault line classification: Softmax outputs the fault probability of each line.

[0110] The structure is the same as the fault type classification head. The number of neurons output by the fully connected layer is equal to the number of lines (e.g., 10 lines correspond to 10 neurons). Softmax generates the fault probability of each line.

[0111] Loss function: Cross-entropy loss for L-line.

[0112] Fault type classification: probability of high impedance / arc / intermittent faults in independent branch outputs.

[0113] Fully connected layer (number of neurons = number of fault types, such as 3 types corresponding to 3 neurons).

[0114] The Softmax activation function transforms the output into a probability distribution.

[0115] (13)

[0116] in, Classifying fault types by head number k The output value of each neuron.

[0117] Training labels: Fault types are encoded using one-hot encoding, for example, high-impedance faults are labeled as [1,0,0].

[0118] Fault location regression: Fully connected layer predicts fault distance (normalized to [0,1]).

[0119] The fully connected layer outputs a single value (such as the normalized fault distance). Mean squared error (MSE) loss is used:

[0120] (14)

[0121] Loss function: Joint optimization of weighted cross-entropy and mean squared error:

[0122] (15)

[0123] In the formula, , , These are the fault line classification loss, fault type classification loss, and fault location regression loss, respectively; weights Adjust according to task importance (e.g., α=0.6, β=0.3, γ=0.1).

[0124] Global-local feature complementarity:

[0125] DenseNet focuses on local fault modes (such as current abrupt changes), while Transformer models waveform periodicity (such as high-frequency oscillations in arc faults). The fusion of the two enhances feature richness.

[0126] Shared feature extraction layer:

[0127] Sharing mechanism: All tasks share the first few layers of DenseNet-121 (such as convolutional layers and dense blocks), and only the last layer is adapted to different tasks independently through branches.

[0128] Backbone network: An improved DenseNet-121 is used as a shared feature extractor, with input being a two-dimensional image generated by GAF (such as a GASF / GADF image).

[0129] Multi-task output headers: Multiple independent fully connected layers (output headers) are added to the end of DenseNet-121, each corresponding to a different task.

[0130] Task 1 (Faulty Line Classification): Output line category probability of the fully connected layer (Softmax activation).

[0131] Task 2 (Fault Type Classification): Probability of high impedance / arc / intermittent faults in fully connected layer output (Softmax activation).

[0132] Task 3 (Fault Location Regression) (Optional): Output the normalized fault distance (linear activation) of the fully connected layer.

[0133] Experiments show that introducing the Transformer improves the accuracy of identifying intermittent faults from 89% to 95%. Sharing the feature extraction layer reduces redundant computation; joint supervision with multi-task labels enhances the model's generalization ability.

[0134] process:

[0135] 1. Four-channel E-GAF images (GASF+GADF for current / voltage) are input to Trans-DenseNet.

[0136] 2. DenseNet extracts local features in the first three dense blocks → Transformer encodes global dependencies → Residual fusion → The fourth dense block further refines the features.

[0137] 3. After the final feature map is processed by global average pooling (GAP), it is input into three fully connected layers to generate prediction results for the line, type, and location.

[0138] S4: Knowledge Distillation Technology

[0139] Lightweight deployment driven by knowledge distillation:

[0140] Teacher-student model architecture: The hybrid deep learning model (Trans-DenseNet) is used as the teacher model. It achieves high-precision fault classification (multi-task output of line, type, and location) by combining the local feature extraction of DenseNet-121 with the global temporal modeling capability of Transformer.

[0141] Knowledge distillation is employed to transfer knowledge from the teacher model to the lightweight student model MobileNetV3. By optimizing the KL divergence loss function (Equation 16) and soft-label supervision (Equation 17), model accuracy is preserved while the number of parameters is reduced.

[0142] (16)

[0143] (17)

[0144] In the formula, T is the temperature coefficient; , , These represent the true labels (hard labels, one-hot encoded), the probability distribution of the teacher model's output after temperature scaling, and the student model's output after being scaled by the same temperature, respectively. The cross-entropy loss function is used to ensure the accuracy of the student model's predictions of the true labels; This refers to the KL divergence loss (Kullback-Leibler Divergence). , These are weighting coefficients, which control the contribution ratios of cross-entropy loss and KL divergence loss, respectively. This is the raw output (logits) of the teacher model.

[0145] The inference speed is 3 times faster than the original Trans-DenseNet, with a latency of ≤10ms, meeting the real-time requirements of the power distribution network.

[0146] like Figure 2 As shown, the traditional GAF (Gaussian Alternating Current) line selection confusion matrix results show a line selection accuracy of 91.2%. Compared to the traditional method, GAF transformation effectively suppresses noise interference, improves the model's adaptability to complex faults such as high resistance and arcing, and significantly improves line selection efficiency through an end-to-end process. Experimental results are as follows... Figure 3 As shown, the multimodal fusion enhancement and improved Gram angle field transform line selection confusion matrix proposed in this invention achieves a line selection accuracy of over 95% even when the signal-to-noise ratio is <10dB. It can achieve high-precision fault location in both simulation and real-world scenarios, providing reliable support for intelligent fault self-healing in new power distribution networks.

[0147] Example 2

[0148] As a second aspect of the present invention, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-described distribution network fault location method based on a multimodal enhancement and hybrid deep learning model. In addition to the processors, memory, and interfaces described above, any data processing device in the embodiments may also include other hardware depending on the actual function of the data processing device, which will not be elaborated further.

[0149] Example 3

[0150] As a third aspect of the present invention, this application also provides a computer-readable storage medium storing computer instructions thereon, which, when executed by a processor, implement the power distribution network fault location method based on the multimodal enhancement and hybrid deep learning model described above. The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0151] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A power distribution network fault line selection method based on multi-modal enhancement and hybrid deep learning model, characterized by the steps of include: Collect zero-sequence current and voltage transient waveform data of each line in the distribution network when a fault occurs; Gram angular field transformation was performed on the zero-sequence current and voltage transient waveform data using a penalized inner product, which was converted into two-dimensional GASF and GADF images of the zero-sequence current and voltage transient waveforms, respectively, and then stitched together to generate a four-channel E-GAF image. Four-channel E-GAF images are input into the Trans-DenseNet model for fault identification. The Trans-DenseNet model integrates local and global features by embedding a Transformer encoder layer in DenseNet-121 and outputs the final fault line selection result.

2. The power distribution network fault line selection method based on multi-modal enhancement and hybrid deep learning model according to claim 1, characterized in that, The method preprocesses the zero-sequence current and voltage transient waveform time series data of each group of samples, specifically including: normalization scaling using sliding window dynamic normalization, and wavelet thresholding for noise reduction.

3. The power distribution network fault line selection method based on multi-modal enhancement and hybrid deep learning model according to claim 1, characterized in that, The Gram angle field transformation based on the inner product with a penalty term is performed on the zero-sequence current and voltage transient waveform data as follows: Polar coordinate transformation is performed on the time series signals of zero-sequence current and voltage transient waveforms to convert the time series data into vectors: wherein denotes a phase angle; is a normalized scaled time series signal ; denotes a radius of the polar coordinates; is a time stamp; denotes a constant coefficient of the regularized polar coordinate generation space; Based on the inner product between vectors with a penalty term, the Gram angle of the zero-sequence current and voltage transient waveforms, and the field GASF and Gram angle difference field GADF are calculated respectively: ， 。 4. The power distribution network fault line selection method based on multi-modal enhancement and hybrid deep learning model according to claim 3, characterized in that, The current data makes ; when common mode interference occurs, makes ; Substitute the two different disturbances into the following inner product formula to calculate GASF and GADF respectively: In the formula, denotes the difference of the angles of adjacent time points.

5. The power distribution network fault line selection method based on multi-modal enhanced and hybrid deep learning model according to claim 1, characterized in that, The Trans-DenseNet model uses DenseNet-121 as the backbone network, retains the four dense blocks of the original DenseNet, and is used to extract local texture features of the image. A Transformer encoder layer is inserted between the third and fourth dense blocks to capture global temporal dependencies using a self-attention mechanism. Line, type, and location prediction results are generated through a multi-task output head.

6. The power distribution network fault line selection method based on multi-modal enhancement and hybrid deep learning model according to claim 5, characterized in that, The specific workflow of the Trans-DenseNet model is as follows: Four-channel E-GAF images are input into the Trans-DenseNet model; Local features are extracted from the first three dense blocks of DenseNet; The Transformer encoder flattens the feature map output from the third dense block of DenseNet into a sequence, adds positional encoding, and performs multi-head attention computation. The Transformer output sequence is reshaped into a feature map and residually connected to the features of subsequent layers of DenseNet, while multi-scale features are preserved through skip connections. The features are further refined using the fourth dense block of DenseNet; The final feature map is processed by global average pooling and then input into the multi-task output head, where three fully connected layers generate prediction results for the line, type, and location, respectively.

7. The power distribution network fault line selection method based on multi-modal enhancement and hybrid deep learning model according to claim 5, characterized in that, The prediction results of the multi-task output head are as follows: Faulty circuit classification: The faulty circuit is classified through a fully connected layer, the number of output neurons of the fully connected layer is equal to the number of circuits, and then the faulty probability of each circuit is output through the Softmax activation function. Fault type classification: via a fully connected layer, the number of neurons is equal to the number of fault types; and then the output is converted into a probability distribution of high resistance / arc / intermittent fault through a Softmax activation function; Fault location regression: the fault distance is predicted through a fully connected layer, and the output is a normalized fault distance.

8. The power distribution network fault line selection method based on multi-modal enhancement and hybrid deep learning model according to claim 1, characterized in that, The Trans-DenseNet model is lightened by using a knowledge distillation technology, a teacher-student model architecture is used, the Trans-DenseNet model is taken as a teacher model, and the knowledge of the teacher model is migrated to a light student model for deployment by using the knowledge distillation technology.

9. A power distribution network fault line selection device based on multi-modal enhancement and hybrid deep learning model, characterized in that, Comprise: One or more processors; Memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the power distribution network fault line selection method based on the multi-modal enhanced and hybrid deep learning model as claimed in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the steps of the power distribution network fault line selection method based on the multi-modal enhanced and hybrid deep learning model as claimed in any one of claims 1-8.

Images