Power distribution network fault line selection method and equipment based on multi-modal enhancement and hybrid deep learning model

Abstract

The invention relates to the technical field of power distribution network fault diagnosis, in particular to a power distribution network fault line selection method and device based on a multi-modal enhancement and hybrid deep learning model, and the method comprises the steps: collecting the zero-sequence current and voltage transient waveform data of each line of a power distribution network when a fault occurs; the one-dimensional waveform data is subjected to Gramer angle field transformation by adopting an inner product with a penalty term, and is converted into a two-dimensional GASF image and a GADF image respectively, and a four-channel E-GAF image is generated by splicing the two-dimensional GASF image and the two-dimensional GADF image; and inputting the E-GAF image into a Trans-DenseNet model to carry out fault discrimination, and by embedding a Transform encoder layer in the DenseNet, fusing local-global features, and outputting a final fault line selection result. Compared with the prior art, the method can adapt to complex faults such as high resistance and electric arc, high-precision fault positioning is achieved, and reliable support is provided for intelligent fault self-healing of a novel power distribution network.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault diagnosis, and in particular to a distribution network fault line selection method integrating multimodal data enhancement, improved Gram angle field transformation and a hybrid deep learning model. Background Art

[0002] Existing distribution network fault detection methods based on Gram Angle Field (GAF), such as Chinese patent applications CN202411303011.7 and CN202411302666.2, can improve diagnostic accuracy by utilizing two-dimensional image features, but still have the following shortcomings: Noise sensitivity: Traditional GAFs have limited ability to suppress high-frequency noise and are particularly susceptible to interference in weak transient signal scenarios.

[0003] Single data source: Relying only on zero-sequence current data, ignoring the potential value of multimodal information such as voltage.

[0004] Insufficient model generalization: Fixed normalization methods are difficult to adapt to complex working conditions, and the classification model has limited ability to distinguish multiple types of faults.

[0005] Real-time bottleneck: Deep models have high computational complexity and are difficult to meet the real-time processing requirements of large-scale distribution networks. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies such as noise sensitivity, single data source and insufficient model generalization, and to provide a distribution network fault line selection method that integrates multimodal data enhancement, improved Gram angle field transform and hybrid deep learning model.

[0007] The purpose of the present invention can be achieved by the following technical solutions: As a first aspect of the present invention, a method for fault line selection in a distribution network based on multimodal enhancement and a hybrid deep learning model is provided, comprising the following steps: Collect zero-sequence current and voltage transient waveform data of each line in the distribution network when a fault occurs; The zero-sequence current and voltage transient waveform data are transformed into two-dimensional GASF and GADF images of the zero-sequence current and voltage transient waveforms respectively using the inner product with penalty term, and then spliced ​​to generate a four-channel E-GAF image. The four-channel E-GAF image is input into the Trans-DenseNet model for fault identification. The Trans-DenseNet model embeds the Transformer encoder layer in the DenseNet-121 to fuse local-global features and outputs the final fault line selection result.

[0008] As a preferred technical feature, the method preprocesses the zero-sequence current and voltage transient waveform time series data of each group of samples collected, specifically including: using sliding window dynamic normalization for normalization and scaling, and using wavelet threshold denoising.

[0009] As a preferred technical feature, the zero-sequence current and voltage transient waveform data are subjected to a Gram angle field transformation based on an inner product with a penalty term, specifically as follows: Perform polar coordinate transformation on the zero-sequence current and voltage transient waveform time series signals and convert the time series data into vectors:

[0010] Where, represents the phase angle; It is the time series signal after normalization and scaling; Represents the radius of polar coordinates; is the timestamp; represents the constant coefficient of the regularized polar coordinate generating space; Based on the inner product with penalty terms between vectors, the Gram angle sum field GASF and the Gram angle difference field GADF of the zero-sequence current and voltage transient waveforms are calculated respectively:

[0011]

[0012] As a preferred technical feature, when differential mode interference occurs, the current data is ; When common mode interference occurs, ; Substitute the above two different interferences into the following inner product formula to calculate GASF and GADF respectively: , .

[0013] Where, Indicates the difference in angles between adjacent time points.

[0014] 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, and is used to extract local texture features of the image; inserts a Transformer encoder layer between the third and fourth dense blocks, and uses the self-attention mechanism to capture global temporal dependencies; and generates line, type, and position prediction results through a multi-task output head.

[0015] As a preferred technical feature, the specific workflow of the Trans-DenseNet model is as follows: Four-channel E-GAF image input Trans-DenseNet model; Extract local features through the first three dense blocks of DenseNet; The Transformer encoder flattens the feature map output by the third dense block of DenseNet into a sequence, adds position encoding, and performs multi-head attention calculation; Reshape the Transformer output sequence into a feature map, perform residual connections with the features of the subsequent layers of DenseNet, and retain multi-scale features through skip connections; The features are further refined through the fourth dense block of DenseNet; The final feature map is globally averaged and then input into the multi-task output head, where three fully connected layers generate line, type, and position prediction results respectively.

[0016] As a preferred technical feature, the prediction results of the multi-task output head are specifically as follows: Faulty line classification: Through a fully connected layer, the number of output neurons in the fully connected layer is equal to the number of lines, and then the Softmax activation function is used to output the fault probability of each line; Fault type classification: Through a fully connected layer, the number of neurons is equal to the number of fault types; the output is then converted into the probability distribution of high resistance / arc / intermittent faults through the Softmax activation function; Fault location regression: The fault distance is predicted through a fully connected layer and the normalized fault distance is output.

[0017] As an optimal technical feature, the Trans-DenseNet model uses knowledge distillation technology for lightweight deployment. By utilizing the teacher-student model architecture, the Trans-DenseNet model is used as the teacher model, and knowledge distillation technology is used to migrate the knowledge of the teacher model to a lightweight student model for deployment.

[0018] As a second aspect of the present invention, a distribution network fault line selection device based on multimodal enhancement and hybrid deep learning model is provided, comprising: one or more processors; a 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 distribution network fault line selection method based on multimodal enhancement and hybrid deep learning model as described above.

[0019] As a third aspect of the present invention, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the distribution network fault line selection method based on multimodal enhancement and hybrid deep learning model as described above.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention employs a multimodal fusion approach, performing Gram angle field transformation on both 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 voltage sag or resonance characteristics at the moment of the fault, enabling in-depth analysis of potential distribution network fault information. The one-dimensional waveform data is converted into two-dimensional GASF and GADF images, respectively, making the information contained in the data more easily exploitable by deep learning.

[0021] 2) This paper proposes a hybrid deep learning model (Trans-DenseNet) that embeds a Transformer encoder layer within DenseNet-121 to construct a local-global feature fusion module. The dense blocks of DenseNet focus on local fault patterns (such as current mutation points), while the Transformer encoder captures global waveform periodicity (such as the high-frequency oscillations of arc faults). These two are fused through a residual connection to enhance feature richness. The resulting feature map, generated by the local-global fusion, is then used through a multi-task output head to predict the fault probability, fault type, and location of each line. This improves the model's adaptability to complex faults such as high resistance and arc faults, while significantly improving line selection efficiency through the end-to-end process.

[0022] 3) This paper also uses knowledge distillation technology to migrate the knowledge of the hybrid deep learning model Trans-DenseNet to the lightweight MobileNetV3, and implements hardware acceleration based on FPGA, which increases the inference speed by 3 times and reduces the latency to less than 10ms, meeting the real-time diagnosis needs of large-scale distribution networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of the distribution network fault line selection method based on multimodal enhanced Gram angle field and hybrid deep learning model of the present invention.

[0024] Figure 2 Confusion matrix of line selection results for traditional Gram Angle Field (GAF)-based distribution network fault detection method; Figure 3 This is the confusion matrix of the line selection results of the distribution network fault line selection method based on the multimodal enhanced Gram angle field and hybrid deep learning model of the present invention. DETAILED DESCRIPTION

[0025] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0026] Example 1 The present invention proposes a distribution network fault line selection method based on multimodal enhanced Gram angular field and hybrid deep learning model. Multimodal data fusion is adopted, and the input signals of the two dimensions of zero-sequence current and voltage are simultaneously processed to improve the fault identification capability. The one-dimensional waveform data is converted into two-dimensional GASF and GADF images using Gram angular field transformation, and spliced ​​into four-channel input to retain the multi-dimensional time series correlation characteristics. By embedding the Transformer encoder layer in DenseNet-121, a local-global feature fusion module is constructed to realize the accurate identification of the fault line. The model is compressed into a lightweight MobileNetV3 through knowledge distillation, and hardware acceleration is realized based on FPGA. The single sample processing delay is ≤8ms, which meets the real-time diagnosis needs of large-scale distribution networks. Figure 1 As shown, the specific steps of the method of the present invention are as follows: Step S1, collecting zero-sequence current and voltage waveform data of each line when a fault occurs; Step S2: Dynamic normalization and wavelet noise reduction are performed on the waveform data. Specifically, the zero-sequence current and voltage time series data are dynamically normalized using a sliding window to suppress local noise, and wavelet threshold denoising is used to enhance signal characteristics. Step S3, performing 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 splicing them to generate a four-channel E-GAF image input; 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.

[0027] In step S3, the one-dimensional time series signal is converted into a two-dimensional image using the Gram's angle field, which maximizes the preservation of the original signal's features without losing them. The Gram's angle field's theoretical basis is to convert one-dimensional data from a Cartesian coordinate system into a polar coordinate system. Then, based on its unique inner product definition, trigonometric functions, such as sums and differences, are used to obtain information representing temporal correlations. This information is then tiled across the image from the upper left to the lower right, generating two types of images: GASF and GADF.

[0028] For time series signals , here Indicates the total number of time points: For time point, .Will The GAF process is as follows: The time series signal of each group of samples The local mean is calculated using a sliding window ( ) and standard deviation ( ) Dynamic normalization processing method scaling, such as formula (1).

[0029] (1) Where, is the sliding window length, To prevent the small constant from dividing by zero. Dynamic normalization is used instead of fixed scaling. The noise resistance of the interval is significantly improved: local statistics suppress noise interference and avoid the influence of global extreme values ​​on normalization. For example, if there is high-frequency noise in the window, It will amplify the local fluctuation of noise and dynamically adjust the scaling ratio. The adaptive ability is significantly improved: window length It can be adjusted according to the signal characteristics (such as short window to capture transient mutations, long window to smooth periodic fluctuations). More stable values: introduce small constants (like ), to prevent the denominator from being zero.

[0030] Formula (2) represents the Gram matrix. The inner product between different vectors can show the correlation between vectors, while the angle between vectors shows the degree of correlation between the same vectors.

[0031] (2) Where, is the Gram matrix; is the inner product operation.

[0032] 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 formula (3): (3) Where, is the timestamp; represents the constant coefficient of the regularized polar coordinate generating space; represents the phase angle; Represents the radius of polar coordinates; is the time series signal after normalization and scaling .

[0033] If the Gram matrix is ​​used to define the inner product, it is difficult to distinguish the Gaussian noise of the output data. Therefore, in order to eliminate the influence of Gaussian white noise, the Gram angle field defines two unique inner product forms with penalty terms. The formulas are as follows (4) and (5): (4) (5) Based on the two different inner product definitions above, we can obtain two different Gramian angle fields: Gramian angle summation fields GASF (gramian angular summation fields) and Gramian angle difference fields GADF (gramian angular difference fields). Formulas (6) and (7) are shown as follows: (6) (7) For current data that are close in time domain, GAF transformation has the ability to suppress signal disturbances. When differential mode interference occurs, there are: ; When common mode interference occurs: , by substituting the above two different interferences into equations (8) and (9), the interference can be automatically removed. The process of GAF generating a two-dimensional image is as follows: Figure 1 shown.

[0034] (8) (9) Where, Represents the difference in angles between adjacent time points. This difference is used to capture the dynamic characteristics of the time series and is a key parameter in constructing the Gram angle field matrix.

[0035] The present invention adopts a multimodal fusion method to expand the data source. The multimodal data specifically includes: Zero-sequence current: reflects the amplitude and waveform distortion of the fault current (such as high-frequency oscillation of arc fault).

[0036] Voltage transient waveform: Captures voltage drops or resonance characteristics at the moment of fault (such as low-amplitude transients of high-resistance faults).

[0037] Fusion method: Perform wavelet denoising and dynamic normalization on the current and voltage signals (as shown above). Then, generate GASF (Equation 6) and GADF (Equation 7) images for the current and voltage signals, respectively, to form a four-channel input.

[0038] The voltage signal is sensitive to weak transient faults (such as high-resistance grounding), making up for the shortcomings of a single current signal.

[0039] In step 4, the present invention proposes a hybrid deep learning model (Trans-DenseNet) to achieve accurate identification of faulty lines.

[0040] 1. Trans-DenseNet model design: The core idea of ​​the hybrid deep learning model is to embed the Transformer encoder layer in DenseNet-121, build a local-global feature fusion module, and achieve end-to-end multi-task learning.

[0041] DenseNet-121 backbone network: retains the four dense blocks of the original DenseNet to extract local texture features of the image (such as mutation points and periodic patterns in GAF images).

[0042] Each dense block contains multiple convolutional layers, which enhance gradient flow and feature expression capabilities through dense connections (feature reuse).

[0043] Transformer encoder insertion position: A Transformer encoder layer is inserted between the third and fourth dense blocks to capture global temporal dependencies using the self-attention mechanism.

[0044] 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: (10) Where, Represents the enhanced sequence input to the Transformer encoder, integrating feature information and location information ; Represents a flattening operation; Represents the three-dimensional feature map output by the DenseNet backbone network (DenseNet-121), usually generated by the first few dense blocks (Dense Block), with a size of H×W×C (height×width×number of channels); Represents positional encoding, which is used to inject position information into the sequence.

[0045] Multi-Head Attention calculation: (11) Where, Obtained by linear transformation of the input sequence, is the attention head dimension.

[0046] 2. Local-global feature fusion: Reshape the Transformer output sequence into a feature map and perform residual connections with the features of the subsequent layers of DenseNet: (12) Where, It is the fused feature map, which combines local spatial features (Dense Block) and global temporal features (Transformer); For the general Input into the subsequent Dense Block to further extract higher-level local features; The three-dimensional feature map output by the DenseNet backbone network (DenseNet-121) is usually generated by the first few dense blocks (Dense Block) and has a size of H×W×C (height×width×number of channels); To transform the sequence output by Transformer Reshape into a three-dimensional feature map and restore to the same Same dimensions H×W×C; is the output sequence of the Transformer encoder, which captures the global temporal dependencies through the self-attention mechanism.

[0047] Skip connections are used to preserve multi-scale features and enhance model robustness. Connection module: DenseNet's intermediate layer features (such as the feature map output by the third Dense Block) Compared with the global features of Transformer encoder ( reshaped feature maps).

[0048] Dense Block: Continue to extract local spatial features (such as mutation points and periodic textures of fault waveforms).

[0049] Reshape(Z_Transformer): Converts the global temporal dependency features (long-range associations modeled by self-attention) output by the Transformer into a three-dimensional tensor of the same dimension as the DenseNet feature map.

[0050] Jump connection method: Residual addition (Add): Add the output of the Dense Block to the reshaped feature map output by the Transformer element by element to form a fusion feature .

[0051] While retaining local details (DenseNet), the global context (Transformer) is introduced to enhance the model's ability to model complex fault modes (such as high resistance and arcing).

[0052] Fusion advantages: Through 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.

[0053] 3. Multi-task output head: Fault line classification: Softmax outputs the fault probability of each line.

[0054] The structure is the same as the fault type classification head. The number of output neurons in the fully connected layer is equal to the number of lines (for example, 10 lines correspond to 10 neurons). Softmax generates the fault probability of each line.

[0055] Loss function: cross entropy loss L circuit.

[0056] Fault type classification: independent branch output high resistance / arc / intermittent fault probability.

[0057] Fully connected layer (the number of neurons = the number of fault types, for example, 3 types correspond to 3 neurons).

[0058] Softmax activation function converts the output into a probability distribution.

[0059] (13) in, Classify the fault type first k The output value of a neuron.

[0060] Training labels: Fault types are encoded using one-hot encoding. For example, the high-resistance fault label is [1, 0, 0].

[0061] Fault location regression: The fully connected layer predicts the fault distance (normalized to [0,1]).

[0062] The fully connected layer outputs a single value (such as the normalized fault distance , using mean square error (MSE) loss: (14) Loss function: Joint optimization of weighted cross entropy and mean square error: (15) Where, 、 、 They are fault line classification loss, fault type classification loss and fault location regression loss; weights Adjust according to the importance of the task (e.g., α=0.6, β=0.3, γ=0.1).

[0063] Global-local feature complementation: DenseNet focuses on local fault modes (such as current mutation points), while Transformer models waveform periodicity (such as high-frequency oscillations of arc faults). The fusion of the two improves feature richness.

[0064] Shared feature extraction layer: 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 independently adapted to different tasks through branches.

[0065] Backbone network: The improved DenseNet-121 is used as a shared feature extractor, and the input is a two-dimensional image generated by GAF (such as GASF / GADF image).

[0066] Multi-task output head: Add multiple independent fully connected layers (output heads) at the end of DenseNet-121, corresponding to different tasks: Task 1 (faulty line classification): The fully connected layer outputs line category probability (Softmax activation).

[0067] Task 2 (fault type classification): The fully connected layer outputs the probability of high resistance / arc / intermittent fault (Softmax activation).

[0068] Task 3 (Fault Location Regression) (Optional): The fully connected layer outputs the normalized fault distance (linear activation).

[0069] Experiments show that the introduction of the Transformer improves the accuracy of intermittent fault recognition from 89% to 95%. Sharing feature extraction layers reduces redundant computation, and multi-task label joint supervision enhances model generalization.

[0070] process: 1. Four-channel E-GAF image (GASF+GADF of current / voltage) is input into Trans-DenseNet.

[0071] 2. DenseNet’s first three dense blocks extract local features → Transformer encodes global dependencies → Residual fusion → The fourth dense block further refines the features.

[0072] 3. After the final feature map passes through global average pooling (GAP), it is input into three fully connected layers to generate line, type, and position prediction results.

[0073] S4: Knowledge Distillation Technology Lightweight deployment driven by knowledge distillation: Teacher-student model architecture: A hybrid deep learning model (Trans-DenseNet) is used as the teacher model. It combines the local feature extraction of DenseNet-121 with the global temporal modeling capabilities of Transformer to achieve high-precision fault classification (multi-task output of line, type, and location).

[0074] Using knowledge distillation technology, the knowledge of the teacher model is transferred to the lightweight student model MobileNetV3. By optimizing the KL divergence loss function (Equation 16) and soft label supervision (Equation 17), the model accuracy is retained while the number of parameters is compressed: (16) (17) Where T is the temperature coefficient; 、 、 They represent the true label (hard label, one-hot encoding), the probability distribution of the teacher model's output after temperature scaling, and the student model's output after the same temperature scaling; Cross-Entropy Loss is a function that ensures the accuracy of the student model’s prediction of the true label. is the KL divergence loss (Kullback-Leibler Divergence); 、 are weight coefficients, which control the contribution ratio of cross entropy loss and KL divergence loss respectively; is the raw output (logits) of the teacher model.

[0075] The inference speed is 3 times faster than the original Trans-DenseNet, with a latency of ≤10ms, meeting the real-time requirements of the distribution network.

[0076] like Figure 2 The figure shows the line selection results of the traditional GAF confusion matrix, with a line selection accuracy of 91.2%. Compared with traditional methods, GAF transformation effectively suppresses noise interference and improves the model's adaptability to complex faults such as high resistance and arcing. At the same time, the end-to-end process significantly improves line selection efficiency. Figure 3 As shown in the figure, the multimodal fusion enhancement and improved Gram angle field transform line selection confusion matrix proposed in the present invention can achieve a line selection accuracy of more than 95% when the signal-to-noise ratio is less than 10dB; high-precision fault location can be achieved in both simulation and actual scenarios, providing reliable support for the intelligent fault self-healing of the new distribution network.

[0077] Example 2 As a second aspect of the present invention, the present application also provides an electronic device, comprising: one or more processors; a 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 above-mentioned distribution network fault line selection method based on multimodal enhancement and hybrid deep learning model. In addition to the above-mentioned processors, memories, and interfaces, any device with data processing capabilities in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0078] Example 3 As a third aspect of the present invention, the present application also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-mentioned method for fault line selection in a distribution network based on multimodal enhancement and a hybrid deep learning model. The computer-readable storage medium can be an internal storage unit of any device with data processing capabilities described in any of the aforementioned 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, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium can also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.

[0079] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A method for fault line selection in distribution network based on multimodal enhancement and hybrid deep learning model, characterized in that the steps include: Collect zero-sequence current and voltage transient waveform data of each line in the distribution network when a fault occurs; The zero-sequence current and voltage transient waveform data are transformed into two-dimensional GASF and GADF images of the zero-sequence current and voltage transient waveforms respectively using the inner product with penalty term, and then spliced ​​to generate a four-channel E-GAF image. The four-channel E-GAF image is input into the Trans-DenseNet model for fault identification. The Trans-DenseNet model embeds the Transformer encoder layer in the DenseNet-121 to fuse local-global features and outputs the final fault line selection result.

2. A distribution network fault line selection method based on multimodal 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 collected samples, specifically including: using sliding window dynamic normalization for normalization and scaling, and using wavelet threshold denoising.

3. The method for fault line selection in a distribution network based on multimodal enhancement and hybrid deep learning model according to claim 1, characterized in that: The Gram angle field transformation based on the inner product with penalty term is performed on the zero-sequence current and voltage transient waveform data, specifically as follows: Perform polar coordinate transformation on the zero-sequence current and voltage transient waveform time series signals and convert the time series data into vectors: Where, represents the phase angle; The time series signal after normalization and scaling ; Represents the radius of polar coordinates; is the timestamp; represents the constant coefficient of the regularized polar coordinate generating space; Based on the inner product with penalty terms between vectors, the Gram angle sum field GASF and the Gram angle difference field GADF of the zero-sequence current and voltage transient waveforms are calculated respectively: ， 。 4. A distribution network fault line selection method based on multimodal enhancement and hybrid deep learning model according to claim 3, characterized in that: When differential mode interference occurs, the current data is ; When common mode interference occurs, ; Substitute the two different interferences into the following inner product formula to calculate GASF and GADF respectively: Where, Indicates the difference in angles between adjacent time points.

5. The method for fault line selection in a distribution network based on multimodal enhancement and hybrid deep learning model according to claim 1, characterized in that: The Trans-DenseNet model uses DenseNet-121 as the backbone network and retains the four dense blocks of the original DenseNet 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 the self-attention mechanism. The multi-task output head generates line, type, and position prediction results.

6. A distribution network fault line selection method based on multimodal 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 image input Trans-DenseNet model; Extract local features through the first three dense blocks of DenseNet; The Transformer encoder flattens the feature map output by the third dense block of DenseNet into a sequence, adds position encoding, and performs multi-head attention calculation; Reshape the Transformer output sequence into a feature map, perform residual connections with the features of the subsequent layers of DenseNet, and retain multi-scale features through skip connections; The features are further refined through the fourth dense block of DenseNet; The final feature map is globally averaged and then input into the multi-task output head, where three fully connected layers generate line, type, and position prediction results respectively.

7. The method for fault line selection in a distribution network based on multimodal 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 line classification: Through a fully connected layer, the number of output neurons in the fully connected layer is equal to the number of lines, and then the Softmax activation function is used to output the fault probability of each line; Fault type classification: Through a fully connected layer, the number of neurons is equal to the number of fault types; the output is then converted into the probability distribution of high resistance / arc / intermittent faults through the Softmax activation function; Fault location regression: The fault distance is predicted through a fully connected layer and the normalized fault distance is output.

8. The method for fault line selection in a distribution network based on multimodal enhancement and hybrid deep learning model according to claim 1, characterized in that: The Trans-DenseNet model uses knowledge distillation technology for lightweight deployment. By utilizing the teacher-student model architecture, the Trans-DenseNet model is used as the teacher model. The knowledge of the teacher model is transferred to the lightweight student model for deployment using knowledge distillation technology.

9. A distribution network fault line selection device based on multimodal enhancement and hybrid deep learning model, characterized in that: include: one or more processors; a 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 distribution network fault line selection method based on multimodal enhancement and hybrid deep learning model as described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the distribution network fault line selection method based on multimodal enhancement and hybrid deep learning model as described in any one of claims 1 to 8 are implemented.

Images