Power distribution network fault line selection method based on image transformation and target detection network

Abstract

The invention discloses a power distribution network fault line selection method based on image transformation and a target detection network. According to the method, after a fault occurs, three-phase voltages of all lines suspected to have the fault are collected, voltage phase-mode transformation is conducted on the three-phase voltages to obtain line-mode components, point symmetry Hilbert transformation is conducted on one-dimensional signals of the line-mode components of all the lines to obtain point symmetry Hilbert transformation images of the lines, then the point symmetry Hilbert transformation images of all the lines are spliced into a complete image, and the complete image is obtained. And inputting the complete image into a target detection network for target detection, wherein a line corresponding to a detection frame marked by the target detection network is a fault circuit. The method has the advantages of high positioning precision, strong anti-interference capability, high speed, strong generalization capability and the like.

Description

Technical Field

[0001] The invention belongs to the technical field of distribution networks, and in particular relates to a fault line selection method for a distribution network. Background Art

[0002] As a key component of the power system, the stability and reliability of the distribution network have a decisive impact on the security of power supply. Against the backdrop of the continued growth in global electricity demand, the continuous expansion of the power grid has brought new challenges to the operation and maintenance of the distribution network. Among them, the accurate and rapid location of distribution network faults is an important part of ensuring the safe operation of the power system.

[0003] Existing distribution network fault line selection methods mainly include impedance measurement method, traveling wave method, zero-sequence current method and machine learning-based method:

[0004] The impedance measurement method determines the fault location by detecting the change in the electrical impedance of the distribution line, which is suitable for simple distribution networks. However, if there are factors such as nonlinear loads, load fluctuations and short circuits in the distribution network, the accuracy of impedance measurement cannot be guaranteed. In addition, the complex structure of the distribution network, such as the existence of multiple branches and multiple feeders, also makes it difficult for the impedance measurement method to accurately determine the specific location of the fault, especially in complex distribution networks, where obvious errors and instability will be shown.

[0005] The traveling wave method uses the propagation characteristics of electromagnetic waves generated by faults to locate faults. In theory, it has high positioning accuracy. However, in practical applications, the reflection and refraction of traveling wave signals in multi-branch lines make signal processing complex and it is difficult to accurately extract traveling wave characteristics. In addition, this method has high requirements on equipment performance and is expensive, so its application is greatly limited.

[0006] The zero-sequence current method determines the fault line by analyzing the zero-sequence current in the line. It has high accuracy when dealing with faults with good symmetry. However, in the case of asymmetric faults or high-resistance grounding faults, the zero-sequence current changes little and the line selection effect is poor. At the same time, the zero-sequence current method is sensitive to the operating conditions of the power grid. Changes in the power grid structure and operating mode will significantly affect the line selection effect of this method.

[0007] With the rapid development of artificial intelligence technology, machine learning algorithms have begun to be introduced into the field of fault line selection in distribution networks. The machine learning-based method is to establish the association between fault characteristics and fault lines by learning historical fault data. However, this method has problems such as poor real-time performance and limited generalization ability in multi-branch and complex distribution networks. The lack of real-time performance is reflected in the long calculation time when the line complexity is high, making it difficult to achieve a fast response. Limited generalization ability means that when faced with different distribution network structures, the model cannot adapt to the complex and changeable grid operation status.

[0008] In summary, the existing distribution network fault line selection technology still has problems such as low positioning accuracy, poor anti-interference ability, insufficient real-time performance and limited generalization ability. Summary of the invention

[0009] The present invention proposes a distribution network fault line selection method based on image transformation and target detection network, the purpose of which is to solve the problems of low positioning accuracy, poor anti-interference ability, insufficient real-time performance and limited generalization ability in the existing distribution network fault line selection method.

[0010] The technical solution of the present invention is as follows:

[0011] A distribution network fault line selection method based on image transformation and target detection network, the method comprising:

[0012] Step 1: After a fault occurs, the three-phase voltages of all lines suspected of having a fault are collected, and the three-phase voltage sampling values ​​collected from each line at each moment are decoupled through voltage phase mode transformation, and the 2-mode component in the voltage phase mode transformation is used as the line mode component of the corresponding line at that moment, thereby obtaining a one-dimensional signal of the line mode component of each line;

[0013] Step 2, performing point-symmetric Hilbert transform on the one-dimensional signal of the line mode component of each line to obtain a point-symmetric Hilbert transform image of the line;

[0014] Step 3: Splice the point-symmetric Hilbert transform images of each circuit into a complete image, and input the complete image into the target detection network for target detection. The circuit corresponding to the detection box marked by the target detection network is the fault circuit.

[0015] As a further improvement of the distribution network fault line selection method based on image transformation and target detection network, the voltage phase model transformation process is:

[0016]

[0017] in, They are the collected three-phase voltages, They are respectively the converted 0-mode component, 1-mode component and 2-mode component.

[0018] As a further improvement of the distribution network fault line selection method based on image transformation and target detection network, step 2 specifically includes:

[0019] Step 2.1, convert the one-dimensional signal of the line mode component into an analytical signal:

[0020] a(t)=x(t)+jH[x(t)];

[0021] Where x(t) is the signal value at time t in the one-dimensional signal of the original line mode component, a(t) is the analytical signal value corresponding to x(t), and H[x(t)] is the Hilbert transform result of x(t): Where γ is an auxiliary variable;

[0022] Step 2.2, mapping the analytical signal to the polar coordinate space in a point-symmetric manner;

[0023] Step 2.3, draw the mapped points in the polar coordinate space, change a new color every 60 degrees during drawing, and use the obtained color space domain image containing RGB three channels as the point-symmetric Hilbert transform image.

[0024] As a further improvement of the distribution network fault line selection method based on image transformation and target detection network, the specific process of step 2.2 is as follows:

[0025] Assume that the amplitude of the i-th analytical signal value in the analytical signal is a env (i), the instantaneous phase is φ(i), then the polar coordinates of the two points of the analytical signal in the polar coordinate space that are symmetrical with respect to the radial line with an angle of θ are (θ(i), r(i)) and where radius r(i) = a env (i), the deflection angle of the first point The deflection angle of the second point θ is the rotation angle of the mirror symmetry plane, τ is the hysteresis coefficient, η is the deflection angle gain, φ max and φ min are respectively the maximum and minimum values ​​of the instantaneous phase in this group of analytical signals.

[0026] As a further improvement of the distribution network fault line selection method based on image transformation and target detection network: the mirror symmetry plane rotation angle θ is 60 degrees.

[0027] As a further improvement of the distribution network fault line selection method based on image transformation and target detection network, the parameter values ​​of the hysteresis coefficient τ and the deflection angle gain η are selected as follows:

[0028] (1) Obtain N analysis signals in normal states and N analysis signals in fault states, where different analysis signals come from different lines;

[0029] (2) Use different combinations of τ and η to obtain multiple sets of different parameter schemes;

[0030] (3) constructing point-symmetric Hilbert transform images for each analytical signal using each parameter scheme, and obtaining N normal images A from the normal state and N fault images B from the fault state for each parameter scheme;

[0031] (4) For each parameter scheme, the normal image A and the fault image B are cross-combined and the similarity distance of each group of normal image A and fault image B is calculated respectively: Where T A (n) is the nth texture parameter feature vector of image A; the texture parameter feature vector refers to a total of 16 texture parameter feature vectors obtained by respectively calculating four statistics of contrast, energy, correlation and homogeneity in the horizontal direction, vertical direction and two diagonal directions of the point-symmetric Hilbert transform image;

[0032] (5) For each parameter scheme, calculate the sum of the corresponding N*N similarity distances D sum =∑D(A,B);

[0033] (6) Finally, the solution with the largest sum of similarity distances is selected as the solution for determining the values ​​of τ and η.

[0034] As a further improvement of the distribution network fault line selection method based on image transformation and target detection network: the target detection network is a network model obtained by improving YOLOv5.

[0035] As a further improvement of the distribution network fault line selection method based on image transformation and target detection network, the improvement of YOLOv5 includes: using a hole convolution module to replace the SPP Block in YOLOv5;

[0036] The dilated convolution module consists of a projection layer, several residual modules and a convolution layer connected in sequence;

[0037] The projection layer includes a 1×1 convolution layer and a 3×3 convolution layer, the former is used to reduce the number of channels to half of the original, and the latter is used to further refine the context information;

[0038] The residual module enhances the feature expression capability through the residual structure, and includes three convolutional layers: the first convolutional layer is a 1×1 convolution, which is used to reduce the number of channels to one-fourth of the original number; the second convolutional layer is a 3×3 dilated convolution, which is used to expand the receptive field; the third convolutional layer is a 1×1 convolution, which is used to restore the number of channels; the feature map generated by the third convolutional layer is also fused with the input features of the residual module; the residual block is repeated 4 times, and each dilation convolution uses a different dilation rate, which is set to 2, 4, 6 and 8 respectively;

[0039] The last convolution layer in the dilated convolution module is a 1×1 convolution, which is used to restore the number of channels.

[0040] As a further improvement of the distribution network fault line selection method based on image transformation and target detection network, the improvement of YOLOv5 includes: adding a channel attention module after each upsampling and downsampling operation.

[0041] As a further improvement of the distribution network fault line selection method based on image transformation and target detection network, the improvement of YOLOv5 includes: using a lightweight module composed of depthwise separable convolution to replace the original standard convolution layer in the residual block of the CSP module;

[0042] The processing process of the depthwise separable convolution is as follows: first, a 3×3 depthwise convolution is used to perform regional convolution on each channel independently, and then a 1×1 point-by-point convolution is used to perform calculations between channels.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The present invention converts a one-dimensional signal into a two-dimensional image by introducing a point-symmetric Hilbert transform. The obtained image has obvious shape and texture features, which can not only reflect the changes in signal amplitude and frequency, but also show the timing correlation and dynamic changes of the signal. Compared with the traditional Hilbert transform, the point-symmetric Hilbert transform effectively enhances the visualization of fault features and improves the sensitivity to power grid signals. This transformation method can reveal the tiny features in complex fault modes, making the deep learning model more accurate in the feature extraction process and improving the accuracy of fault identification.

[0045] 2. The present invention introduces a lightweight module based on YOLOv5, reduces model parameters and reduces the amount of calculation. Compared with the traditional YOLOv5 model, the improved model increases the detection speed by nearly 20%, making it more suitable for the real-time fault line selection requirements of the distribution network, while also reducing the hardware deployment cost.

[0046] 3. The present invention introduces a dilated convolution module based on YOLOv5, which not only significantly increases the receptive field and enhances the detection capability of multi-scale fault features, but also enables the model to respond quickly in a complex and changeable power grid environment, further improving the fault detection speed and real-time performance.

[0047] 4. The point-symmetric Hilbert transform image obtained by the present invention can effectively reduce the impact of external interference on fault signals in the distribution network. Through rotational symmetry and plane mirror processing, the transformation can significantly weaken the interference caused by factors such as load fluctuations and voltage fluctuations, thereby improving the anti-interference and stability of fault characteristics.

[0048] 5. The addition of the channel attention module enables the YOLOv5 model to automatically focus on key channel features, further improving anti-interference and the stability of fault identification.

[0049] 6. The addition of the channel attention module and the multi-scale feature extraction layer makes the model more adaptable to different fault modes and improves the generalization ability of the model.

[0050] 7. The multi-channel feature expression of the point-symmetric Hilbert transform image improves the ability to distinguish fault signal features, thereby reducing the occurrence of overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a structural diagram of the improved YOLOv5 model of the present invention;

[0052] Figure 2 Schematic diagram of the dilated block in the present invention;

[0053] Figure 3 A schematic diagram of a radial distribution network model in an embodiment;

[0054] Figure 4 It is the point-symmetric Hilbert transform image corresponding to the five lines in the embodiment. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0056] A distribution network fault line selection method based on image transformation and target detection network, comprising:

[0057] Step 1: After a fault occurs, collect the three-phase voltages of all lines suspected of having a fault, and decouple the three-phase voltage sampling values ​​collected from each line at each moment through voltage phase mode transformation, and use the 2-mode component in the voltage phase mode transformation as the line mode component of the corresponding line at that moment, thereby obtaining a one-dimensional signal of the line mode component of each line.

[0058] Specifically, the voltage phase mode conversion process is:

[0059]

[0060] in, They are the collected three-phase voltages, They are respectively the converted 0-mode component, 1-mode component and 2-mode component.

[0061] Step 2: Perform point-symmetric Hilbert transform on the one-dimensional signal of the line mode component of each line to obtain a point-symmetric Hilbert transform image of the line.

[0062] Through point-symmetric Hilbert transform, the one-dimensional signal of the line mode component can be mapped to the polar coordinate system, thereby converting the original time series data into a two-dimensional image. Such mapping not only enhances the visualization of the fault signal, but also makes the analysis of fault characteristics more intuitive. Through this transformation, the generated image can clearly show the characteristics of the fault signal, where the texture changes of different images reflect the amplitude and frequency information in the original signal. This method provides strong support for the pattern recognition of fault signals, making the extraction and analysis of fault characteristics more efficient. The specific process includes:

[0063] Step 2.1, convert the one-dimensional signal of the line mode component into an analytical signal:

[0064] a(t)=x(t)+jH[x(t)];

[0065] Where x(t) is the signal value at time t in the one-dimensional signal of the original line mode component, a(t) is the analytical signal value corresponding to x(t), and H[x(t)] is the Hilbert transform result of x(t): Where γ is an auxiliary variable.

[0066] Step 2.2: Map the analytical signal to the polar coordinate space in a point-symmetric manner.

[0067] Assume that the amplitude of the i-th analytical signal value in the analytical signal is a env (i), the instantaneous phase is φ(i), then the polar coordinates of the two points of the analytical signal in the polar coordinate space that are symmetrical with respect to the radial line with an angle of θ are (θ(i), r(i)) and where radius r(i) = a env (i), the deflection angle of the first point The deflection angle of the second point θ is the rotation angle of the mirror symmetry plane, τ is the hysteresis coefficient, η is the deflection angle gain, φ max and φ min are the maximum and minimum values ​​of the instantaneous phase in the group of analytical signals. This mapping method can ensure the correlation of timing.

[0068] Step 2.3, draw the mapped points in polar coordinate space, change a new color every 60 degrees, and use the obtained color space domain image containing RGB three channels as the point symmetric Hilbert transform image. Multi-pixel channels can better characterize the deep features of one-dimensional signals.

[0069] Furthermore, the mirror symmetry plane rotation angle θ is selected to be 60 degrees, at which point-symmetric Hilbert transform image obtained has the best symmetry and readability.

[0070] Furthermore, different τ and η will affect the degree of difference of the image, improve the degree of difference between the point-symmetric Hilbert transform images of the normal line and the faulty line, and thus improve the accuracy of deep learning fault diagnosis. Therefore, the present invention uses the Euclidean distance of the image texture parameters to determine the parameter values ​​of τ and η.

[0071] The parameter values ​​of the hysteresis coefficient τ and the deflection angle gain η are selected as follows:

[0072] (1) Obtain N analysis signals in normal states and N analysis signals in fault states, where different analysis signals come from different lines.

[0073] (2) Multiple sets of different parameter solutions are obtained by using different combinations of τ and η. In this embodiment, in the range of 1≤τ≤10, 20°≤η≤60°, τ is selected with a step length of 2, and η is selected with a step length of 5°.

[0074] (3) Using each parameter scheme, a point-symmetric Hilbert transform image is constructed for each analytical signal. Each parameter scheme obtains N normal images A from the normal state and N fault images B from the fault state.

[0075] (4) For each parameter scheme, the normal image A and the fault image B are cross-combined and the similarity distance of each group of normal image A and fault image B is calculated respectively: Where T A (n) is the nth texture parameter feature vector of image A. The texture parameter feature vector refers to a total of 16 texture parameter feature vectors obtained by calculating the four statistics of contrast, energy, correlation and homogeneity in the horizontal direction, vertical direction and two diagonal directions of the point-symmetric Hilbert transform image. These texture parameter feature vectors can capture the detail changes and texture patterns in the image, providing strong support for further image analysis and pattern recognition.

[0076] (5) For each parameter scheme, calculate the sum of the corresponding N*N similarity distances D sum =∑D(A,B).

[0077] (6) Finally, the solution with the largest sum of similarity distances is selected as the optimal solution for the values ​​of τ and η.

[0078] The largest similarity distance corresponds to the smallest similarity, and the smallest similarity can reflect the largest difference between the point-symmetric Hilbert transform image of the normal line and the point-symmetric Hilbert transform image of the faulty line.

[0079] Using point-symmetric Hilbert transform, a single-dimensional voltage line mode signal containing amplitude and frequency information can be efficiently converted into a symmetrical pattern with significant shape characteristics. The mirror plane of each symmetrical pattern in the image is defined as an image arm, the thickness of the image arm is represented by the arc length of the arm, and each image arm has its own geometric center position. As the amplitude and frequency of the signal change, the distribution characteristics of the points on the image will also change, resulting in point-symmetric Hilbert transform images of different shapes. This method can not only intuitively display the time-frequency characteristics of the signal, but also further analyze the dynamic behavior characteristics of the signal through the morphological changes of the image.

[0080] Different voltage line mode signals will present different image features after point-symmetric Hilbert transform. These features are mainly reflected in the following aspects: ① The color distribution characteristics of the image can reflect the change trend of the signal amplitude and the difference in frequency components; ② The overall shape characteristics of the image, different signal modes will lead to significantly different graphic shapes; ③ The density of point distribution at the intersection of the image arms, reflecting the characteristics of local changes in the signal; ④ The texture characteristics of the image, including the richness of details and local structural information; ⑤ The thickness of each image arm and the distribution of the geometric center show the characteristics of signal symmetry and amplitude distribution; ⑥ The overlapping area of ​​different parts in the image represents the frequency components and phase relationship of the signal. Therefore, by visualizing the signal in the form of a symmetrical graphic through point-symmetric Hilbert transform, it is possible to reveal deep information that is difficult to detect in the one-dimensional time domain signal. Combined with deep learning technology, the analysis of these image features can realize intelligent diagnosis of line faults, thereby significantly improving the accuracy and reliability of fault identification.

[0081] Step 3: Figure 3 and 4 , the point-symmetric Hilbert transform images of each line are spliced ​​into a complete image, and the complete image is input into the target detection network for target detection. The line corresponding to the detection box marked by the target detection network is the fault circuit.

[0082] The target detection network is a network model improved based on YOLOv5. The target detection network includes a CSPDarknet structure, a feature pyramid network FPN and a detection head YOLO Head connected in sequence. The CSPDarknet structure is used to extract fault features, the feature pyramid network FPN is used to further enhance feature information, and the detection head YOLO Head is used to generate a detection box to mark the fault line.

[0083] The YOLO model is a "single-stage" detection model with the advantages of fast detection speed and high accuracy. YOLOv5 inherits and optimizes the advantages of YOLOv1 to YOLOv4, and has significant improvements in detection speed and accuracy. Therefore, the present invention selects YOLOv5 as the basic framework for fault line selection.

[0084] like Figure 1 Based on YOLOv5, the present invention further introduces three modules: Dilated block, ECA block and Lightweight block. The dilated convolution module is used to replace the SPP Block in YOLOv5, the channel attention module is added after each upsampling and downsampling operation, and the lightweight module composed of depthwise separable convolution is used to replace the original standard convolution layer in the residual block of the CSP module.

[0085] Specifically:

[0086] (1) In view of the limitations and deficiencies of the original YOLOv5 detection model in the detection of weak fault signals in the distribution network fault line selection task, a Dilated Block composed of dilated convolutions was introduced to replace the original SPP Block. This improvement effectively reduces the information loss caused by the pooling operation, and by expanding the network's receptive field, it improves the ability to capture fault characteristics of different scales, and can meet the detection requirements of different types of faults in the distribution network. The introduction of the Dilated Block enables the network to show higher detection accuracy when dealing with small fault signals, solves the problem of insufficient detection accuracy of small fault signals, and improves the reliability of fault location.

[0087] like Figure 2 ,The dilated convolution module consists of a projection layer, several residual modules and a convolution layer connected in sequence.

[0088] The projection layer includes a 1×1 convolution layer and a 3×3 convolution layer, the former is used to reduce the number of channels to half of the original, and the latter is used to further refine the context information.

[0089] The residual module enhances the feature expression capability through the residual structure. It contains three convolutional layers: the first convolutional layer is a 1×1 convolution, which is used to reduce the number of channels to one-fourth of the original; the second convolutional layer is a 3×3 dilated convolution, which is used to expand the receptive field; the third convolutional layer is a 1×1 convolution, which is used to restore the number of channels. The feature map generated by the third convolutional layer is also fused with the input features of the residual module, retaining the expanded and original receptive field information. In order to improve the detection effect of fault features of different scales, the residual block is repeated 4 times in this embodiment, and each dilation convolution uses a different dilation rate, which is set to 2, 4, 6 and 8 respectively.

[0090] The last convolution layer in the dilated convolution module is a 1×1 convolution, which is used to restore the number of channels.

[0091] The multi-scale feature map generated by the dilated convolution module can better adapt to the diversity of distribution network faults, especially significantly improve the detection accuracy of small fault signals.

[0092] All convolutional layers in the dilated convolution module are processed with convolution, BN normalization, and SiLU activation.

[0093] (2) In the multi-scale detection process, a channel attention mechanism (Efficient Channel Attention, ECA), namely the channel attention module, is added.

[0094] The channel attention module adjusts the channel weights so that the network can pay more attention to key feature channels, thereby extracting more accurate fault feature information and improving overall detection performance. This attention mechanism can effectively enhance the network's ability to identify complex fault signals, especially when the fault features are relatively weak or there is a lot of interference, which can significantly improve the robustness and accuracy of detection.

[0095] Specifically, in the improved YOLOv5 network, the channel attention module ECA block is introduced to enhance the ability to extract fault features. The addition of ECA Block is mainly used to strengthen the feature fusion operation in the feature extraction process, especially in the processing of multi-scale features in FPN (Feature Pyramid Network), which plays a key role. In FPN, feature maps of different scales are fused through upsampling and downsampling operations to integrate feature information from different levels to form a comprehensive representation of multiple fault features. These operations help capture both large-scale obvious fault features and local small-scale abnormal signals in distribution network faults. However, traditional feature fusion methods are easily affected by noise, resulting in inaccurate extraction of important features.

[0096] In order to solve the above problems, the present invention adds an ECA channel attention mechanism after each upsampling and downsampling operation. ECA Block automatically adjusts the weights of different channels by weighted processing of feature channels, so that the network can pay more attention to feature information that is critical to fault detection, while suppressing the influence of irrelevant or interfering features. In this way, the accuracy of feature fusion is improved, making the network's detection performance of fault signals in the distribution network more outstanding. The addition of ECABlock can not only enhance the ability to capture multi-scale features, but also improve the robustness of detection in complex environments, especially when small fault signals are difficult to be detected by traditional methods. The detection accuracy is significantly improved. This improvement makes the network more adaptable when dealing with multiple fault modes, and can more accurately select fault lines, thereby improving the efficiency and accuracy of distribution network fault location. In summary, by introducing the ECA channel attention module in the feature extraction network FPN of YOLOv5, the quality of feature fusion and the ability of multi-scale detection are significantly improved, so that the detection performance of distribution network fault line selection is comprehensively optimized.

[0097] (3) Although the YOLOv5-L model has strong fault detection capabilities, it is not suitable for lightweight deployment scenarios in actual distribution network applications due to its large number of parameters. In order to solve this problem, the present invention proposes a lightweight module (Lightweight Block) composed of deep separable convolutions to replace the standard convolution layer in the original network. Lightweight Block effectively reduces the number of parameters and calculations of the model by separating spatial convolution and channel convolution, thereby improving the detection speed. While ensuring the accuracy of fault line selection, the model is more suitable for real-time fault monitoring applications in embedded devices or edge computing environments.

[0098] Specifically, in the YOLOv5-L model, high efficiency and lightweight deployment of detection tasks are crucial. However, the 3×3 ordinary convolution used in the residual block in the original YOLOv5-L model carries most of the computation, resulting in a large number of model parameters and high computational overhead. To solve this problem, this paper introduces a lightweight module (Lightweight Block), which replaces the 3×3 ordinary convolution in the residual block with a 3×3 depth-separable convolution to reduce the computational complexity and model parameters.

[0099] The calculation process of ordinary convolution involves both region and channel operations. First, convolution is performed on each region, and then the convolution results of each channel are accumulated to obtain the output feature map of a single channel. This calculation method makes the convolution operation computationally intensive and requires many model parameters.

[0100] The depthwise separable convolution uses an efficient calculation method to separate the operations of regions and channels. First, a 3×3 depthwise convolution is used to perform regional convolution on each channel independently, without the need for inter-channel accumulation operations. Then, a 1×1 point-by-point convolution is used to perform inter-channel calculations. This point-by-point convolution reduces the amount of calculation by 9 times compared to a 3×3 ordinary convolution, significantly reducing the computational overhead and number of parameters of the model.

[0101] In the fault line selection task, lightweight design is of great significance for improving detection speed and reducing resource consumption. By reasonably adjusting the number of residual blocks and replacing the original 3×3 ordinary convolution with 3×3 depthwise separable convolution, the number of model parameters is significantly reduced from about 107 million (94×10^6) and the detection speed is increased from 25 frames per second to 31 frames per second. This improvement improves the real-time performance of fault detection without significantly affecting the detection accuracy, making the model more suitable for resource-constrained embedded devices and edge computing environments. In the fault line selection of distribution networks, the introduction of the improvement of Lightweight Block not only improves the operating efficiency of the model, but also enables fast and accurate fault location in complex power systems. With the reduction of the model calculation amount, the delay of fault detection and line selection is also greatly reduced, which is conducive to the rapid recovery and stable operation of the distribution network. In addition, this lightweight design provides the possibility for subsequent promotion and application in more fault scenarios, further enhancing the robustness and adaptability of the fault line selection algorithm.

[0102] Experimental comparison:

[0103] In order to verify the effectiveness of the present invention, the method is compared with the traditional impedance measurement method, the traveling wave method, the zero-sequence current method and the conventional machine learning models (such as the support vector machine and the original YOLOv5 model).

[0104] Table 1 - Comparison of experimental data of different methods.

[0105]

[0106] From the experimental results in Table 1, it can be seen that the method of the present invention is significantly higher than the traditional method in terms of fault location accuracy and detection speed. The false alarm rate and missed alarm rate are greatly reduced, especially in complex power grids with multiple branches and multiple feeders. In addition, compared with the original YOLOv5, the present invention achieves low-cost deployment by introducing lightweight modules, which is more suitable for application in resource-constrained edge computing environments.

[0107] It should be noted that it is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. The scope of the present invention is defined by the claims rather than the above description.

Claims

1. A distribution network fault line selection method based on image transformation and target detection network, characterized in that: The method includes: Step 1: After a fault occurs, the three-phase voltages of all lines suspected of having a fault are collected, and the three-phase voltage sampling values ​​collected from each line at each moment are decoupled through voltage phase mode transformation, and the 2-mode component in the voltage phase mode transformation is used as the line mode component of the corresponding line at that moment, thereby obtaining a one-dimensional signal of the line mode component of each line; Step 2, performing point-symmetric Hilbert transform on the one-dimensional signal of the line mode component of each line to obtain a point-symmetric Hilbert transform image of the line; Step 3: Splice the point-symmetric Hilbert transform images of each circuit into a complete image, and input the complete image into the target detection network for target detection. The circuit corresponding to the detection box marked by the target detection network is the fault circuit.

2. The method for fault line selection in distribution network based on image transformation and target detection network according to claim 1, characterized in that: The voltage phase mode conversion process is: in, They are the collected three-phase voltages, They are respectively the converted 0-mode component, 1-mode component and 2-mode component.

3. The method for fault line selection in distribution network based on image transformation and target detection network according to claim 1, characterized in that: Step 2 specifically includes: Step 2.1, convert the one-dimensional signal of the line mode component into an analytical signal: a(t)=x(t)+jH[x(t)]; Where x(t) is the signal value at time t in the one-dimensional signal of the original line mode component, a(t) is the analytical signal value corresponding to x(t), and H[x(t)] is the Hilbert transform result of x(t): Where γ is an auxiliary variable; Step 2.2, mapping the analytical signal to the polar coordinate space in a point-symmetric manner; Step 2.3, draw the mapped points in the polar coordinate space, change a new color every 60 degrees during drawing, and use the obtained color space domain image containing RGB three channels as the point-symmetric Hilbert transform image.

4. The method for fault line selection in distribution network based on image transformation and target detection network according to claim 3, characterized in that: The specific process of step 2.2 is as follows: Assume that the amplitude of the i-th analytical signal value in the analytical signal is a env (i), the instantaneous phase is φ(i), then the polar coordinates of the two points of the analytical signal in the polar coordinate space that are symmetrical with respect to the radial line with an angle of θ are (θ(i), r(i)) and where radius r(i) = a env (i), the deflection angle of the first point The deflection angle of the second point θ is the rotation angle of the mirror symmetry plane, τ is the hysteresis coefficient, η is the deflection angle gain, φ max and φ min are the maximum and minimum values ​​of the instantaneous phase in this group of analytical signals respectively.

5. The method for fault line selection in distribution network based on image transformation and target detection network according to claim 4, characterized in that: The rotation angle θ of the mirror symmetry plane is 60 degrees.

6. The method for fault line selection in distribution network based on image transformation and target detection network according to claim 4, characterized in that: The parameter values ​​of the hysteresis coefficient τ and the deflection angle gain η are selected as follows: (1) Obtain N analysis signals in normal states and N analysis signals in fault states, where different analysis signals come from different lines; (2) Use different combinations of τ and η to obtain multiple sets of different parameter schemes; (3) constructing point-symmetric Hilbert transform images for each analytical signal using each parameter scheme, and obtaining N normal images A from the normal state and N fault images B from the fault state for each parameter scheme; (4) For each parameter scheme, the normal image A and the fault image B are cross-combined and the similarity distance of each group of normal image A and fault image B is calculated respectively: Where T A (n) is the nth texture parameter feature vector of image A; the texture parameter feature vector refers to a total of 16 texture parameter feature vectors obtained by respectively calculating four statistics of contrast, energy, correlation and homogeneity in the horizontal direction, vertical direction and two diagonal directions of the point-symmetric Hilbert transform image; (5) For each parameter scheme, calculate the sum of the corresponding N*N similarity distances D sum =∑D(A,B); (6) Finally, the solution with the largest sum of similarity distances is selected as the solution for determining the values ​​of τ and η.

7. The method for fault line selection in a distribution network based on image transformation and target detection network according to any one of claims 1 to 6, characterized in that: The target detection network is a network model improved based on YOLOv5.

8. The method for fault line selection in distribution network based on image transformation and target detection network according to claim 7, characterized in that: Improvements to YOLOv5 include: using a dilated convolution module to replace the SPP Block in YOLOv5; The dilated convolution module consists of a projection layer, several residual modules and a convolution layer connected in sequence; The projection layer includes a 1×1 convolution layer and a 3×3 convolution layer, the former is used to reduce the number of channels to half of the original, and the latter is used to further refine the context information; The residual module enhances the feature expression capability through the residual structure, and includes three convolutional layers: the first convolutional layer is a 1×1 convolution, which is used to reduce the number of channels to one-fourth of the original number; the second convolutional layer is a 3×3 dilated convolution, which is used to expand the receptive field; the third convolutional layer is a 1×1 convolution, which is used to restore the number of channels; the feature map generated by the third convolutional layer is also fused with the input features of the residual module; the residual block is repeated 4 times, and each dilation convolution uses a different dilation rate, which is set to 2, 4, 6 and 8 respectively; The last convolution layer in the dilated convolution module is a 1×1 convolution, which is used to restore the number of channels.

9. The method for fault line selection in distribution network based on image transformation and target detection network according to claim 7, characterized in that: Improvements to YOLOv5 include adding a channel attention module after each upsampling and downsampling operation.

10. The method for fault line selection in distribution network based on image transformation and target detection network according to claim 7, characterized in that: Improvements to YOLOv5 include: using a lightweight module consisting of depthwise separable convolutions to replace the original standard convolutional layer in the residual block of the CSP module; The processing process of the depthwise separable convolution is as follows: first, a 3×3 depthwise convolution is used to perform regional convolution on each channel independently, and then a 1×1 point-by-point convolution is used to perform calculations between channels.

Images