A flexible direct current power distribution network fault diagnosis method and system

Abstract

The application relates to the technical field of power system fault diagnosis, and discloses a flexible DC power distribution network fault diagnosis method and system. The method comprises the following steps: collecting a fault current time sequence signal of a DC side of a flexible DC power distribution network and inputting the fault current time sequence signal into a pre-trained fault diagnosis model; using an interpretable complex time-frequency convolution layer to perform adaptive time-frequency feature extraction on the fault current time sequence signal, obtaining a time-frequency-amplitude tensor, constructing a time-amplitude trajectory, a time-frequency trajectory and a frequency-amplitude trajectory, and respectively performing Gram difference angle field coding on the trajectories to generate a three-channel time-frequency feature matrix; inputting the three-channel time-frequency feature matrix into a parallel double-branch feature extraction module to extract global features and local features; adaptively fusing the extracted global features and local features, classifying the fused features, and outputting a diagnosis result of a fault category. The application improves the accuracy of flexible DC power distribution network fault diagnosis.

Description

Technical Field

[0001] This application relates to the field of power system fault diagnosis technology, and in particular to a method and system for fault diagnosis of flexible DC distribution networks. Background Technology

[0002] With the rapid development of new energy technologies and the large-scale integration of renewable energy sources such as wind power and solar energy, traditional power systems are gradually transforming into more complex flexible DC distribution networks. Flexible DC distribution networks, due to their flexibility and efficiency, have shown great potential, especially in the power transmission of large-scale wind power, photovoltaic power generation systems, and electric vehicles. However, with the expansion of the distribution network scale and the increase in complexity, fault events occur frequently, particularly on the DC side. How to quickly and accurately detect and diagnose faults has become an urgent technical challenge.

[0003] Traditional fault diagnosis methods for flexible DC distribution networks often rely on manually set thresholds, simple waveform analysis, or mathematical derivations of equivalent models. However, these methods are prone to misdiagnosis or missed diagnosis when faced with different control strategies and load characteristics, and are difficult to adapt to complex and ever-changing system conditions. With the rise of deep learning technology, machine learning-based fault diagnosis methods are gradually becoming an effective alternative. Deep learning can automatically extract key features from complex signals, thereby significantly improving the accuracy and generalization ability of fault identification. However, in practical applications, the training and inference process of deep learning models requires a large amount of labeled data, and how to effectively fuse time-frequency domain features and accurately extract the time-frequency features of current signals remains a research challenge. Furthermore, how to extract global and local features and reasonably fuse them without increasing the computational burden remains an important problem to be overcome in smart grid fault diagnosis.

[0004] Therefore, time-frequency feature extraction and adaptive weighted feature fusion methods based on deep learning have become a research hotspot for solving the problem of DC-side fault diagnosis in flexible DC distribution networks. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and system for fault diagnosis of flexible DC distribution networks. Through innovative complex time-frequency convolution, time-frequency-amplitude tensor generation, and dual-branch feature extraction mechanisms, it can efficiently extract and fuse global and local features, thereby improving the accuracy and robustness of fault diagnosis.

[0006] Firstly, this application provides a fault diagnosis method for flexible DC distribution networks, the method comprising:

[0007] Step S1: Collect fault current timing signals on the DC side of the flexible DC distribution network;

[0008] Step S2: Input the fault current time series signal into the pre-trained fault diagnosis model. The fault diagnosis model uses an interpretable complex time-frequency convolutional layer to perform adaptive time-frequency feature extraction on the fault current time series signal to obtain a time-frequency-amplitude tensor.

[0009] Step S3: Construct time-amplitude trajectory, time-frequency trajectory and frequency-amplitude trajectory based on the time-frequency-amplitude tensor, and perform Gram differential angle field encoding on each trajectory to generate a three-channel time-frequency feature matrix;

[0010] Step S4: Input the three-channel time-frequency feature matrix into the parallel dual-branch feature extraction module, and extract global and local features through the Swin-Transformer branch and the shuffled sensing feature branch of the parallel dual-branch feature extraction module, respectively.

[0011] Step S5: Perform adaptive weighted fusion on the extracted global features and local features, classify the fused features, and output the diagnostic results of the fault category.

[0012] Secondly, this application provides a flexible DC distribution network fault diagnosis system, the system comprising:

[0013] The data acquisition module is used to acquire fault current timing signals on the DC side of the flexible DC distribution network;

[0014] The feature extraction module is used to input the fault current time series signal into a pre-trained fault diagnosis model. The fault diagnosis model uses an interpretable complex time-frequency convolutional layer to perform adaptive time-frequency feature extraction on the fault current time series signal to obtain a time-frequency-amplitude tensor.

[0015] The feature transformation module is used to construct time-amplitude trajectory, time-frequency trajectory and frequency-amplitude trajectory based on the time-frequency-amplitude tensor, and to perform Gram differential angle field encoding on each trajectory to generate a three-channel time-frequency feature matrix;

[0016] The feature fusion module is used to input the three-channel time-frequency feature matrix into the parallel dual-branch feature extraction module, and extract global features and local features through the Swin-Transformer branch and the shuffled sensing feature branch of the parallel dual-branch feature extraction module, respectively.

[0017] The fault identification module is used to adaptively weight and fuse the extracted global features and local features, classify the fused features, and output the diagnostic results of the fault category.

[0018] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0019] This application provides a fault diagnosis method and system for flexible DC distribution networks. It adopts an adaptive time-frequency feature extraction method, which performs in-depth analysis of the time-frequency features of the fault current signal through complex convolution kernels, effectively capturing the dynamic changes in different frequency bands. Secondly, by constructing time-amplitude trajectories, time-frequency trajectories, and frequency-amplitude trajectories and performing Gram differential angle field encoding, the time-frequency features of the signal can be transformed into a more easily analyzed form, making the fault features clearer. This multi-dimensional time-frequency feature extraction method not only improves the model's ability to identify different types of faults, but also enhances the system's adaptability under complex operating conditions.

[0020] Then, by employing a parallel structure of Swin-Transformer and shuffled sensing feature branches, combined with weighted fusion of global and local features, the model's ability to capture information at different levels of the signal is further enhanced. Through adaptive weighted fusion of global and local features, the contributions of different features can be balanced, thereby improving the accuracy of fault diagnosis. Without introducing additional convolutional structures, this method effectively controls computational load and inference latency, making it suitable for online real-time fault diagnosis in flexible DC distribution networks. Finally, this method uses a classifier to determine the fault category of the fused features, enabling rapid and accurate identification of different types of faults, ensuring the efficient operation and safety of the system. Therefore, this application can significantly improve the fault diagnosis performance of flexible DC distribution networks, reduce equipment damage and downtime, and enhance system stability and reliability. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a fault diagnosis method for a flexible DC distribution network according to an embodiment of this application;

[0023] Figure 2 This is a schematic diagram of an interpretable complex time-frequency convolutional layer in an embodiment of this application;

[0024] Figure 3 This is a structural diagram of the Swin-Transformer module in the embodiments of this application;

[0025] Figure 4 This is a diagram of the feature fusion and classification decision module in the embodiments of this application;

[0026] Figure 5This is a topology diagram of a dual-ended flexible DC distribution network in the embodiments of this application;

[0027] Figure 6 This is a schematic diagram showing the curves of verification loss and accuracy varying with batch size in the embodiments of this application;

[0028] Figure 7 This is a schematic diagram of the normalized confusion matrix in the embodiments of this application;

[0029] Figure 8 This is a schematic diagram of the structure of a flexible DC distribution network fault diagnosis system according to an embodiment of this application. Detailed Implementation

[0030] This application provides a method and system for fault diagnosis of flexible DC distribution networks. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0031] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of a fault diagnosis method for a flexible DC distribution network in this application includes:

[0032] Step S1: Collect fault current timing signals on the DC side of the flexible DC distribution network.

[0033] Specifically, flexible DC distribution networks are a new type of power distribution system that uses DC power transmission technology. Through an efficient and flexible DC grid architecture, it connects different power sources and loads efficiently, making it suitable for modern power networks such as distributed generation, renewable energy integration, and energy storage systems. Compared with traditional AC distribution networks, flexible DC distribution networks have many advantages, such as low loss, fast power regulation, flexible voltage control, and adaptability to renewable energy integration. Flexible DC distribution networks typically consist of DC buses, converters, DC loads, and transformation devices connected to the AC system. They possess advantages such as high efficiency, flexibility, and strong adaptability. However, because DC-side faults have a significant impact on the system, the diagnosis and handling of DC-side faults are crucial. DC-side faults include, but are not limited to, the following common types: single-pole positive grounding, inter-pole short circuit, or single-pole negative grounding.

[0034] To better identify DC-side faults in flexible DC distribution networks and achieve real-time monitoring, analysis, location, classification, and system recovery of fault types, ensuring a rapid response from the power grid to avoid prolonged outages and equipment damage, this application deploys high-precision current sensors at key locations such as the DC bus, converter input, and load terminals of the flexible DC distribution network. These sensors collect fault current timing signals from the DC side. The sensors possess high response speed and accurate measurement capabilities, capturing instantaneous current changes during fault occurrence. The collected analog current signals are converted into digital signals by high-speed data acquisition equipment, with a sampling frequency set to 4kHz, ensuring each data point has a precise timestamp. To improve data quality, the collected signals can be preprocessed, such as using filters to eliminate noise interference. The system can be configured for multi-channel acquisition and can set fault trigger conditions. When current fluctuations exceed a preset threshold, it automatically records current waveform data within a set time window (e.g., 0.3 seconds) before and after the fault, providing complete information for subsequent diagnosis.

[0035] Step S2: Input the fault current time series signal into the pre-trained fault diagnosis model. The fault diagnosis model uses an interpretable complex time-frequency convolutional layer to perform adaptive time-frequency feature extraction on the fault current time series signal to obtain the time-frequency-amplitude tensor.

[0036] In step S2, the adaptive time-frequency feature extraction of the fault current time-series signal includes: constructing a complex convolution kernel structure consisting of F parallel channels, each channel corresponding to a set of physically meaningful frequency bands with trainable center frequencies; for each channel, performing convolution operations on the fault current time-series signal using a complex convolution kernel function, which consists of a real part and an imaginary part, and is defined as: ,in, For complex convolution kernel function, and These represent the real and imaginary parts of the corresponding kernel function, respectively, where j is the imaginary unit. To enable learnable parameter generation and constraint by physical boundaries, adaptive learning updates are performed during training, including the center frequency and window width scaling factor. For each channel, a modulo operation is performed on the real and imaginary convolution results to obtain the channel's frequency band amplitude response time series. ,in, Let the amplitude response time series be the frequency band k. Let be the combined amplitude of the real and imaginary parts of the k-th frequency band at time n. ;in, , These are the results of the k-th frequency band after convolution with the real part convolution kernel and the imaginary part convolution kernel at time n, respectively. Where N is the kernel length, Indicates the input signal The result of shifting the value in the time domain, where m is the index of the shift. For the k-th frequency band, at discrete time m, given the parameters... The real part of the kernel function is determined. The learnable parameters for the k-th frequency band are constrained by physical boundaries; ,in, For the k-th frequency band, at discrete time m, given the parameters... The imaginary part of the kernel function is determined; the frequency band amplitude response time series of F channels are stacked to obtain the discrete time-frequency-amplitude tensor H. .

[0037] Specifically, the fault current time-series signal is input into a pre-trained fault diagnosis model. This model includes an interpretable complex time-frequency convolutional layer specifically designed for adaptive extraction of time-frequency features from the current signal, such as... Figure 2 The diagram illustrates an interpretable complex time-frequency convolutional layer. To achieve this, the model's time-frequency convolutional layer consists of multiple parallel complex convolutional kernel channels. Each channel corresponds to a set of physically meaningful frequency bands, and the center frequency of each band is trainable. In this convolutional layer, the complex convolutional kernel function consists of a real part and an imaginary part, in the form: , These are learnable parameters that are adaptively learned and updated during training. The training parameters of the complex convolution kernel include the center frequency and the window width scaling factor. The center frequency is the frequency at which the complex convolution kernel focuses in the frequency domain, which determines the frequency range of the convolution kernel's response. The window width scaling factor controls the spectral width of the convolution kernel and affects the fineness of the frequency resolution. The center frequency helps the convolution kernel focus on a specific frequency band, while the window width scaling factor balances frequency and time resolution. By adaptively learning these two parameters, the complex convolution kernel can efficiently extract key time-frequency features from the signal and improve the accuracy of fault diagnosis.

[0038] Each complex convolution kernel performs a convolution operation on its corresponding channel, convolving the input fault current timing signal to obtain the convolution result. Specifically, the input signal... The convolution kernel is shifted in the time domain, with a length of N and a shift index of m. The convolution operation yields the real and imaginary parts of each channel. During this process, for each channel's real and imaginary convolution result, a modulo operation is performed to obtain the frequency band amplitude response time series of that channel, represented as... For each convolutional channel, the frequency band amplitude response time series obtained in this way represents the channel's response to the fault current time-series signal in both the time and frequency domains. Subsequently, the frequency band amplitude response time series of all F channels are stacked to obtain a discrete time-frequency-amplitude tensor. This tensor contains the time-frequency characteristics of all channels. By extracting these time-frequency characteristics, the time-frequency properties of the fault current timing signal are fully revealed, providing rich feature information for subsequent fault diagnosis and analysis.

[0039] The complex convolution kernel function is a continuous-time complex kernel function based on the short-time Fourier transform, defined as: ,in, This is a continuous-time complex function, where t is the current time. Here, f is the window width scaling factor, f is the center frequency, and j is the imaginary unit; the continuous-time complex function is discretized to obtain the discrete-time complex function: , where n is the discrete-time series index, and an odd number is chosen to ensure symmetry. The real and imaginary parts of the discrete-time kernel function constitute the real and imaginary parts of the complex convolution kernel function, respectively.

[0040] Specifically, due to the complex time-frequency characteristics of fault current signals in flexible DC distribution networks, traditional time-domain or frequency-domain methods struggle to simultaneously capture both instantaneous changes and spectral features of the signal. To enhance feature extraction capabilities in the time-frequency domain, this application employs a continuous-time kernel design based on Short-Time Fourier Transform (STFT). The STFT complex convolution kernel, combined with a Gaussian window function and a complex exponential term, adaptively extracts the time-frequency features of the signal at different frequency components, enhancing the time-frequency analysis capability of the fault signal and improving the accuracy and robustness of fault diagnosis. Specifically, the continuous-time kernel function is: , The continuous-time kernel function represents a complex signal. The window width scaling factor controls the balance between frequency and temporal resolution, while f is the center frequency, representing the main frequency components of the convolution kernel response, the complex part of which... It is a complex exponential function that describes the frequency components of a signal, while It is a Gaussian window function that controls the time localization characteristics of the convolution kernel. The entire function combines a Gaussian window and a complex exponential term, which can effectively perform time-frequency analysis of signals. The real and imaginary parts of this complex convolution kernel are composed of the real and imaginary parts of the Gaussian window and the complex exponential function, respectively. The Gaussian window ensures the localization of the signal in the time domain, while the complex exponential term determines the frequency characteristics of the signal. Through this complex convolution kernel, the frequency features of the signal can be adaptively extracted in the time and frequency domains. Especially in the fault diagnosis task of flexible DC distribution networks, the complex convolution kernel can effectively capture signal components of different frequencies.

[0041] To apply this complex convolution kernel to discrete-time signals, discretization is performed, transforming the continuous-time complex kernel function into a discrete-time complex kernel function suitable for digital computation. The discretized complex convolution kernel function is defined as follows: Where n is the index of the discrete time series, and its value ranges from 1 to 2. Let represent the discrete-time sequence of the signal, N be the length of the convolution kernel (an odd value N is chosen to ensure kernel symmetry and centered symmetry in the time domain), and f be the center frequency of the convolution kernel. The Gaussian window term in the discretization process Used to control the balance between frequency and time resolution, smaller A smaller value will make the convolution kernel more concentrated in the frequency domain, but the temporal resolution will be poor; while a larger value will result in a smaller value. The value of this value gives the convolution kernel a wider bandwidth response, enhancing its sensitivity to wideband signals. The real and imaginary parts of the discretized complex convolution kernel are derived from... The real and imaginary parts of the time-frequency response are convolved, and the results are calculated using complex convolution operations to obtain the frequency band amplitude response time series for each channel. Stacking these time series yields a discrete time-frequency-amplitude tensor. This complex convolution kernel design based on short-time Fourier transform enables adaptive extraction of time-frequency features from signals. Especially in complex current signals, it can accurately capture fault signal characteristics across different frequency bands, providing strong support for fault diagnosis in flexible DC distribution networks.

[0042] Step S3: Construct time-amplitude trajectory, time-frequency trajectory and frequency-amplitude trajectory based on time-frequency-amplitude tensor, and perform Gram differential angle field encoding on each trajectory to generate a three-channel time-frequency feature matrix.

[0043] In step S3, constructing the time-amplitude trajectory includes: averaging the time-frequency-amplitude tensor along the frequency band dimension to obtain the average amplitude of the signal at the nth discrete time point across all frequency components. The calculation formula is: Where n is the discrete time series index, k is the frequency band index, and F is the number of frequency bands. For the amplitude corresponding to the k-th frequency band at discrete-time index n, the average amplitudes at multiple time points are combined according to the time series to generate a time-amplitude trajectory, represented as: T represents the total length of the time series, and the time-amplitude trajectory reflects the overall energy trend of the signal over time under the combined energy of all frequency bands.

[0044] Specifically, in the fault diagnosis of flexible DC distribution networks, current signals often have complex time-frequency characteristics. Especially when a fault occurs, the signal exhibits abnormal fluctuations in both time and frequency. Traditional fault diagnosis methods struggle to simultaneously capture the signal's time-domain changes and frequency-domain features. Constructing time-amplitude, time-frequency, and frequency-amplitude trajectories based on the time-frequency-amplitude tensor can comprehensively describe the signal's time-frequency characteristics from multiple dimensions, thereby effectively extracting key features of the fault signal. The time-amplitude trajectory reflects the overall energy trend of the signal over time, the time-frequency trajectory reveals the dynamic changes of the signal's frequency components over time, and the frequency-amplitude trajectory provides the energy distribution of the frequency band across the entire time domain. Gram differential angular field coding (GADF) on these trajectories can further enhance the expressiveness of the features, capture the relationship between the signal at different times and frequencies, and generate a three-channel time-frequency feature matrix. This matrix can provide rich and efficient input for subsequent deep learning models, improving the accuracy and robustness of fault diagnosis.

[0045] The process of constructing the time-amplitude trajectory is as follows: Based on the time-frequency-amplitude tensor H, it is processed along the frequency band dimension. In this process, the mean of the signal at each time point n of the time-frequency-amplitude tensor is first calculated. Specifically, for each discrete time point n, the mean of the amplitude across all frequency bands F at that time is calculated. By averaging the amplitude at each time point, the comprehensive amplitude of the signal at each time point n across all frequency bands can be obtained, thus effectively representing the energy distribution of the signal in each frequency band. After the calculation is completed, the average amplitude TA[n] at multiple times n is combined in time series to obtain the complete time-amplitude trajectory TA. This time-amplitude trajectory reflects the comprehensive trend of the signal's energy across all frequency bands and reveals the overall energy change of the signal over time. Through this trajectory, the frequency component changes of the signal at different time points and the energy change characteristics of the signal can be obtained. This is very important for subsequent fault diagnosis because the characteristics of fault signals usually cause abnormal fluctuations in amplitude at different time points. For example, at a certain time n, a single-pole grounding fault occurs, and the amplitude of some frequency bands of the current may rise or fall rapidly. This fluctuation will be manifested in multiple frequency bands. By averaging the amplitudes of these frequency bands, a comprehensive time-amplitude trajectory can be obtained. This trajectory can effectively reflect the abnormal fluctuations of the current signal when the fault occurs, providing key features for subsequent classification and fault diagnosis. The implementation of this process helps to convert the frequency band energy information in the time-frequency signal into easily analyzable time-amplitude features, providing a powerful feature input for subsequent fault diagnosis. By merging the amplitudes of all frequency bands into a single time-amplitude sequence, the influence of noise can be effectively reduced, the overall trend of the signal can be strengthened, and the accuracy of subsequent feature extraction and classification can be further improved.

[0046] In step S3, constructing the time-frequency trajectory includes: calculating the Softmax frequency weight along the frequency band for each time n, and obtaining the weight coefficients of the time-frequency feature corresponding to the k-th frequency band at the discrete time index n. The calculation formula is: ,in, This is the frequency centroid smoothing factor, used to control the smoothness of the normalized weight distribution, and is set to a fixed value of 0.3. The amplitude corresponding to the j-th frequency band at discrete-time index n; the time-frequency features at F frequency bands are weighted using Softmax. We perform a weighted summation to obtain the time-frequency characteristic value of the signal at time n. The calculation method is as follows: The time-frequency feature values ​​are sorted according to the time series to generate a time-frequency trajectory (TF). The time-frequency trajectory reflects the main frequency trajectory that changes over time.

[0047] Specifically, the process of constructing the time-frequency trajectory is as follows: First, calculations are performed based on the time-frequency-amplitude tensor H. The signal features at each time point n are weighted according to the amplitude of each frequency band to obtain the Softmax frequency weight. The purpose of this process is to quantify the relative importance of each frequency band at that moment, ensuring that frequency bands with larger amplitudes are assigned higher weights in the calculation of time-frequency features. This is achieved by calculating the Softmax frequency weight. It can adaptively adjust the contribution of each frequency band to the time-frequency characteristics based on the amplitude of each band. This weight value is obtained by normalizing the amplitude of all frequency bands after exponential smoothing, ensuring the relative importance of the contribution of each frequency band. Based on the calculated Softmax frequency weight, the time-frequency characteristics of each frequency band are... Its corresponding weight Multiply the results and sum them over all frequency bands to obtain the time-frequency characteristic value of time n. This weighted summation process, by comprehensively considering the information from each frequency band, yields a value reflecting the overall characteristics of the signal at that moment. In this way, information from different frequency bands can be effectively fused, ensuring that the influence of each frequency band on the overall time-frequency characteristics is adjusted according to its importance. Assume that at time n, the time-frequency characteristics of the three frequency bands of the signal are as follows: The frequency band weights obtained through Softmax operation are respectively Then the time-frequency characteristic value at that moment This is a weighted sum of the features of these three frequency bands. Through weighting, the features of important frequency bands will have a greater impact on the overall time-frequency characteristics. Finally, this process is repeated for all times TF. This trajectory reflects the frequency component characteristics of the signal as it changes over time, revealing how the signal's frequency characteristics change over time. Especially when a fault occurs, the signal's frequency characteristics may experience abrupt changes or abnormal fluctuations. Therefore, the time-frequency trajectory is crucial for fault diagnosis, helping to detect the frequency change characteristics caused by the fault, thus providing a key basis for subsequent fault classification. In this way, the time-frequency trajectory can effectively capture important changes in the signal in the time and frequency domains, especially those abrupt changes in frequency components caused by the fault. It provides important time-frequency characteristics of the fault signal, helping to improve the accuracy and robustness of the fault diagnosis model.

[0048] Step S3, which involves constructing the frequency-amplitude trajectory, includes: calculating the mean amplitude along the time dimension for each frequency band to obtain the mean amplitude corresponding to the k-th frequency band. The calculation formula is as follows: The average amplitude values ​​corresponding to the F frequency bands are combined to generate the frequency-amplitude trajectory FA. The frequency-amplitude trajectory reflects the average energy distribution of each frequency band across the entire time domain.

[0049] Specifically, the process of constructing the frequency-amplitude trajectory is as follows: For each frequency band k, the mean of the amplitude H[n,k] of that frequency band is calculated along the time dimension. By averaging the amplitude of each frequency band at all time points, the average amplitude of that frequency band is obtained. The amplitude mean represents the energy distribution of a frequency band across the entire time domain. These amplitude mean values ​​reflect the average energy intensity of each frequency band over the entire time period, helping to identify which frequency bands dominate over long periods or which frequency bands experience significant energy changes over different time periods. After calculating the amplitude mean of all frequency bands, these amplitude mean values ​​are combined to obtain the complete frequency-amplitude trajectory FA. This trajectory reflects the average energy distribution of each frequency band of the signal across the entire time domain, helping to identify the long-term stability or trend of the signal across different frequency bands. By constructing the frequency-amplitude trajectory, it is possible to understand the overall energy change of each frequency band over time. For example, if the amplitude of some frequency bands increases significantly over a period of time, this may indicate abnormal energy fluctuations in that band, thus indicating the possibility of a fault. The frequency-amplitude trajectory provides crucial information for subsequent fault diagnosis, especially by identifying how the frequency components of the signal change over time, thereby further helping to distinguish between normal operation and fault conditions.

[0050] Step S3, generating the three-channel time-frequency feature matrix, includes: resampling the time-amplitude trajectory, time-frequency trajectory, and frequency-amplitude trajectory to length L using linear interpolation, as shown below: , and L is the length of each trajectory; for the resampled time-amplitude trajectories Normalization is performed to obtain the time-amplitude trajectory at time x. The processing formula is: , and They are respectively The minimum and maximum values ​​in the time-amplitude trajectory will be normalized. The angle is obtained by performing an inverse cosine transform, and the transform formula is: ,in, for The angle value at time x. To normalize Perform an inverse cosine transform. Let x be the polar radius of the polar coordinate axis at time x, and construct the trajectory relationship matrix based on the angle. : ,in, Let i be the angle value at time i; for and After normalization and inverse cosine transform, the corresponding trajectory relationship matrix is ​​obtained. , , combined Each trajectory relationship matrix is ​​standardized, and the standardized trajectory relationship matrices are then fused to generate a three-channel time-frequency feature matrix.

[0051] Specifically, for each trajectory—namely, the time-amplitude trajectory TA, the time-frequency trajectory TF, and the frequency-amplitude trajectory FA—a linear interpolation method is used to resample it, unifying the length of each trajectory to L, thus obtaining a new trajectory. , and Where L is the target length of each trajectory, resampling ensures the temporal consistency of the three trajectories in subsequent processing, allowing them to be compared and fused on the same scale. Next, for the time-amplitude trajectories... Normalization is performed, scaling the values ​​to the range of [-1, 1]. Normalization eliminates amplitude differences between different trajectories, making subsequent processing more robust and consistent. The normalized time-amplitude trajectory... For each time step x, normalization aims to make the amplitude changes of the signal more representative, facilitating subsequent angle transformation and feature extraction. The normalized time-amplitude trajectory... An inverse cosine transform is performed to obtain the angle value corresponding to each time step x. This angle value represents the relative position of the signal amplitude within the normalized range at that time. The purpose of the inverse cosine transform is to convert amplitude information into angle information, thereby transforming the time-frequency characteristics of the signal into an angle space that is easier to compare. The Gram difference angular field, i.e., the trajectory relationship matrix, is then constructed using these angle values. This matrix describes the relative relationships between different time points, calculated by examining each pair of time points. and The sine value of the angular difference between the two can effectively capture the similarity between time points. The role of Gram angle field coding is to transform the time-frequency characteristics of the signal into a symmetric matrix, which can establish meaningful correlations between time-frequency characteristics and help reveal the dynamic changes of the signal at different time points.

[0052] Similarly, after normalizing and performing inverse cosine transform on the time-frequency trajectory and the frequency-amplitude trajectory, the corresponding trajectory relationship matrix is ​​obtained. , These matrices represent the relative relationships between time-frequency and frequency-amplitude characteristics in the time-frequency domain, respectively. Finally, these three trajectory relationship matrices... , , Standardization is performed to ensure that the features have the same scale and dimensions. The purpose of standardization is to ensure that features from different sources can be compared on the same scale, avoiding excessive influence of certain features on the model due to differences in dimensions. Finally, the three standardized trajectory relationship matrices are fused together to generate a three-channel time-frequency feature matrix. Each channel corresponds to a different time-frequency feature: time-amplitude, time-frequency, and frequency-amplitude. This fusion method concentrates multiple aspects of the signal's time-frequency features into a single matrix, providing rich input for subsequent feature extraction and classification tasks. This three-channel time-frequency feature matrix generation process helps to comprehensively display the signal's features in the time-frequency domain, enabling the fault diagnosis model to simultaneously capture key information of the signal from multiple dimensions, especially effectively identifying frequency and amplitude changes caused by faults, thereby improving the accuracy and robustness of fault diagnosis.

[0053] Step S4: Input the three-channel time-frequency feature matrix into the parallel dual-branch feature extraction module, and extract global and local features through the Swing-Transformer branch and the shuffling-sensing feature branch of the parallel dual-branch feature extraction module, respectively.

[0054] Step S4 includes: The Swin-Transformer branch adopts a hierarchical structure and includes a sliding window multi-head self-attention mechanism. The three-channel time-frequency feature matrix is ​​input into the Swin-Transformer branch for feature modeling. The resolution is halved and the number of channels is doubled through image merging operations between stages, and global semantic information across regions is obtained to obtain global features. The shuffling perception feature branch includes a lightweight convolutional neural network and a shuffling attention perception module. By integrating spatial attention and channel attention mechanisms and combining channel shuffling operations, it selectively responds to local texture, edge and micro-perturbation features of the image and interacts with cross-channel information to obtain local features.

[0055] Specifically, the three-channel time-frequency feature matrix is ​​input to the parallel dual-branch feature extraction module, which extracts the global and local features of the signal through the Swing-Transformer branch and the shuffling sensing feature branch, respectively. Figure 3The diagram shows the Swin-Transformer module structure. The Swin-Transformer branch adopts a hierarchical structure, with each layer containing a sliding window multi-head self-attention mechanism. This mechanism can capture the global features of the signal at various scales. Through this self-attention mechanism, the model can learn the relationship between different regions of the input features to obtain global semantic information across regions. In each layer, the image merging operation halves the resolution and doubles the number of channels, which helps to improve the model's performance in high-dimensional feature spaces. These operations ensure that global information is effectively modeled at multiple levels, enabling the Swin-Transformer branch to extract the global features of the signal, such as the main frequency components and global patterns. On the other hand, the shuffling-based perceptual feature branch includes a lightweight convolutional neural network (CNN) and a shuffling-attention perceptual module, focusing on the extraction of local features. This branch integrates spatial attention and channel attention mechanisms, enabling selective responses to features such as local textures, edges, and minor perturbations in the image. Through channel shuffling, the shuffling-based perceptual module further enhances the interaction of cross-channel information, allowing for the effective fusion of local information. These operations help the model better capture detailed features in the signal, especially when minor perturbations occur, as changes in local features often reveal early signs of faults. Through these two parallel branches, the model can simultaneously extract global and local features. Global features capture the overall pattern and major changes in the signal, while local features emphasize minor changes in details. The combination of the two provides rich and multi-layered information for subsequent fault classification, thereby improving the accuracy and robustness of fault diagnosis.

[0056] Step S5: Adaptively weightedly fuse the extracted global and local features, classify the fused features, and output the diagnostic results of the fault category.

[0057] Step S5 includes: using an adaptive average pooling operator to pool the global features Align to local features Within the spatial scale, the aligned global features are obtained. Set up a lightweight weighting mechanism to apply the aligned global features. and local features Perform global average pooling to obtain the corresponding channel feature factors. Channel feature factors Perform concatenation along the channel dimension to obtain the concatenated result. Furthermore, a linear mapping is applied to the concatenated result u to obtain the logarithmic weights a and b of the global and local branches. , in, These are the weighting coefficients. The bias parameter is used; a temperature-normalized softmax operation is performed on the logarithmic weights of the global and local branches to normalize the logarithmic weights to the branch weights. The normalization formula is: ,in, For temperature coefficient; use the calculated branch weights to align the global features. and local features Weighted fusion is performed to obtain fusion features. The fused features are input into a global average pooling layer and a linear classifier to obtain the diagnostic results of the fault category.

[0058] Specifically, such as Figure 4 The diagram shows the feature fusion and classification decision module, which uses an adaptive average pooling operator to process global features. Perform spatial scale alignment to align it with local features. With consistent spatial scale, the adaptive average pooling operator can adjust feature maps of different sizes to the same spatial scale, ensuring that global and local features can be compared and fused at the same scale. In this way, the scale of global features can be adjusted to align them with local features in the feature space, resulting in aligned global features. A lightweight weighting mechanism is employed to perform global average pooling on the aligned global and local features. Global average pooling averages the spatial dimensions of each feature map, yielding the global average value for each feature channel, and thus the corresponding channel feature factors. These factors represent the information weights of global and local features in each channel dimension. The channel feature factors g and... The concatenation is performed along the channel dimension to obtain the concatenated result. In this way, the weighted information of global and local features is merged into a unified feature vector. Then, a linear mapping is performed on the concatenated feature vector to obtain the logarithmic weights a and b of the global and local branches, where... , Weighting coefficient and and bias parameters and These features are automatically learned during the training process and optimized using backpropagation and gradient descent. This allows global and local features to be weighted and adjusted according to their importance during fusion. The Softmax operation then uses these weights and biases to calculate the final branch weights. and through temperature coefficient Smooth the temperature coefficients to ensure they are between (0, 1) and sum to 1. By gradually reducing the weights through an annealing strategy to balance exploration and stability, the importance of global and local features is effectively adjusted, improving the accuracy and robustness of the fault diagnosis model. Finally, the calculated branch weights are used to weightedly fuse the aligned global and local features, resulting in a final fused feature, Q. This fused feature Q integrates the weighted information of both global and local features, providing a more comprehensive representation of the signal's time-frequency characteristics. The fused feature is then input into a global average pooling layer and a linear classifier. The classifier diagnoses the fault category and outputs the final fault category result. Assuming a flexible DC distribution network system, after feature extraction and fusion, the fault signal may contain time-frequency features such as amplitude, frequency variations, and local disturbances. Through training, the linear classifier learns the relationship between different fault types, such as single-pole grounding faults, inter-pole short circuits, and overcurrent, and these time-frequency features. For example, a single-pole grounding fault may cause a sudden increase in amplitude in certain frequency bands, while an inter-pole short-circuit fault may manifest as simultaneous and severe fluctuations in amplitude across multiple frequency bands. The classifier uses these characteristic patterns to determine the fault type and outputs the fault category, such as "single-pole grounding fault" or "inter-pole short-circuit fault." This classification process can automatically determine the fault type of the signal at the current moment based on the learning of various fault types during training, thereby performing real-time diagnosis and helping the system to respond and repair faults quickly.

[0059] To verify the effectiveness and engineering feasibility of the DC-side fault diagnosis method for flexible DC distribution networks proposed in this application, a simulation platform based on PSCAD electromagnetic transient simulation was constructed, such as... Figure 5 The diagram shows the topology of the dual-end flexible DC distribution network. The dual-end flexible DC distribution network model systematically includes key components such as AC power supply, VSC converter, DC transformer, AC / DC load modules, and multiple DC transmission lines. The specific simulation parameters are configured as shown in Table 1. The AC power supply voltage is set to 10kV and the capacity is 15MVA. The VSC converter capacity is 7MVA. The DC line length is uniformly set to 5km. The DC load and AC load are 3MW and 4MW, respectively. These parameter settings fully consider the typical scenarios of actual engineering applications and provide a reliable simulation environment for the subsequent generation of fault data.

[0060]

[0061] Table 1 Simulation Model Parameters

[0062] Based on the completed model construction, three typical faults—single-pole positive grounding, single-pole negative grounding, and inter-pole short circuit—were set at key locations such as the near, middle, and far ends of the five DC lines in the simulation model. Simultaneously, current signals under normal system operation were collected, forming four sample sets for different operating conditions. Fault data were collected using a sampling frequency of 4kHz, recording the dynamic waveform of the DC line current within a 0.3-second time window after the fault occurred. Each sample had a fixed length of 1200 data points, resulting in a final dataset containing 500 samples for each operating condition, totaling 4000 sets of data. The detailed composition is shown in Table 2. To ensure the scientific rigor of subsequent model training and evaluation, all samples were randomly divided into three independent subsets—training set, validation set, and test set—in a 7:2:1 ratio.

[0063]

[0064] Table 2. Sample Dataset Details

[0065] During the model training phase, a batch size of 64 and a training epoch of 50 were used as the basic settings, and AdamW was selected as the optimizer with a base learning rate of 100%. This learning rate applies to all network parameters except those of the interpretable complex time-frequency convolutional layers. To further optimize the training process, a weight decay is introduced. A regularization strategy is employed, and a cosine annealing learning rate scheduler is used to dynamically adjust the learning rate. The loss function uses a cross-entropy function with a label smoothing coefficient of 0.05. Gradient norm clipping (up to 1.0) is applied to all learnable parameters during training, and CUDA mixed-precision computation is enabled to significantly improve training efficiency. To address the unique characteristics of training complex convolution kernel parameters, a phased unfreezing strategy is designed: the parameters are frozen in the first 7 training batches, and automatically unfrozen from the 8th batch onwards, with their learning rate set to... This strategy effectively ensures stability in the early stages of training and adaptive optimization of frequency parameters in subsequent training. To suppress overfitting, an early stopping mechanism is also employed during training, with a patience value of 15 rounds and a minimum effective improvement threshold of [value missing]. Training is automatically terminated when the performance on the validation set does not show a significant improvement over multiple consecutive rounds, and the current optimal model weights are saved each time the validation accuracy improves.

[0066] like Figure 6As shown, to verify the curves of loss and accuracy changing with batch size, the model's loss value rapidly decreased from approximately 1.10 after the first training round, and stabilized below 0.20 after the fifth round. The validation set loss exhibited a good trend of synchronous convergence and stabilization. Regarding accuracy, the training accuracy steadily increased from an initial 0.5864 to over 0.99, and the validation accuracy reached 0.9550 in the second round, reaching its optimal value of 0.9775 in the twelfth round. Importantly, both the loss curve and validation set performance remained stable in the later stages of training without significant fluctuations. This indicates that the model has excellent convergence stability and has not experienced overfitting, and also verifies the effectiveness of the phased unfreezing strategy for asymptotically optimizing the frequency parameters of the complex convolution kernel.

[0067] To objectively evaluate the final performance of the model, this application also loads the optimal weights from the validation set and performs comprehensive testing on an independent test set, such as... Figure 7 The normalized confusion matrix results show that 100% accurate identification was achieved for all three types of samples: normal operating conditions, unipolar positive grounding faults, and inter-pole short-circuit faults. The identification accuracy for unipolar negative grounding faults also reached 98%. The final model achieved an overall accuracy of 99.5% on the test set, with a test loss of only 0.0605. This result fully demonstrates that the proposed method can clearly distinguish the characteristic differences of different fault types and has a stable and efficient identification capability for various fault conditions.

[0068] Based on the above experimental results, the fault diagnosis method proposed in this application, based on interpretable complex time-frequency convolution and bi-branch feature fusion, demonstrates outstanding advantages in terms of fast convergence, high-precision classification, and strong generalization ability under limited sample conditions. Through an end-to-end trainable time-frequency analysis front-end and a carefully designed parallel fusion architecture, this method not only significantly improves the accuracy and reliability of fault diagnosis in flexible DC distribution networks, but also shows outstanding practical value in terms of computational efficiency and engineering adaptability, providing an effective technical solution for the intelligent operation and maintenance of complex DC distribution networks.

[0069] The above describes a fault diagnosis method for a flexible DC distribution network according to an embodiment of this application. The following describes a fault diagnosis system for a flexible DC distribution network according to an embodiment of this application. Please refer to [link / reference]. Figure 8 One embodiment of a flexible DC distribution network fault diagnosis system in this application includes:

[0070] The data acquisition module is used to acquire fault current timing signals on the DC side of the flexible DC distribution network;

[0071] The feature extraction module is used to input the fault current time series signal into the pre-trained fault diagnosis model. The fault diagnosis model uses an interpretable complex time-frequency convolutional layer to perform adaptive time-frequency feature extraction on the fault current time series signal to obtain the time-frequency-amplitude tensor.

[0072] The feature transformation module is used to construct time-amplitude trajectory, time-frequency trajectory and frequency-amplitude trajectory based on time-frequency-amplitude tensor, and to perform Gram differential angle field encoding on each trajectory to generate a three-channel time-frequency feature matrix;

[0073] The feature fusion module is used to input the three-channel time-frequency feature matrix into the parallel dual-branch feature extraction module, and extract global and local features through the Swing-Transformer branch and the shuffled sensing feature branch of the parallel dual-branch feature extraction module, respectively.

[0074] The fault identification module is used to adaptively weight and fuse the extracted global and local features, classify the fused features, and output the diagnostic results of the fault category.

[0075] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0076] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A fault diagnosis method for flexible DC distribution networks, characterized in that, The method includes: Step S1: Collect fault current timing signals on the DC side of the flexible DC distribution network; Step S2: Input the fault current time series signal into the pre-trained fault diagnosis model. The fault diagnosis model uses an interpretable complex time-frequency convolutional layer to perform adaptive time-frequency feature extraction on the fault current time series signal to obtain a time-frequency-amplitude tensor. In step S2, the adaptive time-frequency feature extraction of the fault current time-series signal includes: constructing a complex convolution kernel structure consisting of F parallel channels, each channel corresponding to a set of physically meaningful frequency bands with trainable center frequencies; for each channel, performing a convolution operation on the fault current time-series signal using a complex convolution kernel function, the complex convolution kernel function consisting of a real part and an imaginary part, defined as: ,in, For complex convolution kernel function, and These are the real and imaginary parts of the corresponding kernel function, respectively. The imaginary unit, To enable learnable parameter generation and constraint by physical boundaries, adaptive learning updates are performed during training, including the center frequency and window width scaling factor. For each channel, a modulo operation is performed on the real and imaginary convolution results to obtain the channel's frequency band amplitude response time series. ,in, Let the amplitude response time series be the frequency band k. For the first Each frequency band at time The corresponding combined real and imaginary part of the frequency band amplitude, ;in, , The first The result of convolving each frequency band with a real part convolution kernel and an imaginary part convolution kernel at time n, where... Where N is the kernel length, Indicates the input signal The result of shifting the value in the time domain, where m is the index of the shift. For the k-th frequency band, at discrete time m, given the parameters... The real part of the kernel function is determined. The learnable parameters for the k-th frequency band are constrained by physical boundaries; ,in, For the k-th frequency band, at discrete time m, given the parameters... The imaginary part of the kernel function is determined; Stacking the frequency band amplitude response time series of F channels yields the discrete time-frequency-amplitude tensor H. ; Step S3: Construct time-amplitude trajectory, time-frequency trajectory and frequency-amplitude trajectory based on the time-frequency-amplitude tensor, and perform Gram differential angle field encoding on each trajectory to generate a three-channel time-frequency feature matrix; In step S3, constructing the time-amplitude trajectory includes: averaging the time-frequency-amplitude tensor along the frequency band dimension to obtain the average amplitude of the signal at the nth discrete time point across all frequency components. The calculation formula is: Where n is the discrete time series index, For frequency band indexing, Number of frequency bands For the discrete-time index n, the nth The amplitude corresponding to each frequency band is combined with the average amplitude at multiple time points according to the time series to generate a time-amplitude trajectory, which is represented as: T represents the total length of the time series, and the time-amplitude trajectory reflects the overall energy trend of the signal over time under the combined energy of all frequency bands; In step S3, constructing the time-frequency trajectory includes: calculating the Softmax frequency weight along the frequency band for each time n, and obtaining the weight coefficients of the time-frequency feature corresponding to the k-th frequency band at the discrete time index n. The calculation formula is: ,in, This is the frequency centroid smoothing factor, used to control the smoothness of the normalized weight distribution, and is set to a fixed value of 0.

3. For the discrete-time index n, the nth The amplitude corresponding to each frequency band; the time-frequency characteristics at F frequency bands are calculated based on the Softmax frequency weighting. We perform a weighted summation to obtain the time-frequency characteristic value of the signal at time n. The calculation method is as follows: The time-frequency feature values ​​are sorted according to the time series to generate a time-frequency trajectory. , The time-frequency trajectory reflects the main frequency trajectory that changes over time; Step S4: Input the three-channel time-frequency feature matrix into the parallel dual-branch feature extraction module, and extract global and local features through the Swin-Transformer branch and the shuffled sensing feature branch of the parallel dual-branch feature extraction module, respectively. Step S4 includes: the Swin-Transformer branch adopts a hierarchical structure, including a sliding window multi-head self-attention mechanism, inputting the three-channel time-frequency feature matrix into the Swin-Transformer branch for feature modeling, and achieving resolution halving and channel doubling through image merging operations between stages, and obtaining global semantic information across regions to obtain global features; the shuffling-aware feature branch includes a lightweight convolutional neural network and a shuffling-attention aware module, and by integrating spatial attention and channel attention mechanisms, combined with channel shuffling operations, selectively responding to local texture, edge and micro-perturbation features of the image and interacting with cross-channel information to obtain local features; Step S5: Perform adaptive weighted fusion on the extracted global features and local features, classify the fused features, and output the diagnostic results of the fault category.

2. The method according to claim 1, characterized in that, The complex convolution kernel function is a continuous-time complex kernel function based on the short-time Fourier transform, defined as: ,in, It is a continuous-time complex function. For the current time, This is the window width ratio factor. The center frequency is j, where j is the imaginary unit. Discretize the continuous-time kernel function to obtain the discrete-time kernel function: , For discrete-time series indexing, odd numbers are chosen to ensure centered symmetry. The real and imaginary parts of the discrete-time kernel function constitute the real and imaginary parts of the complex convolution kernel function, respectively.

3. The method according to claim 1, characterized in that, Step S3, which involves constructing the frequency-amplitude trajectory, includes: For each frequency band, calculate the mean amplitude along the time dimension to obtain the mean amplitude corresponding to the k-th frequency band. The calculation formula is as follows: The frequency-amplitude trajectory is generated by combining the average amplitude values ​​corresponding to the F frequency bands. , The frequency-amplitude trajectory reflects the average energy distribution of each frequency band across the entire time domain.

4. The method according to claim 1, characterized in that, The step S3 of generating the three-channel time-frequency feature matrix includes: The time-amplitude trajectory, the time-frequency trajectory, and the frequency-amplitude trajectory are resampled to length L using linear interpolation, respectively, as follows: , and L is the length of each trajectory; Resampled time-amplitude trajectory Normalization is performed to obtain the time-amplitude trajectory at time x. The processing formula is: , and They are respectively The minimum and maximum values ​​in the time-amplitude trajectory will be normalized. The angle is obtained by performing an inverse cosine transform, and the transform formula is: ,in, for The angle value at time x. To normalize Perform an inverse cosine transform. Let x be the polar radius of the polar coordinate axis at time x, and construct the trajectory relationship matrix based on the angle. : ,in, Let be the angle value at time i, where ; right and After normalization and inverse cosine transform, the corresponding trajectory relationship matrix is ​​obtained. , , combined Each trajectory relationship matrix is ​​standardized, and the standardized trajectory relationship matrices are then fused to generate a three-channel time-frequency feature matrix.

5. The method according to claim 1, characterized in that, Step S5 includes: The global features are processed using an adaptive average pooling operator. Align to local features Within the spatial scale, the aligned global features are obtained. ; Set up a lightweight weighting mechanism to align global features and local features Perform global average pooling to obtain the corresponding channel feature factors. The channel feature factor Perform concatenation along the channel dimension to obtain the concatenated result. ; and the splicing result Perform a linear mapping to obtain the logarithmic weights a and b of the global and local branches. , in, , These are the weighting coefficients. , These are bias parameters; A temperature-normalized softmax operation is performed on the logarithmic weights of the global and local branches to normalize the logarithmic weights into branch weights. , The normalization formula is: ,in, Temperature coefficient; Use the calculated branch weights to align the global features. and local features Weighted fusion is performed to obtain fusion features. The fused features are input into a global average pooling layer and a linear classifier to obtain the diagnostic results of the fault category.

6. A flexible DC distribution network fault diagnosis system, used to implement the flexible DC distribution network fault diagnosis method as described in any one of claims 1-5, characterized in that, The system includes: The data acquisition module is used to acquire fault current timing signals on the DC side of the flexible DC distribution network; The feature extraction module is used to input the fault current time series signal into a pre-trained fault diagnosis model. The fault diagnosis model uses an interpretable complex time-frequency convolutional layer to perform adaptive time-frequency feature extraction on the fault current time series signal to obtain a time-frequency-amplitude tensor. The feature transformation module is used to construct time-amplitude trajectory, time-frequency trajectory and frequency-amplitude trajectory based on the time-frequency-amplitude tensor, and to perform Gram differential angle field encoding on each trajectory to generate a three-channel time-frequency feature matrix; The feature fusion module is used to input the three-channel time-frequency feature matrix into the parallel dual-branch feature extraction module, and extract global features and local features through the Swin-Transformer branch and the shuffled sensing feature branch of the parallel dual-branch feature extraction module, respectively. The fault identification module is used to adaptively weight and fuse the extracted global features and local features, classify the fused features, and output the diagnostic results of the fault category.

Images