Distribution network line fault detection data processing method, system, equipment and medium

Abstract

The invention relates to the technical field of power system fault diagnosis, and discloses a distribution network line fault detection data processing method, system and device and a medium, and the method comprises the steps: obtaining load data of a target distribution line detection point, and carrying out the first preprocessing of the load data, and obtaining a fusion feature vector; performing second optimization operation on the fusion feature vector to obtain an optimized fusion feature vector; calculating a potential feature matrix according to the optimized fusion feature vector, and mapping the potential feature matrix to a low-dimensional space to obtain a low-dimensional sample set; presetting an adaptive label propagation algorithm, and performing fault category judgment on the low-dimensional sample set based on the adaptive label propagation algorithm; and storing a judgment result in a relational database. The problems that feature extraction is not accurate in a high-noise environment, and a classification model is insufficient in new fault expansion capacity are effectively solved, and fault signal processing robustness is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system fault diagnosis, and in particular to a distribution network line fault detection data processing method, system, device and medium. BACKGROUND

[0002] With the increasing demand for power reliability and safety in modern society, the operation and maintenance of distribution networks have become an important part of power system management. However, due to the wide coverage and complex structure of distribution network lines, they are often affected by external environment, equipment aging and load fluctuation, leading to frequent line faults. In order to timely detect and locate the fault point and reduce the impact of power interruption on users, fault detection technology has become a key research direction in the power industry. At present, traditional fault detection techniques rely on time-domain analysis or frequency-domain feature extraction methods of traveling wave signals, such as wavelet transform or Fourier transform-based fault detection techniques, which identify fault types by analyzing the time-frequency characteristics of line fault signals. Although these methods can achieve good results in specific scenarios, their robustness and accuracy are still insufficient in the face of complex and variable noise environments, especially when fault signals are mixed in a strong noise background, the detection accuracy decreases significantly. In addition, with the continuous expansion of the scale of distribution networks and the rapid growth of fault data, the efficiency and accuracy of traditional methods in processing high-dimensional data are difficult to meet the actual demand.

[0003] Another important bottleneck of the prior art is the lack of adaptability to new fault types. Most traditional methods are based on pre-defined fault types for training and identification, and lack dynamic expansion capability, making it difficult to cope with unknown fault categories that may occur in actual operation. Although deep learning-based fault detection methods such as convolutional neural networks (CNN) and long short-term memory networks (LSTM) have been introduced in recent years, these methods have advantages in processing high-dimensional feature data and complex pattern recognition, but they usually require large-scale labeled samples for training and have weak recognition ability for unlabeled samples and new fault categories. In addition, these methods are easily disturbed in high-noise environments and are difficult to achieve efficient processing and classification of fault signals. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a distribution network line fault detection data processing method, system, device and medium, which can solve the problems of existing distribution network line fault detection techniques in noise suppression, feature extraction and processing of high-dimensional data relying on static models.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a distribution network line fault detection data processing method, comprising:

[0008] Obtaining load data of a target distribution line detection point, and performing first preprocessing on the load data to obtain a fusion feature vector;

[0009] Performing a second optimization operation on the fusion feature vector to obtain an optimized fusion feature vector;

[0010] Calculating a latent feature matrix according to the optimized fusion feature vector, and mapping the latent feature matrix to a low-dimensional space to obtain a low-dimensional sample set;

[0011] The calculation of the latent feature matrix is obtained by designing an encoder;

[0012] A preset adaptive label propagation algorithm is used to judge the fault category of the low-dimensional sample set based on the adaptive label propagation algorithm;

[0013] The judgment result is stored in a relational database.

[0014] As a preferred scheme of the distribution network line fault detection data processing method of the present application, the adaptive label propagation algorithm comprises:

[0015] According to the obtained low-dimensional sample set, the number of nearest neighbors of each sample is set, and the Euclidean distance between each pair of samples is calculated to generate a Euclidean distance matrix;

[0016] An adjacency matrix is constructed according to the calculated Euclidean distance matrix;

[0017] The adjacency matrix represents the preliminary connection relationship between samples, and the value between non-neighbors is set to infinity;

[0018] Based on the adjacency matrix, a shortest path algorithm is used to calculate a geodesic distance matrix between all samples;

[0019] The adjacency matrix is optimized, and the optimized adjacency matrix is normalized to generate a propagation matrix.

[0020] This preferred scheme can more accurately reflect the actual relationship between samples by accurately calculating the geodesic distance between samples, and improve the accuracy of fault category judgment. At the same time, the propagation matrix generated by the normalized processing of the optimized adjacency matrix can more stably propagate labels, reduce the possibility of misjudgment, and thus improve the efficiency and reliability of fault detection.

[0021] As a preferred scheme of the distribution network line fault detection data processing method of the present application, the optimization of the adjacency matrix comprises:

[0022] By calculating the manifold curvature of each sample, the nonlinear complexity of the local area around the sample is evaluated;

[0023] According to the manifold curvature of each sample, a curvature threshold is set, and the number of neighbors thereof is dynamically adjusted:

[0024] According to the adjusted number of neighbors, the adjacency matrix is updated, so that the number of neighbors of each sample matches its manifold complexity, and the local connectivity of the neighborhood is optimized.

[0025] As a preferred scheme of the distribution network line fault detection data processing method, the dynamic adjustment of the number of neighbors includes:

[0026] If the manifold curvature of the sample is greater than or equal to the curvature threshold, it indicates that the structure of the area is complex, and the neighbors need to be allocated;

[0027] If the manifold curvature of the sample is less than the curvature threshold, it indicates that the structure of the area is simple, and the neighbors need to be reduced.

[0028] As a preferred scheme of the distribution network line fault detection data processing method, the mapping of the potential feature matrix to the low-dimensional space to obtain the low-dimensional sample set includes:

[0029] For each sample, the Euclidean distance between the sample and all other samples is calculated, and the Euclidean distance between the sample and all other samples is converted to a similarity value through a Gaussian distribution;

[0030] The width of the Gaussian distribution is dynamically adjusted for each sample, the similarity values of all samples are normalized, and the similarity values of each two samples are averaged to obtain a joint probability distribution in the high-dimensional space;

[0031] Initialize the sample position in the low-dimensional space, calculate the distance between each two samples in the low-dimensional space, and map the distance between each two samples to a similarity value in the low-dimensional space;

[0032] Compare the similarities in the high-dimensional space and the low-dimensional space, calculate the difference between the two distributions to measure the preservation degree of the low-dimensional space to the high-dimensional structure, and use gradient descent method to continuously adjust the position of the sample in the low-dimensional space.

[0033] As a preferred scheme of the distribution network line fault detection data processing method, the second optimization operation includes:

[0034] The number of neurons of the input layer of the dynamic sparse model is determined according to the number of features of the fused feature vector, and the number of neurons of the output layer is configured based on the requirements of the next layer of the dynamic sparse model and the target dimension;

[0035] According to the number of neurons of the input layer and the output layer, the total number of connections between the input layer and the output layer is calculated, and an activation function is selected for the neurons of the input layer and the output layer;

[0036] The sparsity of the dynamic sparse model is set by the total number of connections, and the activated connections are randomly selected and the initial weights are allocated according to the sparsity;

[0037] The fusion feature vector is input to the input layer, and after each topology adjustment, the optimized fusion feature vector is output.

[0038] As a preferred scheme of the power distribution line fault detection data processing method, the load data of the target power distribution line detection point is obtained, including:

[0039] The collection device is installed at the target power distribution line detection point, and a time synchronization device is configured, and the interval between adjacent detection points is set according to actual needs;

[0040] The collection device includes a traveling wave detection device, a current sensor and a voltage sensor;

[0041] The current mutation threshold is preset in each traveling wave detection device;

[0042] If the current value detected by the traveling wave detection device exceeds the current mutation threshold, the collection trigger state is entered, otherwise, no operation is performed;

[0043] The collection trigger state refers to collecting the load data of the current time point by the collection device;

[0044] The load data includes traveling wave signals, current signals and voltage signals.

[0045] In a second aspect, the present application provides a power distribution line fault detection data processing system, comprising:

[0046] A data acquisition and processing module is configured to acquire load data of a target power distribution line detection point, and perform first preprocessing on the load data to obtain a fusion feature vector;

[0047] An optimization module is configured to perform second optimization operation on the fusion feature vector to obtain an optimized fusion feature vector;

[0048] A low-dimensional sample set acquisition module is configured to calculate a latent feature matrix according to the optimized fusion feature vector, and map the latent feature matrix to a low-dimensional space to obtain a low-dimensional sample set;

[0049] The latent feature matrix is obtained by designing an encoder;

[0050] The fault judgment module is configured to preset an adaptive label propagation algorithm, and perform fault category judgment on the low-dimensional sample set based on the adaptive label propagation algorithm.

[0051] The storage module is configured to store the judgment result into a relational database.

[0052] In a third aspect, the present application provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described above when executing the computer program.

[0053] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the method described above when executed by a processor.

[0054] Compared with the prior art, the present application has the following beneficial effects: the present application provides a distribution network line fault detection data processing method, acquires load data of a target distribution line detection point, performs first preprocessing on the load data to obtain a fusion feature vector, performs second optimization operation on the fusion feature vector to obtain an optimized fusion feature vector, calculates a latent feature matrix according to the optimized fusion feature vector, maps the latent feature matrix to a low-dimensional space to obtain a low-dimensional sample set, presets an adaptive label propagation algorithm, performs fault category judgment on the low-dimensional sample set based on the adaptive label propagation algorithm, and stores the judgment result into a relational database. The present application effectively solves the problems of inaccurate feature extraction in a high-noise environment, insufficient expansion capability of a classification model for new faults, and the like, and improves the robustness of fault signal processing.

[0055] Specifically, first, by acquiring load data of a target distribution line detection point and performing preprocessing, key features can be extracted, providing a basis for subsequent analysis. Second, optimization operation is performed on the fusion feature vector, which can further purify feature information and improve the accuracy and reliability of data. Third, the latent feature matrix is mapped to a low-dimensional space to obtain a low-dimensional sample set, which helps to reduce computational complexity and improve processing efficiency. Finally, the adaptive label propagation algorithm is used to perform fault category judgment on the low-dimensional sample set, and the judgment result is stored into a relational database, realizing rapid identification and recording of faults and providing a strong guarantee for stable operation of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0057] Figure 1 A method flow chart of a distribution network line fault detection data processing method is provided for an embodiment of the present application.

[0058] Figure 2 A flow chart of the fault category of the detection sample of the distribution network line fault detection data processing method is provided for an embodiment of the present application.

[0059] Figure 3 An internal structure diagram of an electronic device of the distribution network line fault detection data processing method is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0060] To make the above objectives, characteristics and advantages of the present application more apparent, obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0061] Embodiment 1, refer to Figures 1-3 For the first embodiment of the present application, the embodiment provides a distribution network line fault detection data processing method, comprising:

[0062] In the prior art, there are some problems, such as difficulty in accurately detecting fault signals in a complex noise environment, low efficiency of traditional methods in processing high-dimensional data, and insufficient adaptability to new fault types.

[0063] The present application provides a method that can effectively solve the above-mentioned problems. Next, how to implement the distribution network line fault detection data processing method will be described in detail in combination with multiple embodiments.

[0064] Figure 1 A method flow chart of a distribution network line fault detection data processing method is provided, comprising:

[0065] S101, obtaining load data of a target distribution line detection point, and performing first preprocessing on the load data to obtain a fusion feature vector;

[0066] It should be noted that, in order to successfully implement the distribution network line fault detection data processing, the relevant basic data of the target distribution network line needs to be regularly obtained. Different basic data can be obtained according to different detection points. For example, load data is the basis for distribution network line fault detection, which contains current, voltage, traveling wave signal and other information. These information is crucial for subsequent feature extraction and fault judgment.

[0067] In an optional embodiment, the specific manner of acquiring the load data of the target power distribution line detection point can be to install high-precision sensors and data acquisition devices at key nodes of the target power distribution line to ensure that the changes in the load data can be captured in real time and accurately. These sensors and data acquisition devices should have high stability and anti-interference capability, and can continuously work in harsh environments to ensure the integrity and reliability of the data.

[0068] In an optional embodiment, in order to ensure the timeliness and accuracy of the data, a time synchronization device can also be configured to ensure that the data of all collection points are synchronized in time, thereby facilitating subsequent data analysis and processing. During the collection process, the interval between adjacent detection points can be set according to actual needs to balance the data density and processing efficiency.

[0069] In the embodiments of the present application, acquiring the load data of the target power distribution line detection point includes:

[0070] The collection device and the time synchronization device are installed at the target power distribution line detection point, and the interval between adjacent detection points is set according to actual needs;

[0071] The collection device includes a traveling wave detection device, a current sensor, and a voltage sensor;

[0072] A current mutation threshold is preset in each traveling wave detection device;

[0073] If the traveling wave detection device detects that the current value exceeds the current mutation threshold, the collection trigger state is entered, otherwise, no operation is performed;

[0074] The collection trigger state refers to collecting the load data of the current time point by using the collection device;

[0075] The load data includes a traveling wave signal, a current signal, and a voltage signal.

[0076] Specifically, collecting the load data of the detection point includes,

[0077] The collection device and the time synchronization device are installed at the detection point of the distribution network line, and the interval between adjacent detection points is set according to actual needs;

[0078] The detection point includes a substation outlet point, a line turning point, a branch node, and a device access point in the distribution network line, etc.

[0079] The collection device refers to a traveling wave detection device, a current sensor, and a voltage sensor;

[0080] According to the normal working current of the distribution network line, a current mutation threshold is preset in each traveling wave detection device, and if the traveling wave detection device detects that the current value exceeds the current mutation threshold, the collection trigger state is entered, otherwise, no operation is performed;

[0081] The acquisition trigger state refers to acquiring load data at a current time point by using an acquisition device;

[0082] The load data includes a traveling wave signal, a current signal, and a voltage signal.

[0083] It should be noted that by deploying a traveling wave detection device, a current sensor, and a voltage sensor at a key node of a distribution network line and configuring a time synchronization device, the spatiotemporal consistency of the acquired data is ensured, the acquisition trigger is realized by predefining a current mutation threshold, and the load data is acquired only when an anomaly is detected, thereby effectively avoiding the redundant storage and processing of invalid data, improving the acquisition efficiency and the timeliness of fault response, and providing accurate and reliable raw data basis for subsequent fault detection.

[0084] In an optional embodiment, the first preprocessing is a frame integration operation, including,

[0085] The load data is subjected to frame processing, a closing operation is performed on each frame of load data to fill in small gaps in the load data, an opening operation is performed on the load data after the closing operation to eliminate local peak noise, a result difference value of the closing operation and the opening operation is calculated, a feature vector is formed by combining the difference value of each frame, local peak values in the difference value are found and extracted, the local peak values are taken as key time domain feature values of each frame of load data, and the feature vector of each frame of load data and the extracted key time domain feature values are combined and subjected to normalization processing to form a complete time domain feature vector;

[0086] The closing operation refers to using a structure element S to perform dilation and then erosion to smooth the load data and fill in gaps in the signal;

[0087] The opening operation refers to using a structure element S to perform erosion and then dilation to eliminate local peak noise in the load data;

[0088] In an optional embodiment, the load data is subjected to 6-level decomposition by using a stationary wavelet transform, and Daubechies 10 is selected as a mother wavelet;

[0089] The 6-level decomposition includes applying a low-pass filter and a high-pass filter to each frame of load data to obtain first-level detailed coefficients and approximation coefficients, and recursively decomposing the approximation coefficients of the first level until the detailed coefficients and the approximation coefficients of the 6th level are obtained;

[0090] The detailed coefficients of each level and the approximation coefficients of the 6th level are extracted and subjected to normalization processing, and are sequentially arranged and combined into a frequency domain feature vector. The detailed coefficients of each level reflect high-frequency components of different frequency bands, and the approximation coefficients of the 6th level serve as overall low-frequency components of the load data and provide a stable reference for the overall frequency spectrum;

[0091] In an optional embodiment, the time domain feature vector and the frequency domain feature vector are spliced in sequence to form a fusion feature vector.

[0092] It should be noted that by integrating the load data by framing, combining the closing operation, the opening operation and the stationary wavelet transform, the gaps in the signal can be effectively filled, the local peak noise can be eliminated, and the key time domain feature values and frequency domain features of the load data can be extracted. The closing and opening operations improve the smoothness and noise robustness of the time domain signal, and the stationary wavelet decomposition accurately captures the feature changes of different frequency bands, providing multi-scale information for signal analysis. Finally, the time domain and frequency domain features are spliced in sequence into a fusion feature vector, realizing high-quality and high-resolution feature expression, and providing a stable and accurate data basis for subsequent fault detection.

[0093] S102, performing a second optimization operation on the fusion feature vector to obtain an optimized fusion feature vector;

[0094] In an optional embodiment, the second optimization operation is to further purify and optimize the fusion feature vector to improve the accuracy and efficiency of fault detection. This operation can include but is not limited to feature selection, feature dimension reduction, feature weighting and other steps, aiming to select the most representative features from the fusion feature vector, reduce redundant information, and enhance the ability of the features to distinguish fault types. Through the second optimization operation, the quality of the fusion feature vector can be further improved to provide more reliable data support for subsequent low-dimensional mapping and fault judgment.

[0095] In the embodiments of the present application, the second optimization operation includes:

[0096] The number of neurons of the input layer of the dynamic sparse model is determined according to the number of features of the fusion feature vector, and the number of neurons of the output layer is configured based on the next layer requirement and the target dimension of the dynamic sparse model;

[0097] According to the number of neurons of the input layer and the output layer, the total number of connections between the input layer and the output layer is calculated, and an activation function is selected for the neurons of the input layer and the output layer;

[0098] The sparsity of the dynamic sparse model is set by the total number of connections, and the activated connections are randomly selected and assigned with initial weights according to the sparsity;

[0099] The fusion feature vector is input to the input layer, and after each topology adjustment, the optimized fusion feature vector is output.

[0100] Specifically, using the dynamic sparse model to perform noise classification filtering on the fusion feature vector to obtain the optimized fusion feature vector includes,

[0101] The number of neurons of the input layer of the dynamic sparse model is determined according to the number of features of the fusion feature vector, the number of neurons of the output layer is configured based on the requirement of the next layer of the dynamic sparse model and the target dimension, the total number of connections between the input layer and the output layer is calculated according to the number of neurons of the input layer and the output layer, and ReLU is selected as the activation function of the neurons of the input layer and the output layer;

[0102] The sparsity of the dynamic sparse model is set by the total number of connections, the activated connections are randomly selected according to the sparsity, and initial weights are assigned to the activated connections, and the weights of the unactivated connections are set to 0;

[0103] In an optional embodiment, topology adjustment is performed every n steps, n is determined according to actual requirements, the proportion of the connections to be removed and the connections to be grown is determined according to the total number of connections after each topology adjustment, the low-weight connections are removed by sorting the weights, and the connections to be grown are randomly selected from the unactivated connections, the initial weights of the connections to be grown are set to 0, the weight gradient of each activated connection is calculated by using a loss function, the loss function is designed to minimize the prediction error, and the weight is updated by the gradient descent method;

[0104] The fusion feature vector is input to the input layer, and the optimized fusion feature vector is output after each topology adjustment;

[0105] In an optional embodiment, the SAC algorithm and the TD3 algorithm are introduced for low-noise training and high-noise training, and the weight distribution of different algorithms is set, including setting the time-domain feature weight of the SAC algorithm to 1.0 and the frequency-domain feature weight to 0.5, and setting the frequency-domain feature weight of the TD3 algorithm to 1.0 and the time-domain feature weight to 0.5, so as to prioritize the optimization of related features under different noise levels;

[0106] The root mean square value of the optimized fusion feature vector is calculated, and a noise level threshold is set to determine the noise level;

[0107] In an optional embodiment, if the root mean square value of the optimized fusion feature vector is less than or equal to the noise level threshold, it is determined that the current noise level is low, and the SAC algorithm is activated, and if the root mean square value of the optimized fusion feature vector is greater than the noise level threshold, it is determined that the current noise level is high, and the TD3 algorithm is activated;

[0108] Activating the SAC algorithm means that the SAC algorithm prioritizes the processing of time-domain features, the current time-domain feature weight is increased, and the frequency-domain feature weight remains unchanged, thereby ensuring the priority of time-domain features in a low-noise environment, a loss function of the SAC training process is set to minimize the prediction error, and after every 10 steps of training, the compliance of the loss value is determined, if the loss value is compliant, the next frame is processed, otherwise, iterative training is continued;

[0109] In an optional embodiment, time domain features are prioritized in a low noise environment, which helps to enhance the detection of signal mutations, for example, in fault detection applications, time domain features usually reflect the mutation of current or voltage, which can quickly characterize the occurrence of faults, by setting the weight of time domain features to 1.0 in the SAC algorithm, the sensitivity of the model to signal changes in a low noise environment can be enhanced, and the model can capture fault signals faster and more accurately;

[0110] Activating the TD3 algorithm means that TD3 preferentially enhances the weight of frequency domain features, increases the current weight of frequency domain features, and keeps the weight of time domain features unchanged, to ensure focusing on frequency domain features in a high noise environment. TD3 algorithm uses a delayed update strategy, and the weight is updated every two steps. After 100 steps of training, TD3 locks the weight of frequency domain features;

[0111] In an optional embodiment, frequency domain features are prioritized in a high noise environment, because frequency domain features are usually more stable in a high noise environment. In a high noise environment, time domain mutations are easily masked by noise, while frequency domain features can isolate noise through decomposition, thereby better extracting the core frequency characteristics of the signal. TD3 preferentially allocates the weight of frequency domain features to enable the model to focus on a specific frequency range in a strong noise environment and suppress noise components;

[0112] In an optional embodiment, the maximum number of topology structure adjustments is set, and the final optimized fusion feature vector is output;

[0113] It should be noted that dynamic priority helps to balance the sensitivity and stability of the signal in practical applications. For example, in distribution network fault detection, mutation signals are a direct indicator of fault occurrence, and frequency domain features in a high noise environment can filter out noise interference. Priority allocation enables the model to quickly respond to fault mutations and extract effective signals in a noisy environment.

[0114] It should be noted that by using a dynamic sparse model to filter noise from the fusion feature vector, dynamic adjustment of feature importance and efficient filtering of noise signals are achieved. By introducing SAC and TD3 algorithms, the weight distribution of time and frequency domain features is optimized for different noise levels, enabling the model to quickly capture fault mutation signals in a low noise environment and accurately extract frequency domain features in a high noise environment, improving the adaptability to complex environments. This method balances signal sensitivity and stability through dynamic topology structure adjustment, providing a high-robustness, high-precision optimization solution for distribution network fault detection, effectively addressing noise interference and signal fluctuations in practical applications.

[0115] S103, calculate the latent feature matrix according to the optimized fusion feature vector, and map the latent feature matrix to a low-dimensional space to obtain a low-dimensional sample set;

[0116] In an optional embodiment, the potential feature matrix is a matrix obtained by a specific transformation of the optimized fusion feature vector, which contains rich fault feature information but has a high dimension. Direct processing may face problems such as high computational complexity and large storage requirements. Therefore, it is necessary to map the potential feature matrix to a low-dimensional space to simplify the subsequent processing process while retaining key fault features as much as possible.

[0117] In an optional embodiment, the process of mapping to a low-dimensional space can be realized by a dimension reduction algorithm such as principal component analysis (PCA), linear discriminant analysis (LDA), or t-SNE. In the embodiments of the present application, PCA can be selected as the dimension reduction method because it can effectively extract the main components of the data while reducing the dimension of the data. PCA converts the original data into a new coordinate system through an orthogonal transformation, and the data variance on each coordinate axis of the new coordinate system is arranged from large to small. The first few coordinate axes with the largest variance are selected, that is, the dimension reduction processing of the data is realized.

[0118] In an optional embodiment, the potential feature matrix is taken as the input of the PCA algorithm, which calculates the principal component direction of the data and selects the first few principal component directions to form a low-dimensional space according to the set dimension reduction target (such as the number of dimensions after dimension reduction). Then, the potential feature matrix is projected into this low-dimensional space to obtain a low-dimensional sample set. The low-dimensional sample set retains the main fault feature information in the original data while reducing the dimension of the data, providing a more concise and efficient data basis for subsequent fault detection.

[0119] In the embodiments of the present application, the potential feature matrix is calculated by designing an encoder.

[0120] Specifically, the encoder is used to output the potential feature matrix according to the fusion feature vector, which includes,

[0121] The encoder part is constructed, and the input data is the fusion feature vector;

[0122] In an optional embodiment, the first convolutional layer uses 16 convolutional kernels, each with a size of 4x1 and a step size of 2x1, to extract local features through convolution operation. The size of the convolution output is calculated, and a ReLU activation function is used to introduce nonlinear features for stable training.

[0123] In an optional embodiment, the second convolutional layer has 16 input channels and 32 output channels, and the convolution kernel size, step size, and activation function are the same as those of the first layer. After convolution, a max-pooling operation is performed with a pooling window size of 2x1 and a step size of 2x1, further reducing the feature dimension and enhancing the representation ability of local features.

[0124] In an optional embodiment, the encoder extracts features through a convolutional layer, and the output is flattened into a one-dimensional vector through a Flatten operation. The flattened features are input into a fully connected layer to output a mean vector and a log standard deviation vector, and the latent variable is generated through a reparameterization trick.

[0125] The decoder part is constructed to take the latent variable as input and map it back to a feature vector of the same dimension as the flattened output of the encoder through a fully connected layer.

[0126] In an optional embodiment, the first deconvolutional layer has 32 input channels and 16 output channels. The deconvolutional operation is used to restore the data dimension, and the parameter settings are symmetrical with the convolutional layer in the encoder. The activation function uses ReLU.

[0127] In an optional embodiment, the second deconvolutional layer has 16 input channels and 1 output channel. The parameter settings are consistent with the first deconvolutional layer, and the activation function is Sigmoid, which limits the output value to the interval [0, 1].

[0128] Through the output layer, the decoder outputs a reconstructed feature vector with the same dimension as the fused feature vector.

[0129] In an optional embodiment, the mean squared error is used to measure the difference between the fused feature vector and the reconstructed feature vector L c The KL divergence loss method is used to constrain the difference between the latent variable and the standard normal distribution L k By minimizing L c , it indicates that the model can constrain the distribution of the latent variable to be close to the standard normal distribution, thereby ensuring the controllability and continuity of the latent space. The KL divergence loss method is a measure tool for measuring the difference between two probability distributions. The goal here is to adjust the latent variable generated by the variational autoencoder to the standard normal distribution.

[0130] According to the fault detection requirements, the total number of fault categories is set, and each fused feature vector is taken as each sample. Each sample is manually labeled with its fault category.

[0131] In an optional embodiment, the same class samples are guaranteed to be clustered in the feature space, reducing the distribution range of the intra-class samples. The local separation loss is used to construct a loss term that measures the tightness of the feature distribution of the intra-class samples. The distance between the same class samples and their feature mean vector F o is calculated, and the formula is:

[0132]

[0133] where C represents the total number of fault categories, n i represents the number of samples in the i th category, and fj i u represents the feature vector of the jth sample of the ith category, u i u represents the feature mean vector in the ith category;

[0134] In the embodiments of the present application, the intra-class sample refers to all samples belonging to the same category, for example, in a classification task, assuming that there are two fault types of "fault A" and "fault B", all samples in the "fault A" category are referred to as the intra-class samples of "fault A";

[0135] The feature mean vector refers to averaging each dimension of the feature vector to obtain a new mean vector;

[0136] The main role of the local separation loss is to enhance the aggregation of samples of the same class in the feature space, by minimizing the distance between the intra-class sample and its feature mean, the loss reduces the dispersion of intra-class features, so that the samples of the same category are more closely;

[0137] In an optional embodiment, a non-local separation loss is used to measure the distance F a between the feature distribution centers of different categories, and the formula is:

[0138]

[0139] Wherein, u represents the feature mean vector in all categories;

[0140] The core purpose of the non-local separation loss is to expand the distance between the feature means of different categories, so as to improve the separation degree of inter-class features, by maximizing the discriminability of the category means, the model can more clearly divide the boundaries of different categories, reduce the possibility of classification confusion, this inter-class separation characteristic not only improves the classification accuracy, but also provides support for detecting new fault categories, so that the position of the unlabeled sample in the feature space is more prominent, which is helpful to find the fault features of unknown categories;

[0141] In an optional embodiment, a total loss function J is constructed, and the formula is: J = L c + L k + F o + F a The Adam optimizer is selected, the initial learning rate and batch size are set, the total loss of the current batch is calculated every round of training, the parameters are updated through back propagation, and the training is performed for 100 rounds, and the learning rate is dynamically adjusted every 10 rounds;

[0142] All fusion feature vectors are input into the trained encoder part to generate a latent feature matrix, each row of the latent feature matrix represents a sample, and each table represents the feature value of the sample in a certain dimension.

[0143] It should be noted that by extracting the latent feature matrix of the fusion feature vector through the encoder, and combining the reparameterization technology of the variational autoencoder, the latent variable is generated to retain the key feature information, while the controllability and continuity of the latent space are realized by using the mean square error and KL divergence loss, the local separation loss is introduced to enhance the aggregation of the same type of samples in the feature space, and the non-local separation loss is introduced to expand the separation degree of the feature centers of different categories, so as to improve the clarity of the classification boundary, significantly reduce the confusion rate, and optimize the model by constructing the total loss function, and finally the generated latent feature matrix effectively supports the fault category classification and new fault detection, and enhances the accuracy and expansion ability of fault recognition.

[0144] In the embodiment of the application, mapping the latent feature matrix to a low-dimensional space to obtain a low-dimensional sample set comprises:

[0145] For each sample, the Euclidean distance between the sample and all other samples is calculated, and the Euclidean distance between the sample and all other samples is converted to a similarity value by a Gaussian distribution;

[0146] The width of the Gaussian distribution is dynamically adjusted for each sample, the similarity values of all samples are normalized, and the similarity values of each two samples are averaged to obtain a joint probability distribution in the high-dimensional space;

[0147] The positions of the samples in the low-dimensional space are initialized, the distance between each two samples in the low-dimensional space is calculated, and the distance between each two samples is mapped to a similarity value in the low-dimensional space;

[0148] The similarities in the high-dimensional space and the low-dimensional space are compared, and the difference between the two distributions is calculated to measure the degree of preservation of the high-dimensional structure in the low-dimensional space, and the positions of the samples in the low-dimensional space are continuously adjusted using the gradient descent method.

[0149] Specifically, mapping the latent feature matrix to a low-dimensional space to obtain a low-dimensional sample set using a t-SNE algorithm comprises,

[0150] The t-SNE algorithm is used for dimensionality reduction and visualization of high-dimensional data, and is used in the scene of mapping high-dimensional features to two-dimensional or three-dimensional for visualization analysis;

[0151] In an optional embodiment, for each sample, the Euclidean distance between the sample and all other samples is calculated to measure the difference between the samples, and the distances are converted to similarity values by a Gaussian distribution, the width of the Gaussian distribution is dynamically adjusted for each sample, so that the neighborhood distribution of each sample is uniform, the similarity values of all samples are normalized, and the similarity values of each two samples are averaged to obtain a joint probability distribution in the high-dimensional space, the joint probability represents the relationship between two samples in the local field in the high-dimensional space;

[0152] Initialize the sample positions in the low-dimensional space, usually starting with a random way, in the low-dimensional space, calculate the distance between each two samples, map these distances to similarity values in the low-dimensional space using t-distribution, the similarity values in the low-dimensional space represent the local relationship of the samples in the low-dimensional space after dimensionality reduction;

[0153] In t-distribution, t-distribution can better solve the "crowding problem" in the mapping from high-dimensional to low-dimensional, compared with Gaussian distribution, t-distribution is more suitable to describe the relationship of points in low-dimensional space, especially for those far sample points, its long tail characteristics can provide more reasonable low-dimensional mapping results;

[0154] Compare the similarities in high-dimensional and low-dimensional spaces, measure the preservation of high-dimensional structure in low-dimensional space by calculating the difference between two distributions using KL divergence, continuously adjust the positions of samples in low-dimensional space using gradient descent method to reduce the difference between high-dimensional and low-dimensional spaces;

[0155] In an optional embodiment, during the optimization process, the sample positions in the low-dimensional space will gradually adjust to a balanced state, when the change amplitude of the optimization target is less than the set threshold, stop optimization, output the low-dimensional sample set X, the expression is:

[0156] X={x1,x2,...,x g}

[0157] Where g represents the total number of samples, and each element represents the position in the sample space.

[0158] It should be noted that by mapping the latent feature matrix to the low-dimensional space through the t-SNE algorithm, the high-dimensional similarity is dynamically adjusted using Gaussian distribution and the low-dimensional mapping relationship is optimized using t-distribution, effectively solving the "crowding problem" in the mapping from high-dimensional to low-dimensional, t-SNE algorithm measures the distribution difference between high-dimensional and low-dimensional spaces through KL divergence, and dynamically optimizes the sample positions by combining gradient descent, finally realizes the local relationship preservation of samples in low-dimensional space and the rationality of global distribution, this method enhances the separation and aggregation of features in low-dimensional space, provides clear feature distribution and analysis basis for subsequent classification and new fault discovery, significantly improves the accuracy and visualization effect of fault detection.

[0159] S104, preset adaptive label propagation algorithm, judge the fault categories of the low-dimensional sample set based on the adaptive label propagation algorithm;

[0160] In an optional embodiment, the adaptive label propagation algorithm can propagate label information by utilizing similarity measures between nodes to achieve automatic classification of samples. First, a weighted graph is constructed, where nodes represent samples in the low-dimensional sample set, and edge weights between nodes represent the similarity between samples. Similarity can be determined by calculating the distance between samples, the closer the distance, the higher the similarity, and the larger the edge weight.

[0161] In an optional embodiment, after the weighted graph is constructed, the label of each node is initialized as its own class (if any) or randomly assigned a label. Then, the algorithm starts iteration, and in each iteration, each node updates its label by selecting the most frequent label among its neighbor nodes as its new label. If a node has multiple labels with the same frequency, one of them can be selected as the new label, or other strategies (such as random selection, weight-based selection, etc.) can be used to determine the new label.

[0162] In an optional embodiment, some adaptive mechanisms are introduced in the adaptive label propagation algorithm to optimize the label propagation process. For example, edge weights can be adjusted according to the local density and label distribution of nodes, so that the algorithm can better handle unbalanced data sets and noisy data. In addition, some constraints can be introduced to limit the propagation range of labels to avoid label confusion caused by excessive propagation.

[0163] It should be noted that after multiple iterations, the algorithm will gradually converge, and each node will obtain a relatively stable label. At this time, the fault category to which the node belongs can be determined according to the label of the node. For unknown samples, they can be mapped to a low-dimensional space and then classified using the trained adaptive label propagation algorithm to achieve automatic detection and identification of distribution network line faults.

[0164] In the embodiments of the present application, the adaptive label propagation algorithm includes:

[0165] According to the obtained low-dimensional sample set, the number of nearest neighbors of each sample is set, and the Euclidean distance between each pair of samples is calculated to generate a Euclidean distance matrix;

[0166] According to the calculated Euclidean distance matrix, an adjacency matrix is constructed;

[0167] The adjacency matrix represents the preliminary connection relationship between samples, and the value between non-neighbors is set to infinity.

[0168] Based on the adjacency matrix, a geodesic distance matrix between all samples is calculated using the shortest path algorithm;

[0169] The adjacency matrix is optimized, and the optimized adjacency matrix is normalized to generate a propagation matrix.

[0170] In the embodiment of the present application, the optimization of the adjacency matrix comprises:

[0171] By calculating the manifold curvature of each sample, the non-linear complexity of the local area around the sample is evaluated;

[0172] According to the manifold curvature of each sample, a curvature threshold is set to dynamically adjust the number of neighbors thereof:

[0173] According to the adjusted number of neighbors, the adjacency matrix is updated so that the number of neighbors of each sample matches its manifold complexity, and the local connectivity of the neighborhood is optimized.

[0174] In the embodiment of the present application, the dynamic adjustment of the number of neighbors thereof comprises:

[0175] If the manifold curvature of the sample is greater than or equal to the curvature threshold, it indicates that the structure of the region is complex, and neighbors need to be allocated;

[0176] If the manifold curvature of the sample is less than the curvature threshold, it indicates that the structure of the region is simple, and the number of neighbors needs to be reduced.

[0177] Specifically, the detection of the fault class of the sample by the adaptive label propagation algorithm comprises,

[0178] According to the obtained low-dimensional sample set, the number of nearest neighbors of each sample is set, and the Euclidean distance between each pair of samples is calculated to generate a Euclidean distance matrix, which is used to measure the direct similarity between samples;

[0179] According to the calculated Euclidean distance matrix, an adjacency matrix is constructed to represent the preliminary connection relationship between samples, and the value between non-neighbors is set to infinity. Based on the adjacency matrix, a geodesic distance matrix between all samples is calculated using a shortest path algorithm, which is used to reflect the manifold structure of the data;

[0180] In an optional embodiment, by calculating the manifold curvature (ratio of Euclidean distance to geodesic distance) of each sample, the non-linear complexity of the local area around the sample is evaluated to provide a basis for subsequent dynamic adjustment of the neighbor data;

[0181] According to the manifold curvature of each sample, a curvature threshold is set to dynamically adjust the number of neighbors thereof:

[0182] If the manifold curvature of the sample is greater than or equal to the curvature threshold, it indicates that the structure of the region is complex, and neighbors need to be allocated;

[0183] If the manifold curvature of the sample is less than the curvature threshold, it indicates that the structure of the region is simple, and the number of neighbors needs to be reduced.

[0184] According to the adjusted number of neighbors, the adjacency matrix is updated so that the number of neighbors of each sample matches its manifold complexity, and the local connectivity of the neighborhood is optimized.

[0185] In an optional embodiment, the updated adjacency matrix is normalized to generate a propagation matrix for controlling the intensity of label propagation between samples and ensuring the stability of the propagation process.

[0186] In an optional embodiment, the label matrix is initialized by writing the labels of the labeled samples into the matrix and initializing the label positions of the unlabeled samples as 0. In the label propagation process, the label distribution probability of the unlabeled samples is iteratively updated, each iteration simultaneously referring to the propagation matrix between samples and the label information of the labeled samples, and calculating the change of the label matrix. A convergence threshold is set, and if the change of the label matrix is less than the convergence threshold, the propagation is stopped. After the label matrix converges, each row represents the probability distribution of each sample in all categories. For each sample, the category corresponding to the highest probability value is taken as the current fault category.

[0187] After the propagation is completed, the label distribution of the unlabeled samples is checked, and if the probability of the label distribution of the sample to all known categories is low, it is determined to be a new fault category.

[0188] A new category label is created for the unlabeled sample, which is labeled as a new category, and the sample of the new category is included in the training set, and the model is retrained to recognize the new category.

[0189] It should be noted that through the adaptive label propagation algorithm, the number of neighbors is dynamically adjusted according to the sample manifold structure, the local connectivity of the adjacency matrix is optimized, and the propagation matrix is generated to ensure the stability of the label propagation process, realizing efficient fault classification and identification of new fault categories. Through the calculation of manifold curvature, adaptive neighbor adjustment and iterative update of the label matrix, the local and global characteristics of the data are accurately captured, not only improving the classification accuracy of known categories, but also effectively detecting unknown fault categories, and including new categories in model training, realizing dynamic expansion and adaptive optimization of the model, providing a solution with robustness and scalability for fault detection in complex environments.

[0190] S105, store the judgment result to a relational database.

[0191] It should be noted that the detailed operation can be as shown in the accompanying drawings Figure 2 First, the system will collect load data from each detection point and integrate these data frame by frame to form a fusion feature vector. This is the starting point of the entire process, ensuring a reliable data basis for subsequent analysis.

[0192] Next, the dynamic sparse model is used to filter the noise of the fusion feature vector. This step aims to remove or reduce noise in the data to improve the accuracy of subsequent analysis.

[0193] The system introduces SAC (Soft Actor-Critic) algorithm and TD3 (Twin Delayed Deep Deterministic Policy Gradient) algorithm for training in low-noise and high-noise environments, and sets the weight distribution of different algorithms. This step selects the most suitable algorithm for training according to different noise environments to adapt to different working conditions.

[0194] Set the noise level threshold, and determine whether the root mean square value of the fusion feature vector is less than or equal to the noise level threshold. If the result is "no", i.e. the current noise level is high, activate the TD3 algorithm; if the result is "yes", i.e. the current noise level is low, activate the SAC algorithm. This decision step ensures that the most suitable algorithm is selected for processing in different noise environments.

[0195] According to the algorithm activated in the previous step, set the maximum number of topology structure adjustments, and output the final optimized fusion feature vector. This step further optimizes the feature vector by adjusting the network structure to improve its representation ability.

[0196] Use the encoder to output the reconstructed feature vector according to the fusion feature vector, then construct the total loss function, input all the fusion feature vectors to the trained encoder part, and generate the latent feature matrix. This process further refines and optimizes the feature information through reconstruction and loss function construction.

[0197] Use t-SNE (t-Distributed Stochastic Neighbor Embedding) algorithm to map the latent feature matrix to low-dimensional space to get low-dimensional sample set. t-SNE is a commonly used dimensionality reduction algorithm that can effectively visualize high-dimensional data for subsequent analysis.

[0198] Finally, the adaptive label propagation algorithm is used to detect the fault class of the sample. This step uses the optimized feature vector and low-dimensional sample set obtained in the previous steps to identify and classify the fault class, and completes the functional goal of the entire system.

[0199] In summary, the application provides a distribution network line fault detection data processing method, obtains load data of a target distribution line detection point, and performs first preprocessing on the load data to obtain a fusion feature vector; performs second optimization on the fusion feature vector to obtain an optimized fusion feature vector; calculates a latent feature matrix according to the optimized fusion feature vector, and maps the latent feature matrix to a low-dimensional space to obtain a low-dimensional sample set; a preset adaptive label propagation algorithm is used to judge the fault category of the low-dimensional sample set based on the adaptive label propagation algorithm; and the judgment result is stored in a relational database. The problems of inaccurate feature extraction in a high-noise environment and insufficient expansion capability of a classification model for new faults are effectively solved, and the robustness of fault signal processing is improved.

[0200] In a preferred embodiment of embodiment 2, after storing the judgment result in the relational database, a management operation is required, and all data is stored in the relational database and managed.

[0201] Specifically, the relational database is selected to store and manage data and analysis results, a database table structure is designed to store different types of data, a regular backup task is set, all data in the database is backed up, database users are managed in terms of authority, and data is stored in an encrypted manner.

[0202] By selecting the relational database to store and manage data and analysis results, designing a reasonable table structure to store different types of data, setting a regular backup task and a user authority management mechanism, and storing data in an encrypted manner, the present scheme realizes the efficiency of data storage, the safety of management, and the reliability of backup. The method effectively guarantees the integrity and safety of data, avoids data loss due to unexpected circumstances, ensures access control of sensitive information, and improves the availability and credibility of the system in data management and analysis.

[0203] In embodiment 3, a distribution network line fault detection data processing system is also provided, comprising:

[0204] A data acquisition and processing module is configured to acquire load data of a target distribution line detection point, and perform first preprocessing on the load data to obtain a fusion feature vector.

[0205] An optimization module is configured to perform second optimization on the fusion feature vector to obtain an optimized fusion feature vector.

[0206] A low-dimensional sample set acquisition module is configured to calculate a latent feature matrix according to the optimized fusion feature vector, and map the latent feature matrix to a low-dimensional space to obtain a low-dimensional sample set.

[0207] The latent feature matrix is obtained by designing an encoder;

[0208] The fault judgment module is configured to preset an adaptive label propagation algorithm, and perform fault category judgment on the low-dimensional sample set based on the adaptive label propagation algorithm.

[0209] The storage module is configured to store the judgment result in a relational database.

[0210] The above-mentioned unit modules can be embedded in or independent of the processor in the electronic device in hardware form, or can be stored in the memory in the electronic device in software form, so as to be called and executed by the processor to perform the operations corresponding to the above-mentioned modules.

[0211] The embodiment also provides an electronic device, which can be a terminal, and an internal structure diagram of the electronic device can be as shown in the figure. Figure 3 The electronic device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the electronic device is configured to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the electronic device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The computer program is executed by the processor to implement a power distribution line fault detection data processing method. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen. The input device of the electronic device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the electronic device, or an external keyboard, touchpad or mouse, etc.

[0212] The embodiment also provides a computer readable storage medium having a computer program stored thereon, and the computer program is executed by the processor to implement the following steps:

[0213] Obtain load data of a target power distribution line detection point, and perform first preprocessing on the load data to obtain a fusion feature vector;

[0214] Perform second optimization operation on the fusion feature vector to obtain an optimized fusion feature vector;

[0215] Calculate a latent feature matrix according to the optimized fusion feature vector, and map the latent feature matrix to a low-dimensional space to obtain a low-dimensional sample set;

[0216] The latent feature matrix is obtained by designing an encoder;

[0217] Preset an adaptive label propagation algorithm, and perform fault category judgment on the low-dimensional sample set based on the adaptive label propagation algorithm;

[0218] The result of the judgment is stored in a relational database.

[0219] It should be noted that the above-mentioned embodiments are only used to illustrate but not to limit the technical solutions of the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the field should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.

[0220] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can adopt a computer program product in the form of being implemented on one or more computer usable storage media containing computer usable program codes (including but not limited to disk memory, CD-ROM, optical memory, etc.). The solutions in the embodiments of the present application can be implemented in various computer languages.

[0221] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more blocks.

[0222] These computer program instructions can also be stored in a computer readable memory that can guide the computer or other programmable data processing devices to work in a specific way, so that the instructions stored in the computer readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more blocks.

[0223] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to produce a computer implemented process, so that the instructions executed on the computer or other programmable device provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1steps of the functions specified in the block or blocks.

[0224] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the preferred embodiments by those of skill in the art once they have the benefit of the present disclosure. Therefore, the appended claims are intended to encompass within their scope all possible variations and modifications of the preferred embodiments.

[0225] It is apparent that those skilled in the art can make various changes and modifications to the application without departing from the spirit and scope of the application. It is therefore intended that the application cover all such changes and modifications that are within its scope.

Claims

1. A distribution network line fault detection data processing method, characterized in that: include: Obtaining load data of a target distribution line detection point, and performing a first preprocessing on the load data to obtain a fused feature vector; performing a second optimization operation on the fused feature vector to obtain an optimized fused feature vector; Calculating a latent feature matrix based on the optimized fusion feature vector, and mapping the latent feature matrix to a low-dimensional space to obtain a low-dimensional sample set; The calculated potential feature matrix is ​​obtained by designing an encoder; Presetting an adaptive label propagation algorithm, and performing fault category judgment on the low-dimensional sample set based on the adaptive label propagation algorithm; The judgment results are stored in a relational database.

2. A distribution network line fault detection data processing method according to claim 1, characterized in that: The adaptive label propagation algorithm includes: According to the obtained low-dimensional sample set, the number of nearest neighbors of each sample is set, and the Euclidean distance between each pair of samples is calculated to generate a Euclidean distance matrix; Construct an adjacency matrix based on the calculated Euclidean distance matrix; The adjacency matrix represents the preliminary connection relationship between samples, and the values ​​between non-neighbors are set to infinity; Based on the adjacency matrix, the shortest path algorithm is used to calculate the geodesic distance matrix between all samples; The adjacency matrix is ​​optimized, and the optimized adjacency matrix is ​​normalized to generate a propagation matrix.

3. A distribution network line fault detection data processing method according to claim 2, characterized in that: The optimized adjacency matrix includes: By calculating the manifold curvature of each sample, the nonlinear complexity of the local area around the sample is evaluated; According to the manifold curvature of each sample, the curvature threshold is set and the number of its neighbors is dynamically adjusted: According to the adjusted number of neighbors, the adjacency matrix is ​​updated so that the number of neighbors of each sample matches its manifold complexity and the local connectivity of the neighborhood is optimized.

4. A distribution network line fault detection data processing method according to claim 3, characterized in that: The dynamic adjustment of the number of neighbors includes: If the manifold curvature of the sample is greater than or equal to the curvature threshold, it means that the structure of the area is complex and neighbors need to be assigned; If the manifold curvature of the sample is less than the curvature threshold, it means that the structure of the area is simple and the neighbors need to be reduced.

5. A distribution network line fault detection data processing method according to claim 4, characterized in that: Mapping the potential feature matrix to a low-dimensional space to obtain a low-dimensional sample set includes: For each sample, calculate the Euclidean distance between the sample and all other samples, and convert the Euclidean distance between the sample and all other samples into a similarity value through Gaussian distribution; The width of the Gaussian distribution is dynamically adjusted by each sample, the similarity values ​​of all samples are normalized, and the similarity values ​​of every two samples are averaged to obtain a joint probability distribution in a high-dimensional space; Initialize the sample positions in the low-dimensional space, calculate the distance between each two samples in the low-dimensional space, and map the distance between each two samples to a similarity value in the low-dimensional space; Compare the similarities in high-dimensional space and low-dimensional space, calculate the difference between the two distributions to measure the degree to which the low-dimensional space retains the high-dimensional structure, and use the gradient descent method to continuously adjust the position of samples in the low-dimensional space.

6. A distribution network line fault detection data processing method according to claim 5, characterized in that: The second optimization operation includes: The number of neurons in the input layer of the dynamic sparse model is determined based on the number of features in the fused feature vector, and the number of neurons in the output layer is configured based on the requirements of the next layer of the dynamic sparse model and the target dimension; According to the number of neurons in the input layer and the output layer, the total number of connections between the input layer and the output layer is calculated, and the activation function is selected for the neurons in the input layer and the output layer; The sparse rate of the dynamic sparse model is set by the total number of connections, and the activated connections are randomly selected and assigned initial weights according to the sparse rate; The fused feature vector is input into the input layer, and after each topology adjustment, the optimized fused feature vector is output.

7. A distribution network line fault detection data processing method according to claim 6, characterized in that: The step of obtaining load data of a target distribution line detection point includes: Install data collection devices and configure time synchronization equipment at the target distribution line inspection points, and set the intervals between adjacent inspection points based on actual needs; The acquisition device includes a traveling wave detection device, a current sensor and a voltage sensor; A current mutation threshold is pre-set in each traveling wave detection device; If the traveling wave detection device detects that the current value exceeds the current mutation threshold, it enters the acquisition trigger state, otherwise, no operation is performed; The acquisition trigger state refers to using the acquisition device to collect load data at the current time point; The load data includes a traveling wave signal, a current signal, and a voltage signal.

8. A distribution network line fault detection data processing system, applying the method according to any one of claims 1 to 7, characterized in that: include: A data acquisition and processing module, configured to acquire load data of a target distribution line detection point and perform a first preprocessing on the load data to obtain a fused feature vector; an optimization module, configured to perform a second optimization operation on the fused feature vector to obtain an optimized fused feature vector; A low-dimensional sample set acquisition module is used to calculate a potential feature matrix based on the optimized fusion feature vector, and map the potential feature matrix to a low-dimensional space to obtain a low-dimensional sample set; The calculated potential feature matrix is ​​obtained by designing an encoder; A fault judgment module is configured to preset an adaptive label propagation algorithm and perform fault classification judgment on the low-dimensional sample set based on the adaptive label propagation algorithm; The storage module is used to store the judgment results in a relational database.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the distribution network line fault detection data processing method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a distribution network line fault detection data processing method according to any one of claims 1 to 7 are implemented.

Images