Fault type identification method, system and equipment based on fuzzy waveform

Abstract

This invention discloses a method, system, and device for fault type identification based on fuzzy waveforms, belonging to the field of fault type identification technology. The method includes: collecting multi-source fault data from a distribution network to construct a fault dataset; performing mode decomposition on the fault dataset to extract a set of fault features; determining cluster centers using a clustering algorithm based on the fault feature set to construct fuzzy waveform features; training multiple structurally independent feature forest base classifiers using an ensemble learning algorithm based on the fault dataset and cluster centers; inputting the fuzzy waveform features into the feature forest base classifiers to obtain a preliminary identification result set; and performing fusion reasoning using a probabilistic graphical model algorithm based on the preliminary identification result set and the power grid topology to obtain the fault type identification result. This application aims to address the problem of insufficient identification of complex waveform features in traditional fault identification methods.

Description

Technical Field

[0001] This invention relates to the field of fault type identification technology, and more specifically, to a fault type identification method, system, and device based on fuzzy waveforms. Background Technology

[0002] As a crucial link connecting power sources and users in the power system, the safety and reliability of the distribution network directly affect the power supply quality and the stability of the power system. During the operation of the distribution network, various types of faults frequently occur, such as short circuits, grounding faults, and open circuits. These faults typically manifest as abnormal changes in voltage and current waveforms. Failure to identify the fault type promptly and accurately will lead to delayed fault location and may even trigger large-scale power outages.

[0003] For example, the invention patent application with publication number CN119494023B discloses a method and system for fault location in a smart distribution network, relating to the field of distribution network technology. The method includes: real-time acquisition of voltage amplitude and frequency signals by deploying distributed acquisition devices at each node, calculating spatial distribution patterns to generate a feature template library; generating node state labels matching fault modes by combining the distribution network topology and amplitude-frequency distribution characteristics; designing a fault location model based on a graph convolutional network, introducing spatiotemporal convolution and distribution pattern comparison modules to extract fault modes and predict node fault states; collecting real-time data after a fault occurs, calculating the matching similarity between distribution features and the template library, and comprehensively judging the fault type and specific location; dynamically updating the distributed template library and optimizing the model algorithm by testing the model's performance in different scenarios. This invention is applicable to complex distribution network scenarios, especially under distributed power source access, dynamic topology, and noise interference conditions, enabling efficient fault diagnosis.

[0004] For example, the invention patent announcement with publication number CN118395230B relates to a method for identifying distribution network fault types based on big data, including: acquiring a preprocessed fault training sample set of the distribution network; extracting features from the fault waveform data of the fault training samples to obtain the time-domain, frequency-domain, and entropy features of the fault training samples; performing whitening processing on the time-domain, frequency-domain, and entropy features of the fault training samples to obtain the whitened time-domain features, whitened frequency-domain features, and whitened entropy features of the fault training samples; filtering the whitened time-domain features, whitened frequency-domain features, and whitened entropy features of the fault training samples according to the feature variance and fault category label of all fault training samples to obtain the time-domain attribute features, frequency-domain attribute features, and entropy attribute features of the fault training samples; and using a fault identification model to predict the fault type of the fault training samples based on the time-domain attribute features, frequency-domain attribute features, and entropy attribute features of the fault training samples. This invention effectively improves the accuracy of distribution network fault prediction.

[0005] Existing methods have limitations in adaptability to complex scenarios: On the one hand, when faced with nonlinear distortions and multi-scale oscillating waveforms accompanying faults such as high-resistance grounding and arc discharge, traditional methods rely heavily on static features such as amplitude and frequency, without extracting intrinsic mode function components through empirical mode decomposition or capturing instantaneous dynamic features using Hilbert transform, resulting in the loss of key distortion information and making it difficult to accurately identify complex faults; on the other hand, the boundary ambiguity of fault features due to sensor noise and electromagnetic interference is not effectively characterized, and the deterministic feature processing mode or fixed template matching method used in existing technologies are prone to misjudgment due to feature matching deviations in interference environments, resulting in insufficient anti-interference capability.

[0006] To address the above problems, this invention proposes a solution. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a fault type identification method, system, and device based on fuzzy waveforms. By constructing waveform features through fuzzy C-means clustering and employing a decision mechanism that combines random forest ensemble learning with Bayesian inference, the problem of insufficient identification of complex waveform features by traditional fault identification methods is solved.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a fault type identification method based on fuzzy waveforms, comprising the following steps: collecting multi-source fault data from a distribution network to construct a fault dataset; performing mode decomposition processing on the fault dataset to extract a fault feature set; determining cluster centers using a clustering algorithm based on the fault feature set to construct fuzzy waveform features; training multiple structurally independent feature forest base classifiers using an ensemble learning algorithm based on the fault dataset and cluster centers; inputting the fuzzy waveform features into the feature forest base classifiers to obtain a preliminary identification result set; and performing fusion reasoning using a probabilistic graphical model algorithm based on the preliminary identification result set and the power grid topology to obtain the fault type identification result.

[0009] In a preferred embodiment, the step of collecting multi-source fault data from the distribution network and constructing a fault dataset specifically involves: processing the multi-source fault data using a time synchronization algorithm to obtain time-aligned fault data; and obtaining a fault dataset based on the time-aligned fault data using a data normalization algorithm.

[0010] In a preferred embodiment, the step of performing mode decomposition on the fault dataset to extract the fault feature set specifically involves: performing mode decomposition on the fault dataset to obtain intrinsic mode function components; calculating the information entropy features and instantaneous features of the components; and fusing the information entropy features and instantaneous features to obtain the fault feature set.

[0011] In a preferred embodiment, the step of determining cluster centers and constructing fuzzy waveform features based on the fault feature set using a clustering algorithm specifically involves: determining the optimal number of clusters based on the fault feature set through clustering effectiveness analysis; generating a fuzzy membership matrix by combining the fault feature set and the optimal number of clusters using a fuzzy clustering algorithm; performing feature weighting processing on the fuzzy membership matrix to obtain a weighted feature vector; and converting the weighted feature vector into fuzzy waveform features using a feature reconstruction algorithm.

[0012] In a preferred embodiment, the step of training multiple structurally independent feature forest base classifiers based on the fault dataset and cluster centers using an ensemble learning algorithm specifically involves: dividing the fault dataset into multiple fault data subsets based on the cluster centers; training an ensemble classifier for each fault data subset; and performing weighted optimization on each ensemble classifier using the cluster centers to form the final classifier set.

[0013] In a preferred embodiment, the weighted optimization of each ensemble classifier by cluster centers includes: adjusting the weights of features in the fault data subset based on the membership degree of the cluster centers.

[0014] In a preferred embodiment, the step of determining cluster centers and constructing fuzzy waveform features based on the fault feature set using a clustering algorithm specifically involves: inputting the fuzzy waveform features in parallel into multiple structurally independent feature forest base classifiers; obtaining the corresponding classification result in each feature forest base classifier based on a probability output mechanism; and summarizing the output results of all base classifiers to form a preliminary identification result set.

[0015] In a preferred embodiment, the step of training multiple structurally independent feature forest base classifiers based on the fault dataset and cluster centers using an ensemble learning algorithm specifically involves: constructing a probabilistic graphical model based on the distribution network topology, where each node in the probabilistic graphical model corresponds to a candidate fault type; inputting the preliminary identification result set as evidence into the probabilistic graphical model; obtaining the posterior probability of each candidate fault type through probabilistic inference; and determining the final fault type identification result based on the posterior probability.

[0016] A fault type identification system based on fuzzy waveforms is characterized by comprising: a data acquisition and construction module for acquiring multi-source fault data from a distribution network and constructing a fault dataset; a feature extraction module for performing mode decomposition processing on the fault dataset to extract a set of fault features; a fuzzy waveform construction module for determining cluster centers based on the fault feature set using a clustering algorithm to construct fuzzy waveform features; a classifier training module for training multiple structurally independent feature forest base classifiers using an ensemble learning algorithm based on the fault dataset and cluster centers; a preliminary identification module for inputting the fuzzy waveform features into the feature forest base classifiers to obtain a preliminary identification result set; and a topology association and fusion module for performing fusion reasoning based on the preliminary identification result set and the power grid topology relationship using a probabilistic graphical model algorithm to obtain the fault type identification result.

[0017] A fault type identification device based on fuzzy waveforms, characterized in that it includes at least one processor; and an input / output interface communicatively connected to the at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the fault type identification method based on fuzzy waveforms.

[0018] The technical effects and advantages of the fault type identification method, system, and equipment based on fuzzy waveforms in this invention are as follows:

[0019] This invention proposes a fault type identification method, system, and equipment based on fuzzy waveforms. By extracting intrinsic mode function components through empirical mode decomposition and combining energy entropy and instantaneous features, it achieves comprehensive extraction of multi-scale and multi-dimensional features of fault waveforms, overcoming the limitations of traditional methods such as single feature and missing information. Fuzzy waveform features are constructed through fuzzy C-means clustering, effectively improving the characterization accuracy of complex faults such as high-resistance grounding and arc discharge. Furthermore, multiple base classifiers are trained using the random forest algorithm, and Bagging and random feature subsampling are used to enhance the model's noise resistance, obtaining stable preliminary identification results. Finally, the distribution network topology is introduced, and Bayesian networks are used to fuse and infer the preliminary results. Topological constraints are used to correct misjudgments, improving the rationality and reliability of the identification results. The parallel preliminary identification using random forests and the topological association inference using Bayesian networks leverage both the fitting ability of machine learning to massive amounts of data and the prior knowledge of the power grid topology, achieving complementary advantages of multiple models. This effectively solves the problem of insufficient identification of complex waveform features in traditional fault identification methods. Attached Figure Description

[0020] Figure 1 A schematic diagram of the fault type identification method based on fuzzy waveforms provided in an embodiment of the present invention;

[0021] Figure 2 A schematic diagram of the fault type identification system based on fuzzy waveforms provided in an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the structure of a fault type identification device based on fuzzy waveforms provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0025] Example 1, Figure 1 This invention presents a fault type identification method based on fuzzy waveforms, comprising the following steps:

[0026] S1, collect multi-source fault data of the distribution network and construct a fault dataset;

[0027] S2, perform mode decomposition on the fault dataset to extract the fault feature set;

[0028] S3. Based on the fault feature set, cluster centers are determined through a clustering algorithm to construct fuzzy waveform features;

[0029] S4. Based on the fault dataset and cluster centers, multiple structurally independent feature forest base classifiers are trained using an ensemble learning algorithm.

[0030] S5. Input the fuzzy waveform features into the feature forest base classifier to obtain a preliminary recognition result set;

[0031] S6. Based on the preliminary identification result set and the power grid topology, a probabilistic graphical model algorithm is used for fusion reasoning to obtain the fault type identification result. The probabilistic graphical model algorithm in this embodiment adopts the Bayesian network algorithm.

[0032] This invention proposes a fault type identification method, system, and equipment based on fuzzy waveforms. By extracting intrinsic mode function components through empirical mode decomposition and combining energy entropy and instantaneous features, it achieves comprehensive extraction of multi-scale and multi-dimensional features of fault waveforms, overcoming the limitations of traditional methods such as single features and missing information. Fuzzy waveform features are constructed through fuzzy C-means clustering, effectively improving the characterization accuracy of complex faults such as high-resistance grounding and arc discharge. Furthermore, multiple base classifiers are trained using the random forest algorithm, and Bagging and random feature subsampling are used to enhance the model's noise resistance, obtaining stable preliminary identification results. Finally, the distribution network topology is introduced, and Bayesian networks are used to fuse and infer the preliminary results. Topological constraints are used to correct misjudgments, improving the rationality and reliability of the identification results. The parallel preliminary identification using random forests and the topological association inference using Bayesian networks leverage machine learning's ability to fit massive amounts of data and utilizes prior knowledge of the power grid topology, achieving complementary advantages among multiple models. This effectively solves the problems of insufficient identification of complex waveform features and lack of multi-model collaborative mechanisms in traditional fault identification methods.

[0033] S1 collects multi-source fault data from the distribution network and constructs a fault dataset.

[0034] In this embodiment, the process of collecting multi-source fault data from the distribution network and constructing a fault dataset specifically involves:

[0035] Multi-source fault data are obtained through power distribution network monitoring points;

[0036] A time synchronization algorithm is used to process multi-source fault data to obtain spatiotemporally aligned fault data;

[0037] Based on spatiotemporally aligned fault data, a fault dataset is obtained through a data normalization algorithm.

[0038] Data acquisition devices are deployed at multiple monitoring points in the distribution network to collect multi-source fault data in real time. The multi-source fault data includes three-phase voltage waveform data, three-phase current waveform data, fault time information, and monitoring point location data. Among them, the three-phase voltage and current waveforms can reflect the dynamic changes of electrical quantities when the fault occurs, the fault time information is used to mark the time period when the fault is triggered, and the monitoring point location data provides support for subsequent spatial positioning based on the distribution network topology.

[0039] Because the clocks of sampling devices at different monitoring points differ, directly acquired multi-source signals often exhibit time inconsistencies. Therefore, this embodiment introduces a time synchronization algorithm to process the multi-source fault data, aligning the waveforms of each monitoring point to a unified time reference. Through synchronization processing, the consistency of multi-source signals in the time dimension can be guaranteed, and further combined with the location information of the monitoring points, spatiotemporally aligned fault data can be achieved.

[0040] The time synchronization algorithm is implemented based on the IEEE 1588 precision clock synchronization protocol. Its core steps include: designating a high-precision clock source as the master clock in the network; the master clock periodically sending synchronization messages (Sync) and follow-up messages (Follow_Up) to all slave monitoring devices, where the Follow_Up message contains the precise timestamp of the Sync message transmission; slave devices recording the arrival time of the Sync message and using the timestamp in the Follow_Up message to calculate the network transmission delay and clock offset; and slave devices dynamically adjusting their local clocks based on the calculated offset to achieve microsecond-level or even nanosecond-level time synchronization, thereby ensuring that all monitoring point data has a unified high-precision time reference.

[0041] After obtaining the spatiotemporally aligned data, it is also necessary to address the issues of inconsistent dimensions and amplitude differences among different monitoring points. To this end, this embodiment introduces a normalization algorithm to process the aligned data. The Min-Max normalization method can be used to map waveform signals such as voltage and current to a uniform scale range, eliminating amplitude deviations between different monitoring points, thereby ensuring the effectiveness of subsequent feature extraction and modeling training.

[0042] The Min-Max normalization method maps electrical quantity waveform data of different magnitudes to a unified scale range of [0,1], eliminating dimensional differences while preserving the original distribution trend of the data. The specific implementation steps are as follows: First, for different types of electrical quantities in the spatiotemporally aligned data, their extreme value ranges are determined respectively. For three-phase voltage waveforms, based on the range of 10kV distribution network monitoring equipment (0~12kV) and historical fault data statistics, the minimum and maximum values ​​of all voltage data are taken. For three-phase current waveforms, combined with the range of feeder terminal unit (FTU) (0~600A) and the actual fault current fluctuation range, the corresponding minimum and maximum values ​​are taken. If there is no historical data for a newly commissioned distribution network, the upper and lower limits of the rated range of the monitoring equipment are directly used as the extreme values. If there are abnormal extreme values ​​of over-range caused by instantaneous sensor faults in a certain type of data, the abnormal values ​​are removed by the 3σ principle before the extreme values ​​are determined. Secondly, normalization calculation is performed on each original electrical quantity data point: the original data is subtracted from the minimum value of its corresponding electrical quantity type, and the difference is then divided by the difference between the maximum and minimum values ​​of that type of data, finally yielding the normalized data.

[0043] The data, after time synchronization and normalization, is organized and encapsulated to construct sample instances based on fault occurrence events. Each sample instance includes the three-phase voltage waveform, three-phase current waveform, fault time, and monitoring point location at the time of the fault. By aggregating multiple sample instances, a complete fault dataset is finally formed, providing reliable data support for subsequent feature extraction and fault type identification.

[0044] S2 performs mode decomposition on the fault dataset to extract the fault feature set.

[0045] In this embodiment, the modal decomposition processing of the fault dataset to extract the fault feature set specifically involves:

[0046] Modal decomposition is performed on the fault dataset to obtain the intrinsic mode function components;

[0047] Calculate the information entropy feature and instantaneous feature of the component;

[0048] By fusing information entropy features with instantaneous features, a fault feature set is obtained.

[0049] The empirical mode decomposition algorithm, through an iterative selection process, decomposes the original fault data signal into several intrinsic mode function (IMF) components and a residual term, the mathematical expression of which is as follows:

[0050]

[0051] in, It is the original fault data signal. Indicates the first The intrinsic mode function components of each fault data point. It is a residual signal, usually a single trend term. It is the number of IMFs in the decomposed fault data. This represents the number of iterations for filtering.

[0052] Next, for each fault data IMF component, the energy entropy algorithm is used to calculate the complexity of its energy distribution. Energy entropy characterizes the degree of distribution of signal energy, and its calculation formula is as follows:

[0053]

[0054] in, Indicates the first Energy entropy of IMF component of fault data It is the first Signal energy at a given moment For the first The intrinsic mode function components of each fault data point are calculated as the squared amplitude at that moment. It is the total time length of the component. This represents the number of iterations for filtering.

[0055] The Hilbert transform formula is:

[0056]

[0057] in, It is a transient feature. Hilbert transform of the IMF components of the fault data. Indicates the first The intrinsic mode function components of each fault data point. The imaginary unit, This represents the number of iterations for filtering.

[0058] In this embodiment, the feature fusion algorithm is specifically as follows:

[0059] The component energy entropy features and instantaneous features are concatenated to form an initial fused feature vector containing both component energy entropy features and standardized instantaneous features;

[0060] Then, weights are assigned to each initial feature based on its variance contribution.

[0061] The initial fusion feature vector is optimized by weighted summation to form a fault feature set.

[0062] In this embodiment, assigning weights to each initial feature based on variance contribution is specifically as follows:

[0063] Calculate the variance of each feature in the historical fault samples;

[0064] The variance is normalized and then used as the weight of the corresponding feature.

[0065] It should be noted that the larger the variance, the stronger the ability of the feature to distinguish fault types;

[0066] Furthermore, by optimizing the initial fusion feature vector through weighted summation, the role of key features in fault identification is highlighted, ultimately forming a fault feature set that combines the static energy distribution characteristics and dynamic temporal change characteristics of the fault.

[0067] S3. Based on the fault feature set, cluster centers are determined through a clustering algorithm to construct fuzzy waveform features.

[0068] In this embodiment, the step of determining cluster centers and constructing fuzzy waveform features based on the fault feature set using a clustering algorithm specifically involves:

[0069] Based on the fault feature set, the optimal number of clusters is determined through clustering effectiveness analysis;

[0070] By combining the fault feature set and the optimal number of clusters, a fuzzy membership matrix is ​​generated using a fuzzy clustering algorithm;

[0071] The fuzzy membership matrix is ​​subjected to feature weighting to obtain a weighted eigenvector;

[0072] The weighted feature vectors are converted into fuzzy waveform features using a feature reconstruction algorithm.

[0073] For the feature set extracted from fault data, it is necessary to determine an appropriate number of clusters. To avoid the subjectivity of traditional clustering methods that rely on manually setting the number of clusters, this embodiment introduces the DB index method for adaptive determination of the number of clusters. The specific steps are as follows:

[0074] First, pre-clustering is performed on the fault feature set under different candidate cluster numbers, and then the DB index corresponding to each clustering result is calculated;

[0075] By comparing the DB index values ​​under different candidate cluster numbers, the cluster number that minimizes the DB index is selected as the optimal number of clusters.

[0076] This method can ensure that the final clustering partition has both good intra-cluster compactness and high inter-cluster separation, thereby improving the objectivity and robustness of the clustering effect.

[0077] In this embodiment, the clustering algorithm is the fuzzy C-means clustering algorithm.

[0078] After obtaining the optimal number of clusters, fuzzy C-means clustering is performed on the fault feature set. The specific process is as follows: First, the positions of each cluster center are randomly initialized, and an initial membership degree is assigned to each sample point. Then, during the iteration process, the membership degrees of the cluster centers and samples are updated alternately. When updating the membership degree, the degree of membership to different clusters is calculated based on the relative distance between the sample point and each cluster center, ensuring that the sum of the membership degrees of the same data point in all clusters is 1. When updating the cluster centers, the cluster centers are recalculated by weighted summation of the membership degrees of all sample points to that cluster. These two steps are iterated continuously until the cluster centers converge or the objective function converges, ultimately yielding a fuzzy membership degree matrix. This matrix comprehensively characterizes the degree of belonging of each fault feature sample relative to different cluster centers, reflecting the fuzziness and uncertainty of the feature distribution.

[0079] To further enhance the contribution of features to classification, this embodiment introduces a feature weighting mechanism into the fuzzy membership matrix:

[0080] The weight of each feature is calculated based on the correlation index of each feature dimension in distinguishing different fault types.

[0081] Assign larger weights to features with high discriminative power, and smaller weights to features that are sensitive to noise or have high redundancy.

[0082] The weights are applied to the fuzzy membership matrix to obtain a weighted feature vector, which highlights the discriminative feature dimensions while preserving the fuzzy membership relationships.

[0083] After obtaining the weighted feature vector, this embodiment utilizes a waveform mapping algorithm to map the abstract high-dimensional weighted feature space back to the original waveform space. Specifically, the waveform mapping algorithm works as follows: the fuzzy waveform features of each fault sample are formed by weighted combination of the typical waveforms corresponding to all cluster centers according to their membership degrees. Specifically, each cluster center represents a typical fault feature, which can be converted into a corresponding time-domain waveform through a mapping function; and each fault sample point has different membership degrees on different cluster centers, which characterize the similarity between the sample and various typical waveforms. By weighted superposition of the time-domain waveforms corresponding to each cluster center according to their membership degrees, a new fuzzy waveform can be generated. This waveform retains the main dynamic features of the original waveform while incorporating the categorical fuzziness information obtained from cluster learning. The resulting fuzzy waveform features can more comprehensively reflect the complexity and uncertainty of fault features, thus providing a more reliable input for subsequent fault type identification.

[0084] S4, based on the fault dataset and cluster centers, trains multiple structurally independent feature forest base classifiers through an ensemble learning algorithm.

[0085] In this embodiment, the step of training multiple structurally independent feature forest base classifiers using an ensemble learning algorithm based on the fault dataset and cluster centers specifically involves:

[0086] The fault dataset is divided into multiple fault data subsets based on cluster centers;

[0087] For each subset of fault data, train an ensemble classifier separately;

[0088] The ensemble classifiers are weighted and optimized using cluster centers to form the final classifier set.

[0089] In this embodiment, the ensemble learning algorithm is the random forest algorithm.

[0090] For the fault dataset, several fault data subsets are generated based on the membership distribution of the cluster centers;

[0091] For each subset of fault data, a feature weight vector is calculated based on the cluster center, and the selection probability of candidate features is adjusted based on the feature weight vector.

[0092] In the process of constructing the decision tree, the weighted Gini impurity or information gain is used as the splitting criterion, and the sample membership degree is combined as the sample weight to participate in the calculation.

[0093] When dividing nodes, candidate features are randomly sampled, and the optimal division threshold is searched within the range of the weighted feature set.

[0094] By setting constraints on maximum depth, minimum number of split samples, and minimum number of leaf node samples, and without performing post-pruning, a structurally independent decision tree is obtained.

[0095] Multiple decision trees are integrated into a feature forest base classifier, and the sample sampling method, feature selection ratio and cluster center weight are iteratively adjusted during training until the classification performance meets the preset convergence condition.

[0096] Multiple trained feature forest base classifiers are combined to form a feature forest base classifier set.

[0097] In this embodiment, the weighted optimization of each ensemble classifier through cluster centers includes: adjusting the weights of features in the fault data subset based on the membership degree of the cluster centers.

[0098] During the training process, the fuzzy membership degree of the fault feature set under each cluster center is first obtained;

[0099] Based on the fuzzy membership degree, a corresponding weight coefficient is assigned to each feature in the fault data subset, so that the weight of features with higher membership degree increases during training, and the weight of features with lower membership degree decreases during training.

[0100] Based on this, the adjusted feature weights are introduced into the training process of the feature forest base classifier to make weighted corrections to the splitting criteria and feature selection probabilities of the decision tree.

[0101] Through the above weighted correction, each feature forest base classifier can more fully reflect the fault mode features represented by the cluster centers during the training process, thereby improving the classifier's ability to distinguish different types of faults.

[0102] S5. Input the fuzzy waveform features into the feature forest base classifier to obtain a preliminary recognition result set.

[0103] In this embodiment, the step of determining cluster centers and constructing fuzzy waveform features based on the fault feature set using a clustering algorithm specifically involves:

[0104] The fuzzy waveform features are input in parallel into multiple structurally independent feature forest base classifiers;

[0105] In each feature forest base classifier, the corresponding classification result is obtained based on the probability output mechanism;

[0106] The outputs of all base classifiers are aggregated to form a preliminary recognition result set.

[0107] The fuzzy waveform features are fed one by one into multiple structurally independent feature forest base classifiers;

[0108] Each base classifier contains several decision trees. After inputting features, each tree sequentially performs feature partitioning and path determination from the root node to the leaf node, and finally obtains the corresponding class distribution at the leaf node.

[0109] In each of the feature forest base classifiers, the corresponding classification result is obtained based on a probability output mechanism, which specifically includes:

[0110] Collect the class distribution of leaf nodes of all decision trees within the base classifier;

[0111] The outputs of all decision trees are weighted or averaged to calculate the probability estimate of the base classifier for each candidate fault type.

[0112] The probability estimates of each candidate fault type are combined to form a probability distribution vector, which is then used as the classification result of the base classifier.

[0113] By collecting the probability estimates of each candidate fault type from all feature forest base classifiers, a preliminary identification result set is obtained.

[0114] S6, based on the preliminary identification result set and the power grid topology relationship, uses a probabilistic graphical model algorithm to perform fusion reasoning to obtain the fault type identification result.

[0115] In this embodiment, the step of training multiple structurally independent feature forest base classifiers using an ensemble learning algorithm based on the fault dataset and cluster centers specifically involves:

[0116] A probabilistic graphical model is constructed based on the distribution network topology, and the nodes of the probabilistic graphical model correspond to candidate fault types;

[0117] The preliminary identification results are used as evidence and input into the probabilistic graphical model. The posterior probability of each candidate fault type is obtained through probabilistic reasoning.

[0118] The final fault type identification result is determined based on the posterior probability.

[0119] First, a Bayesian network structure needs to be constructed based on the actual topology of the distribution network. This Bayesian network consists of a set of directed acyclic graphs (DAGs), where nodes represent candidate fault types and their associated state variables, and edges characterize the dependencies between fault types and network topology constraints. For example, nodes located on the same feeder branch may have strong conditional correlations between their fault types, which is reflected in the network structure through directed edges. This process explicitly encodes the spatial association information and electrical logic characteristics of the distribution network into the Bayesian network, enabling fault reasoning to not only rely on classification probabilities but also reflect topological constraints.

[0120] After the Bayesian network structure is established, the preliminary identification result set obtained in step S5 is input into the network as the observation variable. Specifically, the probability distribution vector output by each feature forest base classifier is mapped to the observation node corresponding to the candidate fault type, serving as the basis for updating the prior probability of that type. In this way, the Bayesian network can perform probability fusion of multi-classifier results and form a globally consistent inference process by combining the topological structure. The inference process uses Bayes' theorem to calculate the posterior probability of each candidate fault type, that is, to re-estimate the probability of each type occurring under the condition of input observation information.

[0121] The calculated posterior probability distribution is then used to rank the candidate fault types, and the ranking result is the final fault type identification output. During the ranking process, if the posterior probability of a certain type is significantly higher than that of other types, it is directly output as the final identification result; if the probabilities of multiple types are close, the judgment can be further refined by setting confidence intervals to ensure the stability and interpretability of the results.

[0122] In this embodiment, the step of inputting the preliminary identification result set as an observation variable into a Bayesian network to obtain the probability distribution of each candidate fault type is specifically as follows:

[0123] The preliminary identification results set is normalized to obtain an observation probability vector at a uniform scale;

[0124] Input the observation probability vector into the observation node corresponding to the candidate fault type in the Bayesian network to obtain the prior probability update result of each candidate fault type.

[0125] Based on the prior probability update results and combined with the relevant constraints of fault location in the distribution network topology, the posterior probability distribution of each candidate fault type is inferred.

[0126] First, the preliminary identification result set is normalized to ensure that the outputs of each base classifier are unified within a common scale. This process aims to eliminate scale differences between the output probabilities of different base classifiers, making their probability outputs comparable. Specifically, the probability values ​​of each classifier in the preliminary identification result set are readjusted so that their sum is 1, thus obtaining an observation probability vector at a unified scale. Through normalization, the final observation probability vector reflects the comprehensive probability estimate of each candidate fault type across all base classifiers.

[0127] The normalized observation probability vector is input into the Bayesian network model as an observation variable for inference. A Bayesian network is a graph model based on probability theory that can express and reason about uncertain information through a network structure. In a Bayesian network, each candidate fault type is represented as a node, and the edges between nodes represent their conditional dependencies. By inputting the normalized observation probability vector into the nodes corresponding to the candidate fault types in the Bayesian network, the network updates the prior probabilities of each candidate fault type based on the observation information. This update process is based on Bayes' theorem, which adjusts the probability distribution of each fault type using known observation data (preliminary identification results).

[0128] After obtaining the prior probabilities updated based on the initial identification results, the next step is to infer the fault type by incorporating the distribution network's topology. The distribution network topology typically includes information such as the location of the fault, the network structure, and the connections between nodes. This topological information plays a crucial role in constraining fault type identification. For example, if a candidate fault type appears in a specific area of ​​the distribution network, the Bayesian network can utilize the network's topological knowledge to make a more accurate inference about the fault type. By combining constraint information related to the fault location, the Bayesian network can adjust the probability distribution of each candidate fault type, thereby more accurately inferring the most probable fault type. Based on the Bayesian network's inference results, the posterior probability distribution of each candidate fault type is obtained.

[0129] Example 2, Figure 2 The present invention provides a fault type identification system based on fuzzy waveforms, comprising:

[0130] The data acquisition and construction module is used to collect multi-source fault data from the power distribution network and construct fault datasets.

[0131] The feature extraction module is used to perform mode decomposition on the fault dataset and extract a set of fault features.

[0132] The fuzzy waveform construction module is used to construct fuzzy waveform features by determining cluster centers through a clustering algorithm based on a set of fault features.

[0133] The classifier training module is used to train multiple structurally independent feature forest base classifiers based on the fault dataset and cluster centers through an ensemble learning algorithm.

[0134] The preliminary identification module is used to input the fuzzy waveform features into the feature forest base classifier to obtain a set of preliminary identification results.

[0135] The topology association and fusion module is used to perform fusion reasoning based on the preliminary identification result set and the power grid topology relationship, and obtain the fault type identification result through probabilistic graphical model algorithm.

[0136] Example 3, Figure 3 The present invention discloses a fault type identification device based on fuzzy waveforms, comprising at least one processor; and an input / output interface communicatively connected to the at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the fault type identification method based on fuzzy waveforms.

[0137] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0138] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0139] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0140] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0141] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0142] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault type identification method based on fuzzy waveforms, characterized in that, Includes the following steps: Collect multi-source fault data from the power distribution network and construct a fault dataset; Modal decomposition is performed on the fault dataset to extract a set of fault features. Based on the fault feature set, cluster centers are determined using a clustering algorithm to construct fuzzy waveform features; Based on the fault dataset and cluster centers, multiple structurally independent feature forest base classifiers are trained using an ensemble learning algorithm. The fuzzy waveform features are input into the feature forest base classifier to obtain a preliminary set of recognition results; Based on the preliminary identification results set and the power grid topology, the fault type identification results are obtained by fusion reasoning through probabilistic graphical model algorithm; Extracting the fault feature set, specifically: Modal decomposition is performed on the fault dataset to obtain the intrinsic mode function components; Calculate the information entropy characteristics and instantaneous characteristics of the components; The information entropy features and instantaneous features are fused using a feature fusion algorithm to obtain a fault feature set. The feature fusion algorithm is as follows: The component energy entropy features and instantaneous features are concatenated to form an initial fused feature vector containing both component energy entropy features and standardized instantaneous features; Weights are assigned to each initial feature based on its variance contribution: the variance of each feature in the historical fault samples is calculated; the normalized variance is then used as the weight of the corresponding feature. The initial fused feature vector is optimized by weighted summation to form a fault feature set; Constructing fuzzy waveform features, specifically: Based on the fault feature set, the optimal number of clusters is determined through clustering effectiveness analysis, and the DB index method is introduced for adaptive determination of the number of clusters. By combining the fault feature set and the optimal number of clusters, a fuzzy membership matrix is ​​generated using a fuzzy clustering algorithm; The fuzzy membership matrix is ​​subjected to feature weighting to obtain a weighted eigenvector; The weighted feature vectors are converted into fuzzy waveform features using a feature reconstruction algorithm.

2. The fault type identification method based on fuzzy waveforms according to claim 1, characterized in that, The process of collecting multi-source fault data from the distribution network and constructing a fault dataset involves the following steps: A time synchronization algorithm is used to process multi-source fault data to obtain time-aligned fault data. Based on time-aligned fault data, a fault dataset is obtained through a data normalization algorithm.

3. The fault type identification method based on fuzzy waveforms according to claim 1, characterized in that, The process involves training multiple structurally independent feature forest base classifiers using an ensemble learning algorithm based on the fault dataset and cluster centers. Specifically: The fault dataset is divided into multiple fault data subsets based on cluster centers; For each subset of fault data, train an ensemble classifier separately; The ensemble classifiers are weighted and optimized using cluster centers to form the final classifier set.

4. The fault type identification method based on fuzzy waveforms according to claim 3, characterized in that, The weighted optimization of each ensemble classifier through cluster centers includes: during training, adjusting the weights of features in the fault data subset based on the membership degree of the cluster centers.

5. The fault type identification method based on fuzzy waveforms according to claim 4, characterized in that, The step of determining cluster centers and constructing fuzzy waveform features based on the fault feature set using a clustering algorithm is as follows: The fuzzy waveform features are input in parallel into multiple structurally independent feature forest base classifiers; In each feature forest base classifier, the corresponding classification result is obtained based on the probability output mechanism; The outputs of all base classifiers are aggregated to form a preliminary recognition result set.

6. The fault type identification method based on fuzzy waveforms according to claim 5, characterized in that, The process involves training multiple structurally independent feature forest base classifiers using an ensemble learning algorithm based on the fault dataset and cluster centers. Specifically: A probabilistic graphical model is constructed based on the distribution network topology, and the nodes of the probabilistic graphical model correspond to candidate fault types; The preliminary identification results are used as evidence and input into the probabilistic graphical model. The posterior probability of each candidate fault type is obtained through probabilistic reasoning. The final fault type identification result is determined based on the posterior probability.

7. A system using the fault type identification method based on fuzzy waveforms as described in any one of claims 1-6, characterized in that, include: The data acquisition and construction module is used to collect multi-source fault data from the power distribution network and construct fault datasets. The feature extraction module is used to perform mode decomposition on the fault dataset and extract a set of fault features. The fuzzy waveform construction module is used to construct fuzzy waveform features by determining cluster centers through a clustering algorithm based on a set of fault features. The classifier training module is used to train multiple structurally independent feature forest base classifiers based on the fault dataset and cluster centers through an ensemble learning algorithm. The preliminary identification module is used to input the fuzzy waveform features into the feature forest base classifier to obtain a set of preliminary identification results. The topology association and fusion module is used to perform fusion reasoning based on the preliminary identification result set and the power grid topology relationship, and obtain the fault type identification result through probabilistic graphical model algorithm.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and an input / output interface communicatively connected to the at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the fuzzy waveform-based fault type identification method as described in any one of claims 1 to 6.

Images