Arc fault detection method based on improved empirical mode decomposition algorithm

Abstract

The invention provides an arc fault detection method based on an improved empirical mode decomposition algorithm, and the method comprises the steps: collecting an arc fault signal through building a benchmark test platform, carrying out the preprocessing, and carrying out the decomposition of the signal through employing a completely adaptive noise set empirical mode decomposition algorithm, according to the method, useful intrinsic mode functions containing main characteristics and information of arc fault signals are screened out, then the signals are reconstructed, spectrum characteristics are extracted, and finally, a DBSCAN clustering algorithm is used for carrying out clustering analysis on characteristic vectors so as to judge whether the arc faults occur or not. The modal aliasing phenomenon and the endpoint effect are effectively avoided, and the accuracy of arc fault detection is improved. Meanwhile, through adaptive spectrum segmentation and feature extraction, arc fault signals can be accurately identified under strong noise interference.

Description

Technical Field

[0001] The invention belongs to the technical field of power system protection and fault diagnosis, and specifically relates to an arc fault detection method based on an improved empirical mode decomposition algorithm. Background Art

[0002] Arc faults are a common but dangerous fault type in power systems. They are difficult to detect and can easily cause safety accidents such as fires and equipment damage. Traditional arc fault detection methods are mostly based on qualitative analysis or simple spectrum feature extraction. These methods have poor detection accuracy and robustness under strong noise interference or complex power grid conditions and cannot meet the needs of safe operation of modern power systems.

[0003] As a data-driven signal decomposition method, empirical mode decomposition (EMD) has significant advantages in processing nonlinear and non-stationary signals. However, the EMD method has problems such as mode aliasing and endpoint effect, which will lead to a decrease in the accuracy of the decomposition results. To solve these problems, improved algorithms such as complete ensemble empirical mode decomposition (CEEMDAN) were proposed, but the random noise introduced by them may lead to instability in signal decomposition.

[0004] The fully adaptive noise ensemble empirical mode decomposition (ICEEMDAN) further optimizes the decomposition process, significantly reduces the modal aliasing phenomenon by adaptively adjusting the noise component, and greatly improves the decomposition accuracy and algorithm robustness. However, signal decomposition alone is still not enough to accurately identify arc faults, so it is necessary to combine feature extraction and pattern recognition technology to further improve the detection effect.

[0005] In addition, current clustering algorithms such as K-Means are highly dependent on initial parameters and cannot effectively handle noise points. The density clustering algorithm (DBSCAN) does not require a preset number of clusters and has high robustness to noise and outliers, so its application in signal clustering analysis has gradually attracted attention.

[0006] Based on this, how to combine the improved signal decomposition algorithm with the efficient pattern recognition method to achieve accurate detection of arc fault signals has become one of the key issues in current technical research. In view of the above problems, the present invention proposes an arc fault detection method based on ICEEMDAN decomposition and DBSCAN clustering to solve the shortcomings of the prior art. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide an arc fault detection method based on an improved empirical mode decomposition algorithm, which uses a fully adaptive noise set empirical mode method to decompose the preprocessed signal, screens and analyzes the obtained series of intrinsic mode functions, extracts spectral features through signal reconstruction, calculates its corresponding spectral features, and uses DBSCAN clustering to find a specific cluster that matches the arc fault signal, so as to accurately determine whether an arc fault occurs.

[0008] In order to realize the above technical features, the object of the present invention is achieved as follows: an arc fault detection method based on an improved empirical mode decomposition algorithm, comprising:

[0009] Step 1, signal acquisition:

[0010] Build a benchmark test platform including power supply system and multi-load circuit configuration, simulate multiple fault conditions, and deploy high-resolution sampling devices at key nodes of the circuit to record the current or voltage signal at the moment of arc fault;

[0011] Step 2, signal preprocessing:

[0012] Preprocessing the amplified and denoised arc fault signal first removes the DC component and eliminates the DC offset in the signal, making the signal fluctuate around a zero mean value, and performing normalization processing to scale the signal amplitude to a specific range;

[0013] Step 3, signal ICEEMDAN decomposition:

[0014] The preprocessed signal is subjected to fully adaptive noise set empirical mode decomposition. First, the Z group of standard normally distributed white noise W Z (n) The IMF components are obtained by empirical mode decomposition, and then the standard deviation of the jth IMF component of the white noise is determined to be proportional to the original signal X j-1 The ratio of the standard deviations of (n), i.e., ε j ;

[0015] Step 4, Screening IMFs:

[0016] The useful IMFs containing the main features and information of the arc fault signal are determined by calculating the ratio of the energy of each IMF to the total energy, setting an energy ratio threshold, and discarding IMFs with energy ratios lower than the threshold;

[0017] Step 5, signal reconstruction:

[0018] After screening out the useful IMFs related to the arc fault, the signal is reconstructed by adding the IMF components in order;

[0019] Step 6, feature extraction:

[0020] First, calculate the kurtosis coefficient γ of the reconstructed signal, then perform discrete Fourier transform on it to obtain their spectrum representation, and calculate their corresponding spectrum characteristics and frequency variance Reconstruct the feature vector

[0021] Step 7: Apply DBSCAN algorithm clustering:

[0022] Use DBSCAN algorithm to cluster feature vectors;

[0023] Step 8, arc fault judgment:

[0024] First, the clustering results of normal signals and arc fault signals are obtained through multiple experimental data. The specific cluster matching the arc fault signal is found according to the characteristic statistics of each cluster. If the characteristic value points obtained after reconstructing the signal are distributed near the specific cluster related to the arc fault, it is determined that an arc fault has occurred.

[0025] The calculation formula for removing the DC component in step 2 is:

[0026]

[0027] X i =x i -μ;

[0028] Where μ is the mean value of the signal, x i is the original signal, X i is the signal after DC is removed;

[0029] The normalization formula is:

[0030]

[0031] In the formula, x i is the i-th sample point of the original signal, min(x) and max(x) are the minimum and maximum values ​​of the signal respectively, new min and new max are the minimum and maximum values ​​of the new range.

[0032] In step 3, the jth component of the white noise is multiplied by ε j Then added to the original signal to form Z groups of preprocessed signals:

[0033] X z (n) = X j-1 (n)+ε j ×E j [W z (n)];

[0034] Where: Operator E j (·) represents the jth IMF component of a signal after EMD decomposition. The Z groups of preprocessed signals are decomposed by EMD to obtain IMF 1,j 、IMF 2,j ,…IMF Z,j , and finally calculate the average value of all Z group decomposition results:

[0035]

[0036] After the jth iteration, the original signal X j-1 (n) minus IMF j (n) Get the residual signal X j (n), repeat the above process until the residual signal is small enough or reaches the preset number of iterations, that is, the expression of the signal sequence X(n) after decomposition by ICEEMDAN is as follows:

[0037]

[0038] X M (n) is the residual sequence after the decomposition stops.

[0039] In step 4, the energy of each IMF is first calculated:

[0040]

[0041] In the formula, IMF k [i] is the i-th sample point of the k-th IMF, and M is the total number of sample points;

[0042] Then calculate the total energy of all IMFs: Where K is the total number of IMFs, then the energy ratio of each IMF is calculated: Energy Ratio k It represents the ratio of the energy of the kth IMF to the total energy; according to the energy ratio R k and threshold γ to filter IMFs, if R k >γ, the kth IMF is retained, otherwise it is discarded.

[0043] The series of IMF components obtained by the algorithm decomposition in step 5 are arranged from high to low frequency. These IMF components each contain the characteristic information of different frequency segments in the original signal. The signal is reconstructed in the order of their frequencies from high to low. In the reconstruction process, each IMF component is usually stored in an array or vector, and then a loop or accumulator is used to traverse these IMF components and add them to obtain the reconstructed signal, that is:

[0044]

[0045] Where N is the total number of IMF components, IMF i represents the i-th IMF component.

[0046] The calculation formula of the kurtosis coefficient in step 6 is:

[0047]

[0048] Where: μ is the mean value of the signal, σ 2 is the variance of the signal;

[0049] The discrete Fourier transform is defined as:

[0050]

[0051] Where X[k] is the frequency domain representation, k is the frequency index, and N is the signal length;

[0052] The power spectral density calculation formula is:

[0053]

[0054] Where P[k] is the power spectral density PSD of the signal, and X[k] is the DFT of the signal;

[0055] The frequency variance is calculated as:

[0056]

[0057] In the formula, is the mean of the PSD.

[0058] The present invention has the following beneficial effects:

[0059] The invention provides an arc fault detection method based on an improved empirical mode decomposition algorithm. The preprocessed signal is decomposed by a fully adaptive noise set empirical mode method, and a series of intrinsic mode functions obtained are screened and analyzed. The spectrum features are extracted, and the spectrum representation is obtained by Fourier transform. The corresponding spectrum features are calculated, and then DBSCAN clustering is performed to find a specific cluster matching the arc fault signal to determine whether the arc fault occurs. The ICEEMDAN algorithm is used to decompose the signal, and the mode aliasing phenomenon and the endpoint effect are effectively avoided through adaptive spectrum segmentation, thereby improving the accuracy of arc fault detection. In addition, the EMD is used to decompose the white noise, and a certain IMF component is selected as the noise component, thereby avoiding the excessive noise problem that may be introduced by directly adding Gaussian white noise, and the degree of noise introduction can be better controlled. At the same time, DBSCAN does not need to specify the number of clusters in advance, and can identify the noise points, classify them separately or exclude them, and is robust to noise and outliers. By finding a specific cluster matching the arc fault signal, it is possible to accurately determine whether the arc fault occurs, and it is also simpler and more intuitive. Since arc fault current signals are broadband signals and have a large overlap with the spectrum of interference noise signals generated during normal system operation, this method can effectively identify arc faults under strong noise interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0061] Figure 1 It is a schematic diagram of the process of the present invention.

[0062] Figure 2 Schematic diagram of the ICEEMDAN algorithm flow.

[0063] Figure 3 Time domain waveforms of some IMFs decomposed by ICEEMDAN.

[0064] Figure 4 This is the result diagram obtained by directly extracting feature quantities from the signal and clustering them.

[0065] Figure 5 This is the result graph obtained by extracting and clustering features after the signal is decomposed and reconstructed by ICEEMDAN. DETAILED DESCRIPTION

[0066] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0067] Step 1, signal acquisition:

[0068] Build a benchmark test platform including power supply system and multi-load circuit configuration, simulate multiple fault conditions, and deploy high-resolution sampling devices at key nodes of the circuit to record the current or voltage signal at the moment of arc fault;

[0069] Step 2, signal preprocessing:

[0070] Preprocessing the amplified and denoised arc fault signal first removes the DC component and eliminates the DC offset in the signal, making the signal fluctuate around a zero mean value, and performing normalization processing to scale the signal amplitude to a specific range;

[0071] Step 3, signal ICEEMDAN decomposition:

[0072] The preprocessed signal is subjected to fully adaptive noise set empirical mode decomposition. First, the Z group of standard normally distributed white noise W Z (n) The IMF components are obtained by empirical mode decomposition, and then the standard deviation of the jth IMF component of the white noise is determined to be proportional to the original signal X j-1 The ratio of the standard deviations of (n), i.e., ε j ;

[0073] Step 4, Screening IMFs:

[0074] The useful IMFs containing the main features and information of the arc fault signal are determined by calculating the ratio of the energy of each IMF to the total energy, setting an energy ratio threshold, and discarding IMFs with energy ratios lower than the threshold;

[0075] Step 5, signal reconstruction:

[0076] After screening out the useful IMFs related to the arc fault, the signal is reconstructed by adding the IMF components in order;

[0077] Step 6, feature extraction:

[0078] First, calculate the kurtosis coefficient γ of the reconstructed signal, then perform discrete Fourier transform on it to obtain their spectrum representation, and calculate their corresponding spectrum characteristics and frequency variance Reconstruct the feature vector

[0079] Step 7: Apply DBSCAN algorithm clustering:

[0080] Use DBSCAN algorithm to cluster feature vectors;

[0081] Step 8, arc fault judgment:

[0082] First, the clustering results of normal signals and arc fault signals are obtained through multiple experimental data. The specific cluster matching the arc fault signal is found according to the characteristic statistics of each cluster. If the characteristic value points obtained after reconstructing the signal are distributed near the specific cluster related to the arc fault, it is determined that an arc fault has occurred.

[0083] Furthermore, the calculation formula for removing the DC component in step 2 is:

[0084]

[0085] X i =x i -μ;

[0086] Where μ is the mean value of the signal, x i is the original signal, X i is the signal after DC is removed;

[0087] The normalization formula is:

[0088]

[0089] In the formula, x i is the i-th sample point of the original signal, min(x) and max(x) are the minimum and maximum values ​​of the signal respectively, new min and new max are the minimum and maximum values ​​of the new range.

[0090] Among them, in step 2, removing the DC component and normalizing the signal can eliminate the DC offset and amplitude difference in the signal, making the signal purer and more standardized, thereby improving the accuracy of subsequent signal analysis, and enhancing the robustness of the algorithm, reducing its sensitivity to noise and interference.

[0091] Furthermore, in step 3, the jth component of the white noise is multiplied by ε j Then added to the original signal to form Z groups of preprocessed signals:

[0092] X z (n) = X j-1 (n)+ε j ×E j [W Z (n)];

[0093] Where: Operator E j (·) represents the jth IMF component of a signal after EMD decomposition. The Z groups of preprocessed signals are decomposed by EMD to obtain IMF 1,j 、IMF 2,j ,…IMF Z,j , and finally calculate the average value of all Z group decomposition results:

[0094]

[0095] After the jth iteration, the original signal X j-1 (n) minus IMF j (n) Get the residual signal X j (n), repeat the above process until the residual signal is small enough or reaches the preset number of iterations, that is, the expression of the signal sequence X(n) after decomposition by ICEEMDAN is as follows:

[0096]

[0097] X M (n) is the residual sequence after the decomposition stops.

[0098] Among them, in the above step 3, the ICEEMDAN algorithm is developed based on the complete ensemble empirical mode decomposition (CEEMDAN) of adaptive noise. Unlike CEEMDAN, ICEEMDAN does not directly add Gaussian white noise during the decomposition process, but selects the jth intrinsic mode function (IMF) component of the white noise after empirical mode decomposition (EMD) for decomposition. In this way, ICEEMDAN can better control the degree of noise introduction, thereby improving the decomposition effect of the signal.

[0099] like Figure 2 As shown in , the process approaches a target value or satisfies a specific condition by continuously updating the variable X(n) and calculating its intrinsic mode function (IMF) components. First, X(n) is initialized to a specific value and updated in each iteration by adding a small perturbation related to the expected value of W(n) and its square. During the iteration, a series of IMF components are calculated in each iteration by applying the empirical mode decomposition (EMD) algorithm, and these components are used to update X(n). The iteration process continues until the residual signal is small enough or the preset number of iterations is reached, resulting in the final signal sequence X(n).

[0100] like Figure 3 As shown in the figure, the IMFs obtained by the ICEEMDAN algorithm after the preprocessed signal are expressed in different frequencies and amplitudes. This shows that the ICEEMDAN algorithm can decompose the original signal into multiple components with different frequency and amplitude characteristics. By observing the amplitude changes of IMFs, we can understand the energy proportion of different frequency components in the signal, and further infer information such as the source, propagation mode, and interaction of the signal.

[0101] In step 3, through a large number of experimental measurements, a threshold ε can be set. The threshold is a positive decimal used to measure the energy or amplitude of the residual signal. n |X j (n)| 2 <ε, the residual signal is considered small enough to stop the iteration. On the other hand, in order to prevent the iterative process from failing to converge to a small residual signal in some cases, thereby avoiding an infinite loop. The number of iterations can also be preset as a positive integer N to limit the maximum number of iterations. When k>N, the iterative process will stop regardless of the size of the residual signal.

[0102] Furthermore, in step 4, the energy of each IMF is first calculated:

[0103]

[0104] In the formula, IMF k [i] is the i-th sample point of the k-th IMF, and M is the total number of sample points;

[0105] Then calculate the total energy of all IMFs: Where K is the total number of IMFs, then the energy ratio of each IMF is calculated: Energy Ratio k It represents the ratio of the energy of the kth IMF to the total energy; according to the energy ratio R k and threshold γ to filter IMFs, if R k >γ, the kth IMF is retained, otherwise it is discarded.

[0106] Among them, in step 4, by calculating the proportion of the energy of each IMF to the total energy, it is possible to identify which IMFs are most critical to the representation of the arc fault signal, and retain these IMFs that are most useful for arc fault detection, providing a basis for subsequent signal reconstruction and feature extraction.

[0107] Furthermore, the series of IMF components decomposed by the algorithm in step 5 are arranged from high to low frequency, such as Figure 3 As shown in , these IMF components each contain the characteristic information of different frequency bands in the original signal. The signal is reconstructed in order of their frequencies from high to low. During the reconstruction process, each IMF component is usually stored in an array or vector, and then a loop or accumulator is used to traverse these IMF components and add them together to obtain the reconstructed signal, that is:

[0108]

[0109] Where N is the total number of IMF components, IMF i represents the i-th IMF component.

[0110] Furthermore, the kurtosis coefficient γ in step 6 is used as an important statistic to first measure the sharpness of the signal. It can effectively identify anomalies or extreme values ​​in the signal, which are often closely related to arc faults. Subsequently, in order to deeply analyze the spectral components of the signal, the discrete Fourier transform (DFT) is used to convert the signal from the time domain to the frequency domain. Through DFT, specific frequency features related to arc faults can be clearly identified, which is crucial for fault detection. On this basis, the introduction of power spectral density (PSD) further helps to identify the power distribution of different frequency components in the signal, thereby analyzing the characteristics of the arc fault in more detail. Finally, the frequency variance, as an indicator reflecting the degree of concentration of the signal spectrum, helps to distinguish normal signals from fault signals.

[0111] The calculation formula of the kurtosis coefficient is:

[0112]

[0113] Where: μ is the mean value of the signal, σ 2 is the variance of the signal;

[0114] The discrete Fourier transform is defined as:

[0115]

[0116] Where X[k] is the frequency domain representation, k is the frequency index, and N is the signal length;

[0117] The power spectral density calculation formula is:

[0118]

[0119] Where P[k] is the power spectral density PSD of the signal, and X[k] is the DFT of the signal;

[0120] The frequency variance is calculated as:

[0121]

[0122] In the formula, is the mean of the PSD.

[0123] In step six, The features in the time domain and frequency domain are combined to form a feature set that can fully reflect the characteristics of arc faults. These features not only improve the accuracy of detection, but also enhance the robustness of the system under strong noise interference.

[0124] In step seven, since it is difficult for the feature vector to effectively distinguish normal signals from arc fault signals, and it is difficult to identify specific patterns related to arc faults from mixed signal data, the DBSCAN algorithm is used to cluster the feature vectors. The clustered feature vectors will improve the detection efficiency, reduce the amount of calculation, and do not need to pre-specify the number of clusters. It can cluster dense data sets of any shape, and can find anomalies while clustering. It is insensitive to anomalies in the data set, thereby enhancing the accuracy and reliability of detection.

[0125] like Figure 4 As shown in the figure, the data points are presented in the form of red and black circles, which form a clear group structure on the two-dimensional plane. Some of the points are classified as "category 1", indicating that these points have a high degree of similarity in the feature space. At the same time, noise points and outliers are identified and marked with asterisks. Their positions in the figure are relatively isolated, forming a sharp contrast with the classified points. The kurtosis coefficient, as one of the characteristic dimensions, has an important impact on the clustering results of the data points.

[0126] like Figure 5 As shown in the figure, this figure shows the result of DBSCAN density clustering of the extracted features after ICEEMDAN decomposition and reconstruction, which clearly reveals the clustering relationship between different data points. In the figure, the data points are assigned to two different categories according to the kurtosis coefficient (X-axis) and the distance value (Y-axis): "Category 1" and "Category 2", and are distinguished by different colors. From the distribution of points, the data points of "Category 1" are mainly concentrated in areas with low kurtosis coefficients and wide ranges of distance values, while the data points of "Category 2" are distributed in areas with high kurtosis coefficients and relatively concentrated distance values. This distribution pattern shows that the DBSCAN clustering algorithm clusters data points with similar characteristics together to form obvious clustering areas. Therefore, the DBSCAN algorithm can accurately distinguish normal signals from abnormal signals.

[0127] like Figure 4 and 5As shown in the figure, the clustering of features after decomposition and reconstruction by the ICEEMDAN algorithm shows significant superiority compared with the clustering results of directly extracted features. The ICEEMDAN algorithm effectively extracts the key features in the original signal through a multi-level decomposition and reconstruction process. These features are more discriminative in clustering analysis, making the clustering results clearer and the boundaries between categories more obvious. At the same time, the algorithm shows higher robustness in dealing with noise, can more effectively identify and exclude noise points, and improves the accuracy of clustering analysis. In addition, the ICEEMDAN algorithm also optimizes the feature space, reduces redundant features, and selects the most representative features for clustering, thereby improving the performance of the clustering algorithm. In contrast, the directly extracted features may contain more noise and redundant information, resulting in inaccurate and unstable clustering results.

[0128] In step 8, the detection principle of arc fault is: by collecting current or voltage signals, extracting the intrinsic mode function (IMF) in the signal after preprocessing using algorithms such as empirical mode decomposition, and then extracting characteristic parameters such as kurtosis and root mean square value from IMF, clustering analysis is performed on these characteristic parameters using the DBSCAN clustering algorithm, and classifying the data points into core points, boundary points or noise points according to the number of points in their neighborhood. On this basis, starting from any core point, through the density accessibility relationship, a cluster containing core points, boundary points and indirectly connected data points is gradually expanded. Finally, the DBSCAN algorithm outputs a set containing all clusters, which correspond to the clustering areas of normal signals and arc fault signals. According to the characteristic statistics of each cluster, a specific cluster matching the arc fault signal is found. If the characteristic value points obtained after reconstructing the signal are distributed near the specific cluster related to the arc fault, it is determined that the arc fault has occurred.

[0129] In summary, the present invention provides an arc fault detection method based on an improved empirical mode decomposition algorithm. The method collects arc fault signals by building a benchmark test platform, decomposes the signals using the fully adaptive noise ensemble empirical mode decomposition (ICEEMDAN) algorithm after preprocessing, and screens out useful intrinsic mode functions (IMFs) containing the main features and information of arc fault signals, and then reconstructs the signals and extracts the spectrum features. Finally, the DBSCAN clustering algorithm is used to perform cluster analysis on the feature vectors to determine whether the arc fault occurs. The innovation of the present invention lies in that the ICEEMDAN algorithm is used for signal decomposition, which effectively avoids the mode aliasing phenomenon and the endpoint effect, and improves the accuracy of arc fault detection. At the same time, through adaptive spectrum segmentation and feature extraction, arc fault signals can be accurately identified under strong noise interference. In addition, the application of the DBSCAN clustering algorithm makes the detection process more intelligent and automated, without the need to pre-specify the number of clusters, and is robust to noise and outliers. Experimental results show that the method of the present invention can accurately determine whether an arc fault occurs or not, and has higher detection accuracy and robustness than the method of directly extracting feature quantities for clustering. Therefore, the present invention has broad application prospects in the fields of power systems, electrical equipment fault diagnosis, etc.

[0130] 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 and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An arc fault detection method based on an improved empirical mode decomposition algorithm, characterized in that: include Step 1, signal acquisition: Build a benchmark test platform including power supply system and multi-load circuit configuration, simulate multiple fault conditions, and deploy high-resolution sampling devices at key nodes of the circuit to record the current or voltage signal at the moment of arc fault; Step 2, signal preprocessing: Preprocessing the amplified and denoised arc fault signal first removes the DC component and eliminates the DC offset in the signal, making the signal fluctuate around a zero mean value, and performing normalization processing to scale the signal amplitude to a specific range; Step 3, signal ICEEMDAN decomposition: The preprocessed signal is subjected to fully adaptive noise set empirical mode decomposition. First, the Z group of standard normally distributed white noise W Z (n) The IMF components are obtained by empirical mode decomposition, and then the standard deviation of the jth IMF component of the white noise is determined to be proportional to the original signal X j-1 The ratio of the standard deviations of (n), i.e., ε j ; Step 4, Screening IMFs: The useful IMFs containing the main features and information of the arc fault signal are determined by calculating the ratio of the energy of each IMF to the total energy, setting an energy ratio threshold, and discarding IMFs with energy ratios lower than the threshold; Step 5, signal reconstruction: After screening out the useful IMFs related to the arc fault, the signal is reconstructed by adding the IMF components in order; Step 6, feature extraction: First, calculate the kurtosis coefficient γ of the reconstructed signal, then perform discrete Fourier transform on it to obtain their spectrum representation, and calculate their corresponding spectrum characteristics and frequency variance Reconstruct the feature vector Step 7: Apply DBSCAN algorithm clustering: Use DBSCAN algorithm to cluster feature vectors; Step 8, arc fault judgment: First, the clustering results of normal signals and arc fault signals are obtained through multiple experimental data. The specific cluster matching the arc fault signal is found according to the characteristic statistics of each cluster. If the characteristic value points obtained after reconstructing the signal are distributed near the specific cluster related to the arc fault, it is determined that an arc fault has occurred.

2. The arc fault detection method based on the improved empirical mode decomposition algorithm according to claim 1, characterized in that: The calculation formula for removing the DC component in step 2 is: X i =x i -m; Where μ is the mean value of the signal, x i is the original signal, X i is the signal after DC is removed; The normalization formula is: In the formula, x i is the i-th sample point of the original signal, min(x) and max(x) are the minimum and maximum values ​​of the signal respectively, new min and new max are the minimum and maximum values ​​of the new range.

3. The arc fault detection method based on the improved empirical mode decomposition algorithm according to claim 1, characterized in that: In step 3, the jth component of the white noise is multiplied by ε j Then added to the original signal to form Z groups of preprocessed signals: X z (n)=X j-1 (n)+ε j ×E j [W z (n)]? Where: Operator E j (·) represents the jth IMF component of a signal after EMD decomposition. The Z groups of preprocessed signals are decomposed by EMD to obtain IMF 1,j 、IMF 2,j ,…IMF Z,j , and finally calculate the average value of all Z group decomposition results: After the jth iteration, the original signal X j-1 (n) minus IMF j (n) Get the residual signal X j (n), repeat the above process until the residual signal is small enough or reaches the preset number of iterations, that is, the expression of the signal sequence X(n) after decomposition by ICEEMDAN is as follows: X M (n) is the residual sequence after the decomposition stops.

4. The arc fault detection method based on the improved empirical mode decomposition algorithm according to claim 1, characterized in that: In step 4, the energy of each IMF is first calculated: In the formula, IMF k [i] is the i-th sample point of the k-th IMF, and M is the total number of sample points; Then calculate the total energy of all IMFs: Where K is the total number of IMFs, then the energy ratio of each IMF is calculated: Energy Ratio k It represents the ratio of the energy of the kth IMF to the total energy; according to the energy ratio R k and threshold γ to filter IMFs, if R k >γ, the kth IMF is retained, otherwise it is discarded.

5. The arc fault detection method based on the improved empirical mode decomposition algorithm according to claim 1, characterized in that: The series of IMF components obtained by the algorithm decomposition in step 5 are arranged from high to low frequency. These IMF components each contain the characteristic information of different frequency segments in the original signal. The signal is reconstructed in the order of their frequencies from high to low. In the reconstruction process, each IMF component is usually stored in an array or vector, and then a loop or accumulator is used to traverse these IMF components and add them to obtain the reconstructed signal, that is: Where N is the total number of IMF components, IMF i represents the i-th IMF component.

6. The arc fault detection method based on the improved empirical mode decomposition algorithm according to claim 1, characterized in that: The calculation formula of the kurtosis coefficient in step 6 is: Where: μ is the mean value of the signal, σ 2 is the variance of the signal; The discrete Fourier transform is defined as: Where X[k] is the frequency domain representation, k is the frequency index, and N is the signal length; The power spectral density calculation formula is: Where P[k] is the power spectral density PSD of the signal, and X[k] is the DFT of the signal; The frequency variance is calculated as: In the formula, is the mean of the PSD.

Images