A network-level arc fault diagnosis method based on multi-scale feature fusion

Abstract

The present invention discloses a network-level arc fault diagnosis method based on multi-scale feature fusion, comprising the following steps: 1) current signals of different branches in a current sensor acquisition system; 2) obtaining time domain features, frequency domain features and singular values ​​of current signals of different branches at different scales based on a multi-scale analysis method to construct a feature vector; 3) using a random forest to implement feature importance analysis and selection to obtain an optimal feature combination; 4) using the optimal feature combination and the trained random forest to implement arc fault diagnosis. The present invention extracts fault features based on different scales, provides different perspectives for analyzing features and increases the diversity of features, and uses a random forest algorithm to implement feature selection and fusion, thereby enabling fault information in different types of features to complement each other, and can eliminate redundant information in high-dimensional feature vectors, which is beneficial to improving diagnostic accuracy and reducing the burden on classifiers.

Description

Technical Field

[0001] The invention relates to a network-level arc fault diagnosis method based on multi-scale feature fusion, and belongs to the field of arc fault detection. Technical Background

[0002] With the development of renewable energy, electric vehicles, microgrids, and more electric aircraft, DC power has been used more and more widely. However, in the long-term operation process, the aging of the insulation of the line and the poor contact of the terminal are prone to cause arc faults. In particular, the arc current in the DC distribution network does not have a zero crossing point, which is more likely to produce continuous arcing compared to the AC distribution network. The temperature of the arc column can reach up to 15000℃, which is very easy to ignite nearby combustibles. Therefore, arc faults pose a huge challenge to the safe and stable operation of the DC distribution system. Arc faults are divided into series arc faults and parallel arc faults. When a parallel arc fault occurs in a DC distribution system, the fault line will generate a large short-circuit current and is easily protected by a circuit breaker. When a series arc fault occurs in the system, the current rise of the line where it is located is not obvious due to the introduction of arc impedance, and it is difficult for the circuit breaker to achieve protection. In recent years, the research on series arc fault detection methods has received more and more attention.

[0003] The location where an arc fault occurs is usually accompanied by physical phenomena such as noise, strong light, a large amount of heat, and electromagnetic radiation. Some researchers have implemented series arc fault detection based on the above physical phenomena, but this type of method requires the detection point to be close to the location where the arc fault occurs. Therefore, limited by its small monitoring range, the method based on physical phenomena is difficult to apply to actual systems.

[0004] Most researchers have previously implemented series arc fault detection based on current signals. Some researchers have used the time domain characteristics of current signals, such as standard deviation, to detect arc faults. When an arc fault occurs, the energy distribution of the current signal spectrum will change. Some researchers have extracted the harmonic energy in the range of 48.83kHz-93.99kHz as the characteristic value, and compared the characteristic value with the set threshold to determine whether an arc fault has occurred in the line. When an arc fault occurs, the random volatility of the current signal increases. Some researchers have used the hurst index to measure the randomness of the signal to determine whether an arc fault has occurred in the line. Some researchers have used singular value decomposition to filter out the low-frequency components of the current signal, and then extracted the variance of the filtered current signal as the basis for determining whether an arc fault has occurred. An arc fault will change the size of the characteristic impedance at the fault point in the line. Based on the spread spectrum time domain reflection method, the change of the line characteristic impedance is monitored in real time and the reflected wave head is compared with the set threshold, so as to detect the arc fault. The actual system's working state and the environment in which it is located are complex, and noise interference is easily introduced into the current signal. The arc fault detection method based on a single characteristic threshold is prone to misjudgment. Therefore, more and more researchers are integrating key fault information of different types of features in current signals based on machine learning methods to improve arc fault detection performance, such as artificial neural networks, support vector machines, decision trees, extreme learning machines, and convolutional neural networks.

[0005] Although the feature fusion method based on machine learning can improve the accuracy of arc fault detection, the above research is aimed at arc fault detection of a single branch, that is, it only determines whether there is an arc fault on the current branch. However, the actual distribution system often contains multiple branches, and different lines are interconnected. When an arc fault occurs in a branch in the system, the current signal of the adjacent branch may also be interfered and invaded by high-frequency random noise. Standard UL1699 stipulates that in a system consisting of two parallel lines, when an arc fault occurs in one line, the arc fault circuit breaker installed in the line where the arc fault does not occur cannot make a misjudgment. The use of a single-branch arc fault detection method is prone to false alarms in normal branches and cannot accurately locate the fault point. Therefore, from a system-level perspective, it is necessary not only to detect whether an arc fault occurs in the system, but also to locate the arc fault and avoid false alarms in normal lines.

[0006] At present, some researchers have conducted research on arc fault diagnosis algorithms for distribution networks to accurately detect and locate arc faults. Some researchers have used Kalman filters to identify the parameters of cables in distribution networks to locate series arc faults. This method requires simultaneous detection of the current and terminal voltage of each cable, and uses more sensors. Moreover, the Kalman filter assumes that the noise of the system is white noise, but the noise contained in the arc current is colored noise. Some researchers have used recurrent neural networks to directly process the current of three-phase lines. However, due to excessive superposition of gradients, RNN is prone to gradient vanishing problems during the back propagation process, thus affecting the diagnostic performance. Some researchers have used wavelet transform to achieve high-pass filtering of three-phase current signals for arc fault line selection in three-phase motor systems, and then extracted features based on variational mode decomposition and Wieglar distribution as inputs to support vector machines. However, variational mode decomposition and Wieglar distribution have high computational overhead, which limits the online application of this method. Some researchers have used the time difference of the pulse signal generated by the arc fault in two capacitors connected in parallel at the load end to determine the location of the arc fault in the DC distribution system. However, the high-frequency noise generated during the startup of nonlinear loads can easily interfere with this algorithm. In order to reduce the number of sensors used in the diagnosis process and improve computing efficiency, some researchers have realized the diagnosis of arc faults in distribution networks by analyzing only one branch current signal collected. However, when the load properties of two branches in the system are similar or the same, the characteristics of the arc faults of these two branches in the main circuit current signal will be similar or the same, making it difficult to accurately locate the arc fault.

[0007] In summary, the shortcomings of the methods mentioned above are mainly reflected in two aspects: (1) Some algorithms only analyze the information of one branch and cannot be applied to arc fault diagnosis of multi-branch systems; (2) Some arc fault diagnosis methods based on system-level analysis have low accuracy and cannot effectively adapt to complex noise under different working conditions. Summary of the invention

[0008] In order to improve the accuracy of arc fault diagnosis and avoid false alarms of normal lines, the present invention proposes a network-level arc fault diagnosis method based on multi-scale feature fusion.

[0009] In order to achieve the above-mentioned invention object, the present invention specifically adopts the following technical solutions:

[0010] Step 1): Collect current signals of each branch in the power distribution system and perform preprocessing;

[0011] Step 2): extracting the time domain features, frequency domain features and singular values ​​of the roughness sequence of the current signal of each branch at different scales based on a multi-scale analysis method, and constructing a feature vector;

[0012] Step 3): Analyze the importance of features based on random forest;

[0013] Step 4): Based on the analysis results of feature importance, all features are arranged from large to small according to the importance value, and the features with the highest importance value are used to construct feature vectors and train the random forest classifier. The trained random forest is verified using the test set to obtain the test accuracy, and the feature vector corresponding to the highest test accuracy is selected to form the optimal feature combination. The optimal feature combination and the trained random forest are used to realize arc fault diagnosis.

[0014] The beneficial effects of the present invention are:

[0015] 1. Multi-scale analysis can show the complexity of signals at different time scales. Compared with a single scale, multi-scale analysis can analyze the characteristics of high-frequency and low-frequency components in the signal, extract key fault information from different levels and enhance the richness of subsequent features. Feature extraction at different scales can mine more valuable information, provide different perspectives for analyzing features and increase the diversity of features, which is conducive to improving the accuracy of subsequent classification diagnosis.

[0016] 2. At the same time, the random forest algorithm is used to achieve feature selection and fusion. This makes it possible to comprehensively utilize the fault information in different types of features, achieve complementary advantages, eliminate the interference of redundant information in high-dimensional feature vectors on the diagnosis results, enhance the anti-noise ability, and thus help distinguish complex fault types, improve diagnostic accuracy, and reduce the burden on the classifier. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic diagram of the method flow of Example 1 of the present invention;

[0018] Figure 2 is a block diagram of a DC power distribution system according to Embodiment 2 of the present invention;

[0019] Figure 3 is a schematic diagram of frequency band segmentation according to Embodiment 2 of the present invention;

[0020] Figure 4 is the confusion matrix of the offline diagnosis result of Example 2 of the present invention;

[0021] Figure 5 It is a waveform diagram of the online diagnosis result of Example 2 of the present invention. DETAILED DESCRIPTION

[0022] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Example 1

[0024] This embodiment provides a network-level arc fault diagnosis method based on multi-scale feature fusion, which is applicable to a distribution system having at least two branches, such as Figure 1 As shown, the following steps are included:

[0025] Step 1): Collect and pre-process the current signals of each branch in the power distribution system. The pre-processing method is: subtract the average value from the collected current signals of each branch in the power distribution system, and avoid the interference of DC components on arc fault diagnosis by subtracting the average value.

[0026] Step 2): Use a multi-scale analysis method to obtain the roughness sequence of the current signal of each branch at different scales. The roughness sequence of the current signal at different scales can be expressed as:

[0027]

[0028] Among them, u i is the data point of the current signal, N=1024 is the original length of the current signal; τ=1, 2, ... is the time scale; That is, it represents the roughness sequence of the current signal of branch k at the time scale τ. When τ = 1, It is equivalent to the original signal; as τ increases, it is equivalent to low-pass filtering the signal.

[0029] Step 3): Construct a feature vector by extracting the time domain features, frequency domain features and singular values ​​of the roughness sequence at different scales of the current signal.

[0030] Time domain features: The time domain features of each roughness sequence extracted are 7-dimensional, including standard deviation, skewness index, kurtosis index, waveform factor, peak factor, pulse factor and margin factor.

[0031] Frequency domain features: To extract the frequency domain features of the current signal, first divide the frequency in the frequency band into multiple frequency bands (FBF) on average. Then calculate the power sum in each of the multiple frequency bands to obtain multiple frequency domain features corresponding to each roughness sequence.

[0032] Singular value: To extract the singular value of the signal, we first need to transform the one-dimensional roughness sequence Converted into a two-dimensional matrix, where Len is the data length. The matrix conversion method adopted by the present invention adopts the truncation method, as shown in formula (2).

[0033]

[0034] In the formula, u nm Representation sequence The m×nth data point in n=INT(Len / m), INT(·) is a rounding down operation. After processing the matrix M using the singular value decomposition algorithm, m singular values ​​can be obtained.

[0035] Step 3): Analyze the importance of features based on random forest.

[0036] The feature importance analysis method based on random forest is mainly based on the Gini coefficient of the feature.

[0037] First, through random forest training, the current Gini index of a node is obtained

[0038]

[0039] Where E is the number of categories; p er is the proportion of the e-th category in the r-th node (r∈R, R is the number of nodes in the current decision tree);

[0040] Then select a feature w to divide the current node r, and get two data sets after division. Then get the Gini index corresponding to these two data sets. The difference ΔIG of the Gini index before and after the division of feature w at node r can be obtained. wr

[0041] ΔIG wr =IG r -IG r1 -IG r2 (4)

[0042] Among them, IG r1 With IG r2 The Gini indix of the two new branches obtained after division; ΔIG wr The larger it is, the greater the influence of this feature on node division and the higher its importance;

[0043] The feature importance of feature w in decision tree t is expressed as

[0044] VIM tw =∑ r∈R ΔIG wr (5)

[0045] The overall feature importance VIM of feature w in a random forest containing T decision trees w It can be expressed as

[0046]

[0047] About VIM w Normalize to get the final importance of feature w

[0048]

[0049] Among them, W * is the total number of features; the feature w corresponds to The larger the value, the more important the feature is.

[0050] Step 4) Selection of the optimal combination of features based on random forest.

[0051] Based on the analysis results of feature importance, all features are arranged from large to small according to the importance value. The features ranked top k (k≤369) in importance are used to construct feature vectors and train the random forest classifier. The trained random forest is verified using the test set to obtain the test accuracy. When k is traversed from 1 to 369, 369 sets of test accuracy can be obtained. Among them, if the feature vector composed of the features ranked top s in importance has the highest test accuracy, then the feature vector composed of the features ranked top s in importance is the optimal feature combination.

[0052] Step 5): Use the optimal feature combination and the trained random forest to realize network-level DC series arc fault diagnosis.

[0053] Example 2

[0054] A network-level arc fault diagnosis method based on multi-scale feature fusion comprises the following steps:

[0055] Step 1): First build Figure 2 The DC power distribution system shown in the figure contains three branches. Connecting the arc generator to the system can simulate the occurrence of series arc faults in different lines. Then the current sensor is used to collect the current signals of the three branches and transmit them to the microprocessor through the data acquisition board. The original current signal fragments (chips) of the three branches collected are constructed into a sample. Each chip contains 1024 data points, and a total of 12,000 samples are collected. There are a total of 5 states in the DC distribution network, namely normal, series arc fault in branch 1, series arc fault in branch 2, series arc fault in branch 3, and series arc fault in the main circuit. Each state contains 2,400 samples.

[0056] Step 2): First, preprocess each sample, that is, subtract the average value of each chip. Then use the multi-scale analysis method to obtain the roughness sequence of each chip at three scales. The roughness sequence of the current signal at different scales can be expressed as:

[0057]

[0058] Among them, u iis the data point of the current signal, N=1024 is the original length of the current signal; τ=1, 2, 3 is the time scale; That is, it represents the roughness sequence of the current signal of branch k at the time scale τ;

[0059] Step 2): Extract the roughness sequence of the current signal at different scales The time domain features, frequency domain features and singular values ​​of construct feature vectors, where:

[0060] Time domain features: The time domain features of each roughness sequence extracted are 7-dimensional, including standard deviation, skewness index, kurtosis index, waveform factor, peak factor, pulse factor and margin factor.

[0061] Frequency domain features: Divide the frequencies within the signal frequency band into FC frequency bands (FBF) on average. By calculating the power sums within FC frequency bands respectively, FC frequency domain features corresponding to each roughness sequence can be obtained respectively. Taking the original current signal as an example, based on a sampling frequency of 40kHz, the maximum frequency that can be analyzed using Fourier transform is 20kHz. If the FC value is set too large, the number of features will increase and the computational complexity will increase. If the FC value is too small, the differences between different frequency bands cannot be effectively distinguished. Therefore, considering the computational efficiency and the discrimination performance comprehensively, such as Figure 3 As shown, the frequency in the 20kHz range is evenly divided into 10 frequency bands (FBF), that is, FC = I0. By calculating the power sum in the 10 frequency bands, we can get the power of each frequency band. The corresponding 10 frequency domain features.

[0062] Singular value: To extract the singular value of the signal, we first need to transform the one-dimensional roughness sequence Converted into a two-dimensional matrix, where Len is the data length. The matrix conversion method of the present invention adopts the truncation method, as shown in formula (2).

[0063]

[0064] In the formula, n=INT(Len / m), INT(·) is a rounding operation. After processing the matrix M using the singular value decomposition algorithm, m singular values ​​can be obtained. For each sample, the original signal of the first time scale The 32nd order singular value sv can be obtained i (i is the order, i = 1, 2, ..., 32); for the second time scale You can get 22nd order sv i (i=1,2,…,22); for the second time scale You can get 18th order sv i (i=1, 2,…, 18).

[0065] From the above analysis, it can be seen that for a sample containing three branch current signals, a 369-dimensional feature vector can be obtained after preprocessing, multi-scale analysis and feature extraction.

[0066] Step 3): Analyze the importance of features based on random forest.

[0067] The feature importance analysis method based on random forest is mainly based on the Gini coefficient of the feature. During the training process, the current Gini index of a node can be obtained.

[0068]

[0069] Wherein, E is the number of categories. In the embodiment of the present invention, there are five states in the DC distribution network, namely, normal, series arc fault in branch 1, series arc fault in branch 2, series arc fault in branch 3, and series arc fault in the main circuit. er is the proportion of the e-th category in the r-th node (r∈R, R is the number of nodes in the current decision tree).

[0070] Then select a feature w to divide the current node r, and you can get the two data sets after the division and the Gini index corresponding to the two data sets, thereby getting the difference ΔIG of the Gini index before and after the division of feature w at node r. wr

[0071] ΔIG wr =IG r -IG r1 -IG r2 (4)

[0072] Among them, IG r1 With IG r2 are the Gini coefficients of the two new branches obtained after division; ΔIG wr The larger it is, the greater the influence of this feature on node division and the higher its importance;

[0073] The feature importance of feature w in decision tree t is expressed as

[0074] VIM tw =∑ r∈R ΔIG wr (5)

[0075] The overall feature importance VIM of feature w in a random forest containing T decision trees w It can be expressed as

[0076]

[0077] About VIM wNormalize to get the final importance of feature w

[0078]

[0079] Among them, W * is the total number of features. The feature dimension extracted in the embodiment listed in the present invention is 369, so W * =369. The feature w corresponds to The larger the value, the more important the feature is.

[0080] Step 4) Selection of the optimal combination of features based on random forest.

[0081] Based on the analysis results of feature importance, all features are arranged from large to small according to the importance value. The feature vector composed of the top k features in importance is used to train the random forest classifier based on the data set and test the corresponding diagnostic accuracy. Among them, the feature number k changes from 1 to 369. As the number of features increases, the diagnostic accuracy shows a trend of first increasing and then decreasing. When the feature dimension k is 106, the feature vector constructed based on the top 106 features in importance can achieve the maximum diagnostic accuracy. The results show that smaller dimensions will cause the loss of fault information. More dimensions will introduce redundant information and interfere with diagnosis. Therefore, the feature vector is composed of the top 106 features in importance among the original 369-dimensional features.

[0082] Step 5): Use the optimal feature combination and the trained random forest to realize network-level DC series arc fault diagnosis.

[0083] This embodiment can analyze the characteristics of high-frequency and low-frequency components in the signal by adopting a multi-scale analysis method, and can extract key fault information from different levels and enhance the richness of subsequent features. Secondly, feature selection and fusion are realized based on random forests, so that fault information in different types of features can be comprehensively utilized to achieve complementary advantages, and the interference of redundant information in high-dimensional feature vectors on the diagnosis results can be eliminated, and the anti-noise ability can be enhanced, which is conducive to distinguishing complex fault types, improving diagnostic accuracy and reducing the burden on the classifier.

[0084] Figure 4 The confusion matrix of the offline diagnosis results of the method proposed in this invention is given, and the overall diagnostic accuracy is 98.93%. The vertical axis is the actual state label, and the horizontal axis is the state label obtained after diagnosis. The label distribution of different states is shown in Table 1. Under normal conditions, the diagnostic accuracy of the algorithm proposed in this article is 99.80%, which can effectively reduce the economic losses caused by system downtime. The method proposed in this article can correctly determine the existence of a series arc in the system with an accuracy of 99.84%, and can diagnose the location of the arc fault with an accuracy of 98.71%.

[0085] Table 1 Label allocation in different states

[0086]

[0087] In order to further verify the real-time performance of the proposed algorithm, the proposed algorithm was embedded in Figure 2 In the ARM processor shown in FIG. 1 , in order to enhance the robustness of the algorithm, when the diagnosis results for three consecutive times indicate that a SAF occurs at a certain location, it is finally determined that an arc fault occurs at that location.

[0088] Figure 5 Table 2 shows the diagnosis results under 6 different working conditions: DC-DC startup (R1), load voltage change (R2), arc fault in branch 1 (R3), arc fault in branch 2 (R4), arc fault in branch 3 (R5), arc fault in the main circuit (R6). R1 ​​and R2 are normal conditions, which include transient processes. R3, R4, R5 and R6 are arc faults at different locations in the system.

[0089] The current waveform in the transient process of R1 and R2 contains high-frequency harmonic disturbance components, but the label representing the system state is always 1, and no misdiagnosis occurs. This paper conducted 30 tests in the two cases of R1 and R2, and the diagnostic accuracy was 100%, as shown in Table 2.

[0090] Table 2 Online diagnosis results

[0091]

[0092] Figure 5 (c), (d), (e), and (f) respectively give the diagnosis results of arc faults at different locations in the system, indicating that the algorithm proposed in this paper can timely diagnose the location of the arc fault within 250ms. The diagnosis time of the algorithm proposed in this paper can meet the requirements of standard UL1699B. The method proposed in this paper can not only detect arc faults but also accurately locate arc faults, and obtain more abundant fault information compared to single-branch arc fault detection. At the same time, the computational efficiency still meets the requirements of the single-branch arc fault detection standard, proving that the computational efficiency of the algorithm proposed in this paper is acceptable.

[0093] The experiments were performed 30 times for the four cases of R3, R4, R5 and R6. Only in the case of R3 was there one experiment in which the correct judgment was not made. The accuracy of the cases of R4, R5 and R6 was 100%. The results of the online experiments show that the computational efficiency of the method based on multi-scale feature fusion proposed by the present invention can meet the requirements of the standard and has excellent reliability, robustness and certain engineering application value. If the number of branches in the system increases, the method proposed by the present invention can be made equally applicable by adding sensors to collect the current of the corresponding branches and adding the fault characteristics of the corresponding branches in the feature vector.

[0094] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A network-level arc fault diagnosis method based on multi-scale feature fusion, characterized in that The following steps are involved: Step 1): Collect current signals of each branch in the power distribution system and perform preprocessing; Step 2): extracting the time domain features, frequency domain features and singular values ​​of the roughness sequence of the current signal of each branch at different scales based on a multi-scale analysis method, and constructing a feature vector; Step 3): Analyze the importance of features based on random forest; Step 4): Based on the analysis results of feature importance, all features are arranged from large to small according to the importance value, and the features with the highest importance value are used to construct feature vectors and train the random forest classifier. The trained random forest is verified using the test set to obtain the test accuracy, and the feature vector corresponding to the highest test accuracy is selected to form the optimal feature combination. The optimal feature combination and the trained random forest are used to realize arc fault diagnosis; In step 2), the roughness sequence at different scales is expressed as: Among them, u i is the data point of the current signal, N is the original length of the current signal; τ=1,2,… is the time scale; represents the roughness sequence of the current signal of branch k at the time scale τ; when τ = 1, It is equivalent to the original signal; as τ increases, it is equivalent to low-pass filtering the signal; In step 2), extracting the singular values ​​of the signal includes: Step 2-1): Use the truncation method to convert the one-dimensional roughness sequence Transformed into a two-dimensional matrix M, as shown in formula (2): In the formula, u nm Representation sequence The m×nth data point in n=INT(Len / m), INT(·) is the rounding down operation, Len is the roughness sequence Length; Step 2-2): Use the singular value decomposition algorithm to process the matrix M to obtain m singular values; Step 3) includes: Step 3-1): Get the current Gini index of a node through random forest training Where E is the number of categories; p er is the proportion of the e-th category in the r-th node, r∈R, R is the number of nodes in the current decision tree; Step 3-2): Select a feature w to partition the current node r, obtain two data sets and the Gini index corresponding to the two data sets, and the difference ΔIG of the Gini index before and after the feature w is partitioned at the node r wr It is expressed as: ΔIG wr =I.G. r -IG r1 -IG r2 (4) Among them, IG r1 With IG r2 are the Gini indexes corresponding to the two data sets respectively; ΔIG wr The larger it is, the greater the influence of this feature on node division and the higher its importance; Step 3-3): The feature importance of feature w in decision tree t is expressed as BECAUSE tw =∑ r∈R ΔIG wr (5) The overall feature importance VIM of feature w in a random forest containing T decision trees w Expressed as About VIM w Normalize to get the final importance of feature w Among them, W * is the total number of features; the feature w corresponds to The larger the value, the more important the feature is.

2. The network-level arc fault diagnosis method based on multi-scale feature fusion according to claim 1 is characterized in that In step 1), the preprocessing includes subtracting the average value from the collected current signals of each branch, and the interference of the DC component on the arc fault diagnosis is avoided by subtracting the average value.

3. The network-level arc fault diagnosis method based on multi-scale feature fusion according to claim 1 is characterized in that In step 2), the time domain characteristics of the roughness sequence include standard deviation, skewness index, kurtosis index, waveform factor, peak factor, pulse factor and margin factor.

4. The network-level arc fault diagnosis method based on multi-scale feature fusion according to claim 1 is characterized in that In step 2), extracting the frequency domain features of the current signal includes: evenly dividing the frequencies within the signal frequency band into multiple frequency bands, respectively calculating the power sum in each frequency band, and obtaining multiple frequency domain features corresponding to each roughness sequence.

Images