CN116908618A - A method for diagnosing AC series arc fault in low-voltage distribution network - Google Patents

Abstract

The application aims to provide a low-voltage distribution network alternating current series arc fault diagnosis method, which is used for collecting alternating current signals of two periods at an outlet of an ammeter at a certain sampling rate; the method comprises the steps of dividing a load waveform into a resistive load waveform, a dimmer load waveform and a complex type load waveform by using the spectrum amplitude and extremum weighted fuzzy entropy value difference of the normalized current waveform through a load waveform classification technology; for resistive load waveforms and dimmer load waveforms, five time-frequency features are extracted and arc fault diagnosis is performed using a conventional machine learning classifier. Aiming at complex load waveforms, a one-dimensional new data set is constructed by using data random fusion operation, and a convolutional neural network called a multichannel attention mechanism is trained to finish the identification of arc faults. Compared with the existing detection algorithm, the alternating current series arc fault diagnosis method provided by the application has the advantages of good adaptability and high accuracy.

Description

A method for diagnosing AC series arc fault in low-voltage distribution network

Technical field

The invention belongs to the technical field of low-voltage distribution network arc fault diagnosis, and specifically relates to a low-voltage distribution network AC series arc fault diagnosis method.

Background technique

With economic development, the scale of low-voltage distribution networks is getting larger and larger, and the rapid increase in distribution lines has led to a rapid increase in the possibility of electrical accidents. The stability of the distribution network has been greatly challenged. The distribution network is an important part of the power network. As the terminal where the power system supplies power to users, the stability of its operation directly affects the reliability of power supply to users. When the insulation layer of low-voltage distribution lines is damaged or aged due to long use or external force, it is likely to cause series and parallel arcs, and fault arcs can easily cause electrical fires. It can be seen that the harm caused by arc fault cannot be underestimated. Not only that, when a fault arc occurs, its own impedance and radiation effects will cause harmonic interference, affect power quality, damage electrical equipment, and affect power supply economy.

At present, the existing low-voltage distribution network AC series arc fault diagnosis algorithms are relatively diverse, which can be mainly summarized as time-frequency domain detection methods and neural network diagnosis methods. Time-frequency domain detection extracts the characteristics of the current signal in the time and frequency domains, and uses a classifier or threshold method to complete the identification of AC series arc faults. The neural network diagnosis method uses the one-dimensional waveform of the data or converts the data into images and inputs them into various networks for training and testing. Under the influence of different loads, fault arc current waveforms are diverse, and there are similarities between the normal and fault waveforms of various loads, resulting in low recognition accuracy and poor robustness of the current algorithm. In addition, due to the uncertainty of neural network feature learning, using one-dimensional current data to input into the network for training will result in learning redundant information or causing normal and fault features to be submerged, resulting in poor diagnostic effect of the neural network detection algorithm. .

Contents of the invention

Therefore, in view of the defects and shortcomings of the existing technology, the present invention proposes a new method for diagnosing AC series arc faults in low-voltage distribution networks, specifically a two-level classification series arc fault detection algorithm. The advantage of this algorithm is to first classify twelve load waveforms through a classification strategy to weaken the waveform similarity between loads, thereby reducing the number of misjudgment samples caused by similar waveforms and improving the accuracy of arc fault diagnosis. At the same time, a data random fusion mechanism is proposed to enhance the original waveform features, and a neural network with an attention mechanism is constructed for diagnosis.

It collects two cycles of AC current signals at the meter outlet at a certain sampling rate; through a load waveform classification technology, the difference between the spectrum amplitude of the normalized current waveform and the extreme value weighted fuzzy entropy value is used to classify the load waveform into Resistive load waveforms, dimmer load waveforms and complex load waveforms; for resistive and dimmer load waveforms, five time-frequency features are extracted and conventional machine learning classifiers are used for arc fault diagnosis. For complex load waveforms, random data fusion operations are used to construct a one-dimensional new data set and a multi-channel attention mechanism convolutional neural network is trained to complete the identification of arc faults. Compared with existing detection algorithms, the AC series arc fault diagnosis method proposed by the present invention has good adaptability and high accuracy.

The technical solutions specifically adopted by the present invention to solve the technical problems are:

A low-voltage distribution network AC series arc fault diagnosis method, which is characterized by using the difference between the spectrum amplitude of the normalized current waveform and the extreme weighted fuzzy entropy value to distinguish the load waveform into a resistive load waveform and a dimmer load waveform. and complex load waveforms; for resistive and dimmer load waveforms, extract multiple time-frequency features and use machine learning classifiers for arc fault diagnosis; for complex load waveforms, use random data fusion operations to construct a one-dimensional new data set and Train a multi-channel attention mechanism convolutional neural network to complete the identification of arc faults.

Furthermore, using the difference between the spectrum amplitude of the normalized current waveform and the extreme value weighted fuzzy entropy value to classify the load waveform into a resistive load waveform, a dimmer load waveform and a complex load waveform specifically includes the following process:

The two-cycle AC current signal X(i) at the meter outlet is collected at a certain sampling rate and normalized according to equation (1);

Among them, X _L is the normalized waveform sequence, n is the number of sampling points in two periods, X _{max and}

Obtain the normalized waveform, use Fourier transform, calculate the fundamental wave amplitude fc and make a judgment. If fc is greater than β ₀ , it is the waveform of a resistive load; where β ₀ is the threshold for classifying resistive load waveforms, β ₀ =0.95;

If it is not a resistive load waveform, calculate the extreme weighted fuzzy entropy F _ew value to determine whether the load waveform is a dimmer load waveform. The specific calculation steps are as follows:

1) Calculate the difference sequence of two-cycle current waveform samples according to equation (2);

X(n)＝abs(X _L (i+1)-X _L (i)),i＝1,2,3···n-1 (2)

Among them, X(n) is the obtained difference sequence, and abs(.) is the absolute value operation;

2) Take the four largest peaks in the difference sequence X(n) and calculate _Pav according to equation (3);

Among them, M(1), M(2), M(3), and M(4) are the four largest peaks respectively;

3) Calculate the fuzzy entropy value FE;

4) Divide the calculated P _av and FE values to obtain the extreme weighted fuzzy entropy F _ew value, as shown in equation (4);

If the calculated F _ew value is less than β ₁ , then the waveform is the current waveform of the dimmer load, β ₁ is the classification threshold that can separate the dimmer load waveform, and β ₁ =0.38;

If fc is less than β ₀ and F _ew is greater than β ₁ , the load waveform is classified into the complex load waveform pool.

1. A low-voltage distribution network AC series arc fault diagnosis method according to claim 2, characterized in that:

The specific steps to calculate the fuzzy entropy value are as follows:

Construct the current waveform {X(i):1<i<N} with length N into an m-dimensional vector according to equation (5), that is

Y _i ^m ={X(i),X(i+1),···,X(i+m-1)}-X ₀ (i) (5)

In the formula, N is the number of samples, where 1≤i≤N-m+1;

According to equation (6), the similarity between sample Y _i ^m and sample Y _j ^m is calculated as D _ij ^m

In the formula, d _ij ^m is the maximum absolute difference, 1≤j, j≤Nm and i≠j;

Define the function according to equation (7)

In the formula, similarity tolerance r, membership function

Construct an m+1 dimensional vector according to equation (8)

Calculate the fuzzy entropy of sampling point N according to equation (9)

FE(m,n,r,N)=lnφ ^m (n,r)-lnφ ^m+1 (n,r) (9).

Furthermore, five time-frequency domain features were extracted for the resistive and dimmer load waveforms, namely pulse factor C _if , margin factor C _mf , kurtosis C _k , fundamental wave relative content proportion F ₁ , third harmonic The relative content proportion F ₃ is as follows:

Among _them , The amplitude of the fifth harmonic component after transformation,

To utilize machine learning classifiers for arc fault diagnosis.

Furthermore, for complex load waveforms, a random data fusion operation is used to construct a one-dimensional new data set and a multi-channel attention mechanism convolutional neural network is trained to complete the identification of arc faults;

The specific steps for performing random data fusion operations on complex load waveforms and constructing a one-dimensional new data set are as follows:

1) In order to retain the utilization of the overall information, two periodic data point samples are selected as an overall waveform;

2) Divide the local waveform information and divide an entire sample into four parts, each part containing half a cycle of waveform data, and each part of the local waveform contains a peak/trough or zero rest position;

3) Construct one-dimensional new data, randomly select two pieces of local information and splice them after the overall waveform to construct one-dimensional new data.

Further, the multi-channel attention mechanism convolutional neural network is a multi-channel convolutional neural network that combines the multi-head attention mechanism and the GRU unit:

The network constructs three data extraction channels, and sets different structural parameters to extract features from data of different lengths to ensure that each data segment can extract the optimal feature group; among them, the first channel is used to extract features from two data segments. Features are extracted from the overall waveform of the cycle. A multi-head attention module is first embedded in the beginning of the network. The attention mechanism is used to eliminate noise and redundant data in the data and focus on features to improve the robustness of the model.

Furthermore, the calculation process of the multi-head attention mechanism is as follows:

1) Divide the input current waveform into multiple segments in parallel;

2) Generate three corresponding matrices for each piece of input data, namely the query matrix, the key matrix, and the sum value matrix;

3) Use the dot product of the query matrix and the key matrix to obtain the weight score of the data;

4) The weight score is normalized by softmax and then multiplied by the value matrix to obtain the final data with weighted attention;

Q＝AW ^Q (16)

K＝AW ^K (17)

V＝AW ^V (18)

Among them, Q, K and V represent query matrix, key matrix and value matrix respectively, d _k is the dimension of Q, K and V, A is the input matrix, W ^Q , W ^K and W ^V are obtained by linear transformation of the input matrix. coefficient matrix;

5) Combine the learned multiple attention segments to obtain global attention;

After passing through the attention layer, the features are re-extracted through the convolution layer, pooling layer, BN layer and Drouput layer, and a GRU unit is connected to prevent over-fitting during training; finally, a BN layer and a Drouput layer are passed through The relu activation function outputs a feature vector;

The second and third channels are used to extract features from local waveform segments of the half cycle; directly through the convolution layer, pooling layer, BN layer and Drouput layer, and connected to a GRU unit, and finally output the feature vector through the relu activation function ; The Drouput layer is used to help the network prevent overfitting, and the activation function of the convolution layer uses leakyrelu.

Furthermore, when the sampling rate changes, the number of overall waveform sample points and the number of local waveform sample points will change. The number and size of the convolutional layers in the three channels are compared and adjusted according to the number of input data points. The parameters of the network are trained through Obtain.

Compared with the existing technology, the present invention and its preferred solutions have at least the following outstanding advantages:

(1) A fast and efficient load waveform classification algorithm is innovatively designed to achieve the initial screening of load types where series arc fault waveforms occur. This technology effectively reduces the computational stress of subsequent neural network models, reduces the number of misjudgment samples caused by similar waveforms between loads, and helps improve the overall diagnostic accuracy.

(2) Innovatively propose a data random fusion mechanism for arc fault diagnosis. On the basis of using the overall waveform of the one-dimensional original data, the local waveform is extracted and fused with the overall information to construct a new one-dimensional data set. This mechanism enhances the utilization of waveform information, effectively amplifies the weak characteristics of arc faults, and improves the recognition accuracy of the back-end detection algorithm.

(3) Innovatively improve the convolutional neural network (CNN), integrate the multi-head attention module and the GRU unit into the CNN, and construct multiple feature extraction channels to separately analyze the overall waveforms and local waveforms of different lengths. Expand adaptive feature extraction and propose a MC-MGCNN network. The network has the ability to efficiently learn a variety of load waveforms and identify whether arc faults occur.

Description of the drawings

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments:

Figure 1 is an overall flow chart of the low-voltage distribution network AC series arc fault diagnosis algorithm according to the embodiment of the present invention;

Figure 2 is a flow chart of the first-level classification technology according to the embodiment of the present invention;

Figure 3 is a diagram of data random fusion operation according to the embodiment of the present invention;

Figure 4 is a MC-MGCNN network structure diagram according to the embodiment of the present invention;

Figure 5 is a flow chart of MC-MGCNN network parameter training according to the embodiment of the present invention;

Figure 6 is a flow chart of the second-level classification technology S3 according to the embodiment of the present invention;

Figure 7 is a schematic diagram of the resistive load waveform classification threshold β ₀ according to the embodiment of the present invention;

Figure 8 is a schematic diagram of the dimmer load waveform classification threshold _β1 according to the embodiment of the present invention;

Figure 9 is a schematic diagram of a confusion matrix according to an embodiment of the present invention.

Detailed ways

In order to make the features and advantages of this patent more obvious and easy to understand, examples are given below and explained in detail as follows:

It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meanings commonly understood by one of ordinary skill in the art to which this application belongs.

It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular forms are also intended to include the plural forms unless the context clearly indicates otherwise. Furthermore, it will be understood that when the terms "comprises" and/or "includes" are used in this specification, they indicate There are features, steps, operations, means, components and/or combinations thereof.

The design and specific usage steps of the proposed low-voltage distribution network AC series arc fault diagnosis method are introduced below through a specific embodiment:

The load types applicable to the proposed technology include resistive loads, motor loads, power supply loads, gas discharge lamp loads, dimmer loads and the above-mentioned mixed loads. The overall process of construction and implementation of the proposed technology is shown in Figure 1. First, the series arc fault data is obtained according to step S1, and then the first-level load waveform is classified into step S2, which is divided into simple load waveforms and complex load waveforms. After that, step S3 is used to perform the second-level fault diagnosis classification.

(1) S1: Collect the two-cycle AC current signal X(i) at the meter outlet at a certain sampling rate, and perform normalization processing according to equation (1);

Among them, X _L is the normalized waveform sequence, n is the number of sampling points in two periods, and X _max and X _min are respectively the maximum value and the minimum value in the two period current signals.

(2) First level classification S2:

Step S21: Obtain the normalized waveform in S1, use Fourier transform, calculate the fundamental wave amplitude fc and make a judgment. If fc is greater than β ₀ , it is the waveform of a resistive load. Where β ₀ is the threshold for classifying resistive load waveforms, β ₀ =0.95;

Step S22: If it is not a resistive load waveform, calculate the extreme weighted fuzzy entropy F _ew value to determine whether the load waveform is a dimmer load waveform. The specific calculation steps are as follows:

1) Calculate the difference sequence of two-cycle current waveform samples according to equation (2);

X(n)＝abs(X _L (i+1)-X _L (i)),i＝1,2,3···n-1 (2)

Among them, X(n) is the obtained difference sequence, and abs(.) is the absolute value operation;

2) Take the four largest peaks in the difference sequence X(n) and calculate _Pav according to equation (3);

Among them, M(1), M(2), M(3), and M(4) are the four largest peaks respectively;

3) Calculate the fuzzy entropy (FE) value. The specific steps are as follows:

①Construct the current waveform {X(i):1<i<N} with length N into an m-dimensional vector according to equation (4), that is

Y _i ^m ={X(i),X(i+1),···,X(i+m-1)}-X ₀ (i) (4)

In the formula, N is the number of samples, where 1≤i≤N-m+1;

②According ^to equation (5), the similarity between sample Y _im and sample Y _j ^m is calculated as D _ij ^m

In the formula, d _ij ^m is the maximum absolute difference, 1≤j, j≤Nm and i≠j;

③Define the function according to equation (6)

In the formula, similarity tolerance r, membership function

④ Repeat the above steps to construct an m+1 dimensional vector according to equation (7)

⑤ Calculate the fuzzy entropy of sampling point N according to equation (8)

FE(m,n,r,N)＝lnφ ^m (n,r)-lnφ ^m+1 (n,r) (8)

4) Divide the calculated P _av and FE values to obtain the extreme weighted fuzzy entropy F _ew value, as shown in equation (9);

If the calculated F _ew value is less than β ₁ , then the waveform is the current waveform of the dimmer load, β ₁ is the classification threshold that can separate the dimmer load waveform, and β ₁ =0.38.

Step S23: If fc is less than β ₀ and F _ew is greater than β ₁ , the load waveform is classified into the complex load waveform pool, and the resistive and dimmer load waveforms are called simple load waveforms.

(3) Second level classification S3:

Step S31: Obtain the load waveform classification result of S2;

Step S32: Extract five time-frequency domain features for the simple load waveform, namely pulse factor C _if , margin factor C _mf , kurtosis C _k , fundamental wave relative content proportion F ₁ , and third harmonic relative content proportion F ₃ , its formula is as follows:

Among _them , The amplitude of the fifth harmonic component after transformation,

Step S33: Use conventional machine learning classifiers including but not limited to random forest, SVM, XGBoost, etc. to diagnose arc faults for simple load waveforms;

Step S34: Perform data random fusion operations on complex load waveforms to construct a one-dimensional new data set. The specific steps are as follows:

1) Preserve the utilization of the overall information and select two periodic data point samples as an overall waveform;

2) Divide the local waveform information and divide an entire sample into four parts, each part containing half a period of waveform data, and each part of the local waveform will contain a peak/trough or zero rest position;

3) Construct one-dimensional new data, randomly select two pieces of local information and splice them after the overall waveform to construct one-dimensional new data.

The overall operation diagram is shown in Figure 3.

Step S35: For the load waveform of the complex load pool, based on the convolutional neural network (CNN), design a multi-channel convolutional neural network (MC-MGCNN network) that combines the multi-head attention mechanism and the GRU unit. The specific improvements to the network are as follows:

The network constructs three data extraction channels, and sets different structural parameters to extract features from data of different lengths, ensuring that the optimal feature group can be extracted from each data segment. Among them, the first channel is used to extract features from the overall waveform of two cycles. A multi-head attention module is first embedded in the beginning of the network. The attention mechanism is used to eliminate noise and redundant data in the data, focus on features, and improve the model's performance. robustness. The calculation process of the multi-head attention mechanism is as follows:

1) Divide the input current waveform into multiple segments in parallel;

2) Generate three corresponding matrices for each piece of input data, namely query matrix (Query), key matrix (Key), and value matrix (Value);

3) Use the dot product of the query matrix and the key matrix to obtain the weight score of the data;

4) The weight score is normalized by softmax and then multiplied by the value matrix to obtain the final data with weighted attention;

Q＝AW ^Q (16)

K＝AW ^K (17)

V＝AW ^V (18)

Among them, Q, K and V represent query matrix, key matrix and value matrix respectively, d _k is the dimension of Q, K and V, A is the input matrix, W ^Q , W ^K and W ^V are obtained by linear transformation of the input matrix. coefficient matrix.

5) Combine the learned multiple attention segments to obtain global attention.

After passing through the attention layer, features are re-extracted through the convolution layer, pooling layer, BN layer and Drouput layer, and a GRU unit is connected to prevent over-fitting during training. Finally, a BN layer and a Drouput layer are passed through, and the feature vector is output through the relu activation function.

The second and third channels are used to extract features from local waveform segments of half cycles. Directly through the convolution layer, pooling layer, BN layer and Drouput layer, and connected to a GRU unit, and finally output the feature vector through the relu activation function. The Drouput layer can help the network prevent overfitting, and the activation function of the convolution layer uses leakyrelu. The overall structure of the MC-MGCNN network is shown in Figure 4, and the specific network structure is shown in Table 1.

When the sampling rate changes, the number of overall waveform sample points and the number of local waveform sample points will change. The number and size of the convolutional layers in the three channels can be appropriately compared and adjusted according to the number of input data points. The MC-MGCNN network The parameters are obtained through training, and the training process is shown in Figure 5:

Table 1 MC-MGCNN network structure

Step S36: Use the MC-MGCNN network to diagnose the complex load pool waveform.

The following provides a calculation example analysis to further introduce the solution of the present invention:

In this calculation example, the data samples obtained are all from step S1. The case analysis selected 11 types of loads, including 9 single loads and 2 mixed types of loads. The specific parameters are shown in Table 2, in which the dimmer load is divided into small angle (dimming angle is 60°) and large angle (dimming angle is 300°). Each load collected 900 groups under normal and fault conditions, and a total of 21,600 groups of data samples were collected.

Table 2 Experimental load introduction

The current transformer CPL8100A is clamped on the distribution line at the outlet of the electric meter, and is used with the DSOX4024A oscilloscope to collect the current signal. First, discuss the threshold value of the first-level load waveform classification. The selection of the resistive load waveform classification threshold is shown in Figure 7. When β ₀ is set to 0.95, the current waveforms of the resistive load and other loads can be separated to the greatest extent. . After the resistive load waveform is separated, continue to calculate the F _ew value to classify the dimmer load waveform and the remaining single load waveforms. As shown in Figure 8, when β ₁ is set to 0.38, the boundary between the dimmer load waveform and other load waveforms is most obvious.

A total of 1,500 sets of samples of hot water kettles and incandescent lamps were used for testing. The final accuracy was 100%. The current waveforms of both were correctly classified as resistive load waveforms. A total of 1,500 sets of waveform samples from two angles of the dimmer load were selected for testing. Finally, 1,495 sets were correctly classified, with an accuracy rate of 99.7%. The load waveform was successfully determined to be the dimmer load waveform. In order to continue to verify the effectiveness of load waveform classification, two combined load types (switching power supply vacuum cleaner in parallel and incandescent lamp in parallel with fluorescent lamp) were extracted for verification. In the end, both load waveforms were successfully classified into the complex load waveform pool, and the accuracy is 100%.

Next, the effect of the second-level diagnostic classification is verified. Due to different load types, the resistive and dimmer load waveforms are first diagnosed and verified. In this case, the XGBoost classifier is used for verification. The diagnostic accuracy of the resistive load waveform is 100%, and the diagnosis of the dimmer load waveform is accurate. The rate is 99.9%. At the same time, the diagnosis time of both is around 33.4ms. The results show that the classified load waveform can diagnose whether an arc fault has occurred quickly and with high accuracy.

Secondly, the effect of the MC-MGCNN network used for complex load waveform pool diagnosis is verified, and the confusion matrix shown in Figure 9 is obtained. It can be obtained from the confusion matrix that the final multi-classification accuracy is 99.02%, but if only the normal and fault classification states are considered, the final accuracy can reach 99.9%. This model can 100% identify the normal conditions of various load waveforms. However, under fault conditions, 8 samples of the vacuum cleaner parallel switching power supply combined load waveform were misjudged as vacuum cleaner load waveforms, and 3 samples of the vacuum cleaner load waveform were misjudged. It was misjudged to be the combined load waveform of a parallel switching power supply of a vacuum cleaner. This is because the arc current of the combined load is the sum of the branch currents of the two loads. It is affected by the arc characteristics of the vacuum cleaner load. The arc current of the combined load of the vacuum cleaner parallel switching power supply shows a waveform similar to a triangle wave, and at zero break The parts have large burrs and distortion, which makes the two waveforms have a certain similarity, causing the model to be unable to correctly identify the waveform type of the load. Finally, the diagnosis time of this model is around 50.1ms.

Those skilled in the art will understand that embodiments of the present application may be provided as methods, systems, or computer program products. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment that combines software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each process and/or block in the flowchart illustrations and/or block diagrams, and combinations of processes and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions executed by the processor of the computer or other programmable data processing device produce a use A device for realizing the functions specified in one process or multiple processes of the flowchart and/or one block or multiple blocks of the block diagram.

These computer program instructions may also be stored in a computer-readable memory that causes a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction means, the instructions The device implements the functions specified in a process or processes of the flowchart and/or a block or blocks of the block diagram.

These computer program instructions may also be loaded onto a computer or other programmable data processing device, causing a series of operating steps to be performed on the computer or other programmable device to produce computer-implemented processing, thereby executing on the computer or other programmable device. Instructions provide steps for implementing the functions specified in a process or processes of a flowchart diagram and/or a block or blocks of a block diagram.

The above are only preferred embodiments of the present invention, and are not intended to limit the present invention in other forms. Any skilled person familiar with the art may make changes or modifications to equivalent changes using the technical contents disclosed above. Example. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the protection scope of the technical solution of the present invention.

This patent is not limited to the above-mentioned best embodiments. Under the inspiration of this patent, anyone can come up with various other forms of AC series arc fault diagnosis methods for low-voltage distribution networks. Anyone who is within the scope of the patent application of this invention can Equal changes and modifications shall fall within the scope of this patent.

Claims (8)

1. The low-voltage distribution network alternating current series arc fault diagnosis method is characterized in that the load waveform is divided into a resistive load waveform, a dimmer load waveform and a complex type load waveform by utilizing the spectrum amplitude and extremum weighted fuzzy entropy value difference of the normalized current waveform; aiming at the resistive load waveform and the dimmer load waveform, extracting various time-frequency characteristics and performing arc fault diagnosis by using a machine learning classifier; aiming at complex load waveforms, a one-dimensional new data set is constructed by using data random fusion operation, and a multichannel attention mechanism convolutional neural network is trained to finish the identification of arc faults.

2. The method for diagnosing an arc fault in ac series connection of a low voltage distribution network according to claim 1, wherein the method comprises the steps of:

the load waveform is divided into a resistive load waveform, a dimmer load waveform and a complex type load waveform by utilizing the difference of the frequency spectrum amplitude and the extremum weighted fuzzy entropy value of the normalized current waveform, and the method specifically comprises the following steps:

collecting alternating current signals X (i) of two periods at the outlet of the ammeter at a certain sampling rate, and carrying out normalization processing according to a formula (1);

wherein X is _L Is normalized toWaveform sequence, n is the number of sampling points of two periods, X _max 、X _min Respectively a maximum value and a minimum value in the two periodic current signals;

acquiring normalized waveform, calculating fundamental wave amplitude fc by Fourier transform, and judging if fc is larger than beta ₀ Then the waveform is a resistive load; wherein beta is ₀ Is the threshold value of the classified resistive load waveform, beta ₀ ＝0.95；

If the waveform is not the resistive load waveform, the extremum weighted fuzzy entropy F is calculated _ew The value, judge whether this load waveform is dimmer load waveform, the concrete calculation steps are as follows:

1) Calculating a difference sequence of two-period current waveform samples according to the formula (2);

X(n)＝abs(X _L (i+1)-X _L (i)),i＝1,2,3···n-1 (2)

wherein X (n) is the obtained difference sequence, and abs ()'s is the absolute value operation;

2) Taking four largest peaks in the difference sequence X (n) and calculating P according to the formula (3) _av ；

Wherein M (1), M (2), M (3), M (4) are respectively four maximum peaks;

3) Calculating a fuzzy entropy value FE;

4) To calculate P _av Obtaining extremum weighted fuzzy entropy F by dividing FE value _ew A value represented by formula (4);

if F is calculated _ew A value less than beta ₁ The waveform is the current waveform of the dimmer load, beta ₁ To be able to separate the classification threshold of the dimmer load waveform, and beta ₁ ＝0.38；

If fc is less than beta ₀ And F _ew Greater than beta ₁ The load waveform is classified into a complex load waveform pool.

3. The method for diagnosing an arc fault in ac series connection in a low voltage distribution network according to claim 2, wherein the method comprises the steps of:

the specific steps for calculating the fuzzy entropy value are as follows:

the current waveform { X (i): 1< i < N } with length N is constructed into m-dimensional vector according to the formula (5), namely

Y _i ^m ＝{X(i),X(i+1),···,X(i+m-1)}-X ₀ (i) (5)

Wherein N is the number of samples, wherein i is more than or equal to 1 and less than or equal to N-m+1;

obtaining sample Y according to (6) _i ^m And sample Y _j ^m Similarity is D _ij ^m

Wherein d _ij ^m The maximum absolute difference value is 1-j, j-N-m and i is not equal to j;

defining a function according to (7)

In the formula, the similarity tolerance r and the membership function

Construction of an m+1-dimensional vector according to (8)

Obtaining fuzzy entropy with sampling point N according to (9)

FE(m,n,r,N)＝lnφ ^m (n,r)-lnφ ^m+1 (n,r) (9)。

4. A method for diagnosing an ac series arc fault in a low voltage distribution network according to claim 3, wherein:

five time-frequency domain features, respectively pulse factor C, are extracted for resistive and dimmer load waveforms _if Margin factor C _mf Kurtosis C _k Relative content of fundamental wave F ₁ Third harmonic relative content ratio F ₃ The formula is as follows:

wherein X is _L For current sample data, n is the signal length, fc is the amplitude of the fundamental component after fourier transform, 3rdhc is the amplitude of the third harmonic component after fourier transform, 5rdhc is the amplitude of the fifth harmonic component after fourier transform,

to make arc fault diagnosis using a machine learning classifier.

5. The method for diagnosing an arc fault in ac series connection in a low voltage distribution network according to claim 4, wherein the method comprises the steps of:

aiming at complex load waveforms, constructing a one-dimensional new data set by using data random fusion operation and training a multichannel attention mechanism convolutional neural network to finish the identification of arc faults;

the specific steps of constructing a one-dimensional new data set are as follows:

1) To preserve the utilization of the overall information, two data point samples of the cycle are selected as an overall waveform;

2) Dividing local waveform information, dividing an integral sample into four parts, wherein each part comprises half-period waveform data, and each part of local waveform comprises a wave crest/wave trough or zero rest part;

3) And constructing one-dimensional new data, randomly selecting two pieces of local information, splicing the two pieces of local information on the whole waveform, and constructing the one-dimensional new data.

6. The method for diagnosing an arc fault in ac series connection of a low voltage distribution network according to claim 1, wherein the method comprises the steps of:

the multichannel attention mechanism convolutional neural network is a multichannel convolutional neural network integrating a multi-head attention mechanism and GRU units:

three data extraction channels are constructed by the network, and different structural parameters are respectively set for extracting features of data with different lengths so as to ensure that each data segment can extract an optimal feature group; the first channel is used for extracting features of the whole waveform of two periods, a multi-head attention module is embedded in the beginning part of the network, noise and redundant data in the data are eliminated by using an attention mechanism, and the features are focused on, so that the robustness of the model is improved.

7. The method for diagnosing an arc fault in ac series connection in a low voltage distribution network according to claim 6, wherein the method comprises the steps of:

the calculation process of the multi-head attention mechanism is as follows:

1) Dividing an input current waveform into a plurality of sections in parallel;

2) Generating three corresponding matrixes for each section of input data, wherein the three corresponding matrixes are respectively a query matrix, a key matrix and a value matrix;

3) Obtaining the weight fraction of the data by utilizing the dot product of the query matrix and the key matrix;

4) The weight fraction is normalized by softmax and multiplied by a value matrix to obtain the final data with weight attention;

Q＝AW ^Q (16)

K＝AW ^K (17)

V＝AW ^V (18)

wherein Q, K, V represent a query matrix, a key matrix and a value matrix, d, respectively _k The dimensions of Q, K, V, A being the input matrix, W ^Q 、W ^K 、W ^V Is a coefficient matrix obtained by performing linear transformation on an input matrix;

5) Combining the learned multiple sections of attention to obtain global attention;

after passing through the attention layer, extracting features through the convolution layer, the pooling layer, the BN layer and the output layer, and connecting a GRU unit at the same time to prevent fitting in training; finally, a characteristic vector is output through a BN layer and a Droutput layer and through a relu activation function;

the second and third channels are used for extracting features from the half-period local waveform segments; directly passing through a convolution layer, a pooling layer, a BN layer and a Droutput layer, connecting with a GRU unit, and finally outputting a feature vector through a relu activation function; the Droutput layer is used to help the network prevent overfitting, and the activation function of the convolution layer uses the leakyrelu.

8. The method for diagnosing an arc fault in ac series connection in a low voltage distribution network according to claim 7, wherein the method comprises the steps of: when the sampling rate is changed, the number of the integral waveform sample points and the number of the local waveform sample points are changed, the number and the size of the convolution layers in the three channels are compared and adjusted according to the number of the input data points, and the parameters of the network are obtained through training.