CN112016473B - Power distribution network high-resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism - Google Patents

Power distribution network high-resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism Download PDF

Info

Publication number
CN112016473B
CN112016473B CN202010894524.5A CN202010894524A CN112016473B CN 112016473 B CN112016473 B CN 112016473B CN 202010894524 A CN202010894524 A CN 202010894524A CN 112016473 B CN112016473 B CN 112016473B
Authority
CN
China
Prior art keywords
sample
lstm
phase voltage
network
neural network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010894524.5A
Other languages
Chinese (zh)
Other versions
CN112016473A (en
Inventor
高博
李帷韬
邓雅丽
丁津津
吴刚
汪玉
李奇越
汪勋婷
孙伟
彭思遥
孙辉
张峰
何开元
陈洪波
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Electric Power Research Institute of State Grid Anhui Electric Power Co Ltd
Hefei University of Technology
Original Assignee
Electric Power Research Institute of State Grid Anhui Electric Power Co Ltd
Hefei University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Electric Power Research Institute of State Grid Anhui Electric Power Co Ltd, Hefei University of Technology filed Critical Electric Power Research Institute of State Grid Anhui Electric Power Co Ltd
Priority to CN202010894524.5A priority Critical patent/CN112016473B/en
Publication of CN112016473A publication Critical patent/CN112016473A/en
Application granted granted Critical
Publication of CN112016473B publication Critical patent/CN112016473B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Abstract

The invention discloses a power distribution network high resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism, which comprises the following steps: 1. classifying the fault phase voltage time-series data set into a marked sample set and an unmarked sample set; 2. updating the two types of sample sets according to the similarity of the unmarked sample set and the marked sample set; 3. training a 1NN classifier by using a labeled sample set, and performing prediction labeling on an unlabeled sample set by using the 1NN classifier; 4. constructing an LSTM-CNN neural network based on an attention mechanism; 5. fusing the feature vectors of the LSTM and the CNN by using an attention mechanism network; 6. and updating the neural network parameters of the LSTM-CNN through a gradient descent back propagation algorithm. The invention enables the fault detection model to have time and space characteristic expression capability, thereby improving the high-resistance grounding fault detection rate of the power distribution network.

Description

Power distribution network high-resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism
Technical Field
The invention relates to the field of relay protection of a power distribution network, a high-resistance grounding fault diagnosis technology and a deep learning technology, in particular to a high-resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism.
Background
The power distribution network is close to users, the operation environment is complex, and the power distribution network is easy to contact with branches, buildings or the ground during operation, so that high-resistance faults are generated. For such a fault, the change of the electrical quantity is not obvious due to the large grounding resistance, and the conventional protection is difficult to operate reliably. Most high-resistance faults can not seriously affect a power distribution network, but if the faults exist for a long time, the system can generate a new grounding point due to overvoltage generated by the faults, so that the accidents are further expanded, and even fire disasters can be caused by electric arcs accompanied by the faults when the faults occur, so that the personal and property safety is threatened. High-resistance fault identification is one of the very challenging difficult problems in the field of power distribution network relay protection.
A power grid intelligent scheduling and control education department key laboratory (Shandong university) Wemingjie et al researches a power distribution network high-resistance ground fault detection method (power system automation, 2020,44(14): 164. charge 175.) based on a zero-sequence current waveform interval slope curve, describes waveform nonlinear distortion by adopting least square linear fitting based on analysis of a 10kV power grid field actual measurement high-resistance fault waveform, simultaneously inhibits interference of irregular waveform distortion on the interval slope curve by adopting a Grubbs method, and further ensures correct extraction of fault characteristics by an algorithm. However, the method is difficult to determine a reasonable setting threshold value aiming at complex operating environment and various fault characteristics, does not have a universal judgment rule, and is poor in sensitivity.
A high-resistance ground fault identification method (an electrical measurement and instrument 2020,57(02):52-56.) for a power distribution network based on PSO and a Bayes classifier is researched by Wenying and Pimping, Chenxiangyu and other people of a SpA-Tujin power supply company of Liaoning Power saving Limited of China, and the method firstly adopts discrete wavelet transformation to construct a time-frequency matrix of voltage and current of the power distribution network and extracts characteristic quantities reflecting the high-resistance ground fault. The method optimizes the characteristic space of the electric quantity data, and improves the classification accuracy and the calculation timeliness. But the selection of the mother wavelet function directly affects the quality of the wavelet transform extracted feature information. Generally, for specific distribution network topology and parameters, a fixed mother wavelet needs to be selected through repeated trial and error, and the method can only achieve a local optimal effect.
Suwen clever, Zhu Xingyu and the like at the university of Fuzhou invented a high-resistance ground fault detection method for a power distribution network based on wavelet transformation and a neural network (publication number: CN 109613402A). The method utilizes an evolved neural network to improve the traditional detection method. The evolved neural network is an intelligent system based on a dynamic connection structure, and the topological structure of the system can be adjusted through incremental learning so as to incorporate new information. The method utilizes discrete wavelet transformation to process fault signals, and the fault signals are input into an evolved neural network, so that the high-resistance grounding fault of the power distribution network is detected. But the unmarked samples in the high-resistance fault detection of the power distribution network are large in quantity and easy to obtain, the marked samples are small in quantity and difficult to obtain, and the unmarked data are much more than the marked data. The method does not adopt unmarked data for training, can not truly represent the distribution characteristics of real data, a classifier can not correctly find the classification boundary of the real data, the method can not fully utilize a sample set, and the fault diagnosis does not achieve the best effect.
Disclosure of Invention
The invention aims to avoid the defects of the prior art, and provides a high-resistance grounding fault diagnosis method for a power distribution network based on semi-supervised learning and attention mechanism, so that when massive high-resistance fault data are faced, fault sample information is fully acquired through the semi-supervised learning mechanism, and characteristics of a sample set on time domain and frequency domain are fused by using an attention mechanism network, thereby improving the high-resistance grounding fault detection rate of the power distribution network and meeting the actual requirement of rapidness and accuracy.
The invention adopts the following technical scheme for solving the technical problems:
the invention relates to a method for diagnosing a high-resistance grounding fault of a power distribution network based on semi-supervised learning and attention mechanism, which is characterized by comprising the following steps of:
step 1: when a group of high-resistance ground faults of the power distribution network are obtained, a fault phase voltage time sequence of a secondary side end of a transformer is obtained and is used as a training sample, and the training sample is recorded as TS ═ TS 1 ,TS 2 ,···,TS i ,···,TS N }; wherein N represents the number of training samples, TS i Representing the ith fault phase voltage time sequence in the training sample, and comprising:
TS i ＝{(t 1 ,x i,1 ),(t 2 ,x i,2 ),···,(t j ,x i,j ),···,(t q ,x i,q )}q represents the length of the ith fault phase voltage time sequence in the training sample, t j Represents the jth time, x, in the training sample i,j Representing the voltage value corresponding to the ith fault phase voltage time sequence in the training sample at the jth time;
step 2, denoising the fault phase voltage time sequence set TS to obtain a preprocessed fault phase voltage time sequence set; classifying the preprocessed fault phase voltage time sequence set into a fault phase voltage time sequence sample set D containing unmarked fault phase voltage u And a time series sample set D containing the marked fault phase voltage p ；
Step 3, defining a variable k, and initializing k to be 1; define the set of unlabeled samples after the kth update as
Figure BDA0002658014910000021
And is initialized to D u The set of marked samples after the kth update is
Figure BDA0002658014910000022
Initialisation to D p ；
Step 4, calculating the unmarked sample set after the kth updating
Figure BDA0002658014910000023
The jth fault phase voltage time sequence sample and the kth updated marked sample set
Figure BDA0002658014910000024
Similarity of (2) R j Thereby obtaining a similarity vector
Figure BDA0002658014910000025
j＝1,2,...m k ，m k For the unlabeled sample set after the kth update
Figure BDA0002658014910000026
The number of samples of (a);
step 5, defining the k-th updateSet of unlabeled samples
Figure BDA0002658014910000027
Single sample and labeled sample set
Figure BDA0002658014910000028
The similarity threshold of (a) is epsilon;
judging the similarity vector R k If all the similarity degrees are smaller than epsilon, executing a step 7 if yes, otherwise, executing a step 6;
step 6, selecting the unmarked sample set after the kth updating
Figure BDA0002658014910000031
Neutralization of labeled sample sets
Figure BDA0002658014910000032
And manually marking the sample with the maximum similarity, and adding the manually marked sample into the marked sample set
Figure BDA0002658014910000033
And obtaining the unmarked sample set after the (k + 1) th update
Figure BDA0002658014910000034
Simultaneous unlabeled sample set
Figure BDA0002658014910000035
Deleting the artificially marked samples to obtain an unmarked sample set updated for the (k + 1) th time
Figure BDA0002658014910000036
Assigning k +1 to k, and returning to the step 4;
step 7, utilizing the marked sample set after the kth updating
Figure BDA0002658014910000037
Training the nearest neighbor 1NN classifier to obtain a trained 1NN classifier for the unknownLabeling a sample set
Figure BDA0002658014910000038
Classifying;
step 8, collecting the unmarked samples after the kth updating
Figure BDA0002658014910000039
Inputting the trained 1NN classifier, and taking the obtained prediction label as a real label of an unlabeled sample;
then, the unlabeled samples with the real labels are collected
Figure BDA00026580149100000310
And labeled sample sets
Figure BDA00026580149100000311
Merging to obtain a high-resistance fault training set T containing N samples;
step 9, constructing an LSTM-CNN neural network based on an attention mechanism;
the LSTM-CNN neural network comprises an LSTM network, a CNN network, a feature fusion layer based on an attention mechanism and an SCN classifier; the number of input nodes of the LSTM network is q multiplied by 1, and the output dimension of the hidden layer is m multiplied by 1; the output dimension of the CNN network hidden layer is m multiplied by 1;
step 10, defining the current iteration number of the network as mu, and initializing mu to 1; maximum number of iterations is mu max (ii) a Carrying out the random initialization for the mu time on the parameters of each layer in the network so as to obtain an LSTM-CNN neural network of the mu time iteration;
step 11, defining a variable i, and initializing to i-1;
step 12, carrying out Fourier transform on the ith fault phase voltage time sequence sample in the high-resistance fault training set T to obtain an image sample X of the ith fault phase voltage time sequence i ；
Step 13, the image sample X i Inputting the CNN network in the LSTM-CNN neural network of the mu iteration to obtain a feature vector C with dimension of m multiplied by 1 i,μ ；
Step 14, selecting the ith fault phase voltage time sequence sample from the high-resistance fault training set T, inputting the ith fault phase voltage time sequence sample into an LSTM network in the LSTM-CNN neural network of the mu iteration, and obtaining a feature vector with dimension q multiplied by m
Figure BDA00026580149100000312
Wherein the content of the first and second substances,
Figure BDA00026580149100000313
inputting the n-th time step representing the ith fault phase voltage time sequence sample into an LSTM network in an LSTM-CNN neural network of the mu-th iteration, and outputting a feature vector with the dimension of m multiplied by 1;
step 15, utilizing the formula (1-1) -formula (1-3) to align the feature vector C i,μ And the feature vector H i,μ Performing fusion to obtain m × 1 feature vector F with dimension i,μ ；
Figure BDA0002658014910000041
Figure BDA0002658014910000042
Figure BDA0002658014910000043
In the formula (1-1) -formula (1-3),
Figure BDA0002658014910000044
the weight matrix of the feature fusion layer of the LSTM-CNN neural network of the mu iteration is a dimension of m multiplied by m;
Figure BDA0002658014910000045
is a bias term for the feature fusion layer of the LSTM-CNN neural network for the μ iteration;
Figure BDA0002658014910000046
representing a feature vector with dimension of m multiplied by 1 output by an LSTM network in an LSTM-CNN neural network of the mu iteration when an ith fault phase voltage time sequence sample is input at the ith time step; alpha is alpha i,μ,n In the feature fusion layer of the LSTM-CNN neural network representing the μ iteration
Figure BDA0002658014910000047
Corresponding fusion weight coefficients;
step 16, the feature vector F i,μ Inputting the SCN classifier in the LSTM-CNN neural network of the mu iteration to obtain an output result t' i,μ Will output the result t' i,μ And the desired output t i Performing subtraction to obtain the output error e of the SCN classifier i,μ ；
Step 17, assigning i +1 to i, and then judging whether i is greater than N; if yes, continuing to execute the step 18, otherwise, returning to the step 12;
step 18, calculating the output root mean square error of the SCN classifier in the LSTM-CNN neural network of the mu iteration to be
Figure BDA0002658014910000048
Step 19, judging mu is more than mu max And e μ ＜e 0 Whether the two are true at the same time; if yes, obtaining the LSTM-CNN neural network A of the mu iteration μ And the method is used for diagnosing the high-resistance grounding fault of the power distribution network, otherwise, after the mu +1 is assigned to the mu, the LSTM-CNN neural network A of the mu iteration is updated according to a gradient descent algorithm μ Then, step 11 is performed, wherein e 0 An error threshold is manually set.
Compared with the prior art, the invention has the beneficial effects that:
1. according to the method, a semi-supervised learning mechanism is adopted to carry out full labeling on massive fault phase voltage time sequence data, the problem that a deep learning network needs a large number of known samples to carry out training is solved, inconvenience caused by factors such as difficulty in obtaining high-resistance fault labeling data is reduced, and the detection rate of the method for the high-resistance ground fault of the power distribution network is improved.
2. The invention utilizes the LSTM network and the CNN network to respectively extract the characteristics of the fault phase voltage time sequence data on the time dimension and the frequency domain, fuses the time-frequency characteristics of the data by adopting an attention mechanism mode and dynamically weights the changed time characteristics, thereby enhancing the time-frequency characteristic expression capability of the method and ensuring the reliability of the high-resistance earth fault diagnosis of the power distribution network.
3. The random configuration network is adopted to replace a softmax layer in the traditional convolutional neural network, the learning capability and the generalization capability of a learner are enhanced, the global approximation of a high-resistance fault state of multi-scale characteristics of fault phase voltage time sequence data is realized, and the identification precision of the high-resistance ground fault of the power distribution network is improved.
Drawings
FIG. 1 is a flow chart of a high-resistance grounding fault diagnosis method based on a power distribution network according to the invention;
FIG. 2 is a timing waveform of a sample time series of a faulted phase voltage of the present invention;
FIG. 3 is a Fourier transform plot of samples of a time series of faulted phase voltages of the present invention;
fig. 4 is a schematic diagram of a high-resistance ground fault diagnosis model of the power distribution network.
Detailed Description
In this embodiment, referring to fig. 1, a method for diagnosing a high impedance ground fault of a power distribution network based on semi-supervised learning and attention mechanism is performed according to the following steps:
step 1: when a group of high-resistance ground faults of the power distribution network are obtained, a fault phase voltage time sequence of a secondary side end of a transformer is obtained and is used as a training sample, and the training sample is recorded as TS ═ TS 1 ,TS 2 ,···,TS i ,···,TS N }; where N represents the number of training samples, TS i Representing the ith fault phase voltage time sequence in the training sample, and comprising: TS (transport stream) i ＝{(t 1 ,x i,1 ),(t 2 ,x i,2 ),···,(t j ,x i,j ),···,(t q ,x i,q ) Q represents the length of the ith fault phase voltage time sequence in the training sample, t j Representing the jth time, x, in the training sample i,j Representing the voltage value corresponding to the ith fault phase voltage time sequence in the training sample at the jth time;
step 2, denoising the fault phase voltage time sequence set TS to obtain a preprocessed fault phase voltage time sequence set; classifying the preprocessed fault phase voltage time sequence set into a fault phase voltage time sequence sample set D containing unmarked fault phase voltage time sequence samples u And a time series sample set D containing marked fault phase voltage p ；
Step 3, defining a variable k, and initializing k to be 1; define the set of unlabeled samples after the kth update as
Figure BDA0002658014910000051
And is initialized to D u The set of marked samples after the kth update is
Figure BDA0002658014910000052
Initialisation to D p ；
Step 4, calculating the unmarked sample set after the kth updating
Figure BDA0002658014910000061
The jth fault phase voltage time sequence sample and the kth updated marked sample set
Figure BDA0002658014910000062
Similarity of (2) R j Thereby obtaining a similarity vector
Figure BDA0002658014910000063
j＝1，2，...m k ，m k As the set of unlabeled samples after the kth update
Figure BDA0002658014910000064
The number of samples of (a);
step 5, defining the unmarked sample set after the kth update
Figure BDA0002658014910000065
Single sample and labeled sample set
Figure BDA0002658014910000066
The similarity threshold of (a) is epsilon;
judging the similarity vector R k If all the similarity degrees are smaller than epsilon, executing a step 7 if yes, otherwise, executing a step 6; in a specific implementation, the similarity threshold is set to be ∈ ═ 0.8;
step 6, selecting the unmarked sample set after the kth updating
Figure BDA0002658014910000067
Neutralization of labeled sample sets
Figure BDA0002658014910000068
And manually marking the sample with the maximum similarity, and adding the manually marked sample into the marked sample set
Figure BDA0002658014910000069
Obtaining the unmarked sample set after the (k + 1) th update
Figure BDA00026580149100000610
Simultaneous unmarked sample set
Figure BDA00026580149100000611
Deleting the artificially marked samples to obtain an unmarked sample set updated for the (k + 1) th time
Figure BDA00026580149100000612
Assigning k +1 to k, and returning to the step 4;
step 7, utilizing the marked sample set after the kth updating
Figure BDA00026580149100000613
Training a nearest neighbor 1NN classifier to obtain a trained 1NN classifier which is used for collecting unlabeled samples
Figure BDA00026580149100000614
Classifying;
step 8, the unmarked sample set after the kth updating
Figure BDA00026580149100000615
Inputting the trained 1NN classifier, and taking the obtained prediction label as a real label of an unlabeled sample;
then, the unlabeled samples with the real labels are collected
Figure BDA00026580149100000616
And labeled sample sets
Figure BDA00026580149100000617
Merging to obtain a high-resistance fault training set T containing N samples;
in the embodiment, a intercepted part of a fault phase voltage time sequence data sample in a high-resistance fault training set T is shown in a table I; a timing waveform diagram of the faulted phase voltage time series samples is shown in figure 2.
Sample cutting part of watch phase voltage time sequence data
T(s) 1/60 2/60 3/60 4/60 5/60 6/60 7/60 8/60 9/60 10/60
V(v) -26.2228 -9.5195 5.5134 17.2057 25.5574 39.4768 56.7368 71.7698 85.6892 94.0408
T(s) 11/60 12/60 13/60 14/60 15/60 16/60 17/60 18/60 19/60 20/60
V(v) 102.9492 107.9602 109.6305 114.6415 120.766 128.5609 135.799 143.5939 150.8319 161.9675
T(s) 21/60 22/60 23/60 24/60 25/60 26/60 27/60 28/60 29/60 30/60
V(v) 168.6488 172.5462 173.6598 170.8759 165.8649 155.8429 148.0481 140.81 134.1287 125.2203
Step 9, constructing an LSTM-CNN neural network based on an attention mechanism;
the LSTM-CNN neural network comprises an LSTM network, a CNN network, a feature fusion layer based on an attention mechanism and an SCN classifier; the number of input nodes of the LSTM network is qx 1, and the output dimension of the hidden layer is mx 1; the output dimension of the CNN network hidden layer is mx 1;
step 10, defining the current iteration number of the network as mu, and initializing mu to 1; maximum number of iterations is mu max (ii) a Carrying out the random initialization for the mu time on the parameters of each layer in the network so as to obtain an LSTM-CNN neural network of the mu time iteration;
step 11, defining a variable i, and initializing to i-1;
step 12, carrying out Fourier transform on the ith fault phase voltage time sequence sample in the high-resistance fault training set T to obtain an image sample X of the ith fault phase voltage time sequence i (ii) a In an embodiment, a fourier transform diagram of a time series sample of a fault phase voltage in the high resistance fault training set T is shown in fig. 3.
Step 13, image sample X i Inputting the CNN network in the LSTM-CNN neural network of the mu iteration to obtain a feature vector C with dimension of m multiplied by 1 i,μ ；
Step 14, selecting the ith fault phase voltage time sequence sample from the high-resistance fault training set T, inputting the ith fault phase voltage time sequence sample into an LSTM network in an LSTM-CNN neural network of the mu iteration to obtain a feature vector with dimension q multiplied by m
Figure BDA0002658014910000071
Wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0002658014910000072
inputting the n-th time step representing the ith fault phase voltage time sequence sample into an LSTM network in the LSTM-CNN neural network of the mu iteration, and outputting a feature vector with dimension of m multiplied by 1;
step 15, feature vector C is aligned by using formula (1-1) -formula (1-3) i,μ And feature vector H i,μ Performing fusion to obtain m × 1 feature vector F with dimension i,μ ；
Figure BDA0002658014910000073
Figure BDA0002658014910000074
Figure BDA0002658014910000075
In the formula (1-1) -formula (1-3),
Figure BDA0002658014910000076
the weight matrix of the feature fusion layer of the LSTM-CNN neural network of the mu iteration is a dimension of m multiplied by m;
Figure BDA0002658014910000077
is a bias term for the feature fusion layer of the LSTM-CNN neural network for the μ iteration;
Figure BDA0002658014910000078
representing a feature vector with dimension of m multiplied by 1 output by an LSTM network in an LSTM-CNN neural network of the mu iteration when an ith fault phase voltage time sequence sample is input at the ith time step; alpha is alpha i,μ,n In the feature fusion layer of the LSTM-CNN neural network representing the μ th iteration
Figure BDA0002658014910000081
Corresponding fusion weight coefficients;
step 16, feature vector F i,μ Inputting the SCN classifier in the LSTM-CNN neural network of the mu iteration to obtain an output result t' i,μ Will output the result t' i,μ And the desired output t i Performing subtraction to obtain the output error e of the SCN classifier i,μ (ii) a In a specific embodiment, a schematic diagram of a high-resistance ground fault diagnosis model of the power distribution network constructed by the method in the steps 9 to 16 is shown in fig. 4;
step 17, assigning i +1 to i, and then judging whether i is greater than N; if yes, continuing to execute the step 18, otherwise, returning to the step 12;
step 18, calculating the output root mean square error of the SCN classifier in the LSTM-CNN neural network of the mu iteration to be
Figure BDA0002658014910000082
Step 19, judging mu is more than mu max And e μ ＜e 0 Whether the two are true at the same time; if yes, obtaining the LSTM-CNN neural network A of the mu iteration μ And the method is used for diagnosing the high-resistance grounding fault of the power distribution network, otherwise, after the mu +1 is assigned to the mu, the LSTM-CNN neural network A of the mu iteration is updated according to a gradient descent algorithm μ Then, step 11 is performed, wherein e 0 An error threshold is manually set. In a specific embodiment, the maximum number of iterations μ of the network is set manually max ＝500，e 0 ＝0.01。

Claims (1)

1. A power distribution network high-resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism is characterized by comprising the following steps:
step 1: when a group of high-resistance ground faults of the power distribution network are obtained, a fault phase voltage time sequence of a secondary side end of a transformer is obtained and is used as a training sample, and the training sample is recorded as TS ═ TS 1 ,TS 2 ,…,TS i ,…,TS N }; wherein N represents the number of training samples, TS i Representing the ith fault phase voltage time sequence in the training sample, and comprising: TS (transport stream) i ＝{(t 1 ,x i,1 ),(t 2 ,x i,2 ),…,(t j ,x i,j ),…,(t q ,x i,q ) Q represents the length of the ith fault phase voltage time sequence in the training sample, t j Represents the jth time, x, in the training sample i,j Representing the voltage value corresponding to the ith fault phase voltage time sequence in the training sample at the jth time;
step 2, denoising the fault phase voltage time sequence set TS to obtain the preprocessed fault phase voltage timeA sequence set; classifying the preprocessed fault phase voltage time sequence set into a fault phase voltage time sequence sample set D containing unmarked fault phase voltage time sequence samples u And a time series sample set D containing the marked fault phase voltage p ；
Step 3, defining a variable k, and initializing k to be 1; define the set of unlabeled samples after the kth update as
Figure FDA0002658014900000011
And is initialized to D u The set of marked samples after the kth update is
Figure FDA0002658014900000012
Initialisation to D p ；
Step 4, calculating the unmarked sample set after the kth updating
Figure FDA0002658014900000013
The j th fault phase voltage time sequence sample and the k time updated marked sample set
Figure FDA0002658014900000014
Similarity of (2) R j Thereby obtaining a similarity vector
Figure FDA0002658014900000015
j＝1,2,...m k ，m k For the unlabeled sample set after the kth update
Figure FDA0002658014900000016
The number of samples of (a);
step 5, defining the unmarked sample set after the kth update
Figure FDA0002658014900000017
Single sample and labeled sample set
Figure FDA0002658014900000018
The similarity threshold of (a) is epsilon;
judging the similarity vector R k If all the similarity degrees are less than epsilon, executing a step 7, otherwise, executing a step 6;
step 6, selecting the unmarked sample set after the kth updating
Figure FDA0002658014900000019
Neutralization of labeled sample sets
Figure FDA00026580149000000110
And manually marking the sample with the maximum similarity, and adding the manually marked sample into the marked sample set
Figure FDA00026580149000000111
And obtaining the unmarked sample set after the k +1 time of updating
Figure FDA00026580149000000112
Simultaneous unmarked sample set
Figure FDA00026580149000000113
Deleting the artificially marked samples to obtain an unmarked sample set updated for the (k + 1) th time
Figure FDA00026580149000000114
Assigning k +1 to k, and returning to the step 4;
step 7, utilizing the marked sample set after the kth updating
Figure FDA00026580149000000115
Training a nearest neighbor 1NN classifier to obtain a trained 1NN classifier which is used for collecting unlabeled samples
Figure FDA0002658014900000021
Classifying;
step 8, collecting the unmarked samples after the kth updating
Figure FDA0002658014900000022
Inputting the trained 1NN classifier, and taking the obtained prediction label as a real label of an unlabeled sample;
then collecting the unlabeled sample with the real label
Figure FDA0002658014900000023
And labeled sample sets
Figure FDA0002658014900000024
Merging to obtain a high-resistance fault training set T containing N samples;
step 9, constructing an LSTM-CNN neural network based on an attention mechanism;
the LSTM-CNN neural network comprises an LSTM network, a CNN network, a feature fusion layer based on an attention mechanism and an SCN classifier; the number of input nodes of the LSTM network is q multiplied by 1, and the output dimension of the hidden layer is m multiplied by 1; the output dimension of the CNN network hidden layer is mx 1;
step 10, defining the current iteration number of the network as mu, and initializing mu to 1; maximum number of iterations mu max (ii) a Carrying out the random initialization for the mu time on the parameters of each layer in the network so as to obtain an LSTM-CNN neural network of the mu time iteration;
step 11, defining a variable i, and initializing to i-1;
step 12, carrying out Fourier transform on the ith fault phase voltage time sequence sample in the high-resistance fault training set T to obtain an image sample X of the ith fault phase voltage time sequence i ；
Step 13, the image sample X i Inputting the CNN network in the LSTM-CNN neural network of the mu iteration to obtain a feature vector C with dimension of m multiplied by 1 i,μ ；
Step 14, selecting the ith fault phase voltage time sequence sample from the high-resistance fault training set T and inputting the ith fault phase voltage time sequence sample into the LSTM-CNN neural network of the mu iterationObtaining feature vector with dimension q × m by LSTM network in network
Figure FDA0002658014900000025
Wherein the content of the first and second substances,
Figure FDA0002658014900000026
inputting the n-th time step representing the ith fault phase voltage time sequence sample into an LSTM network in an LSTM-CNN neural network of the mu-th iteration, and outputting a feature vector with the dimension of m multiplied by 1;
step 15, utilizing the formula (1-1) -formula (1-3) to align the feature vector C i,μ And the feature vector H i,μ Performing fusion to obtain m × 1 feature vector F with dimension i,μ ；
Figure FDA0002658014900000027
Figure FDA0002658014900000028
Figure FDA0002658014900000031
In the formula (1-1) -formula (1-3),
Figure FDA0002658014900000032
the weight matrix of the feature fusion layer of the LSTM-CNN neural network of the mu iteration is a dimension of m multiplied by m;
Figure FDA0002658014900000033
is a bias term for the feature fusion layer of the LSTM-CNN neural network for the μ iteration;
Figure FDA0002658014900000034
representing the ith faulted phase voltage time seriesInputting the characteristic vector with dimension of mx 1 output by the LSTM network in the LSTM-CNN neural network of the mu iteration at the first time step of the sample; alpha is alpha i,μ,n In the feature fusion layer of the LSTM-CNN neural network representing the μ th iteration
Figure FDA0002658014900000035
Corresponding fusion weight coefficients;
step 16, the feature vector F is processed i,μ Inputting the SCN classifier in the LSTM-CNN neural network of the iteration of the mu time to obtain an output result t' i,μ Will output the result t' i,μ And the desired output t i Performing subtraction to obtain the output error e of the SCN classifier i,μ ；
Step 17, assigning i +1 to i, and then judging whether i is greater than N; if yes, continuing to execute the step 18, otherwise, returning to the step 12;
step 18, calculating the output root mean square error of the SCN classifier in the LSTM-CNN neural network of the mu iteration to be
Figure FDA0002658014900000036
Step 19, judging mu is more than mu max And e μ ＜e 0 Whether the two are true at the same time; if yes, obtaining the LSTM-CNN neural network A of the mu iteration μ And the method is used for diagnosing the high-resistance grounding fault of the power distribution network, otherwise, after the mu +1 is assigned to the mu, the LSTM-CNN neural network A of the mu iteration is updated according to a gradient descent algorithm μ Thereafter, step 11 is performed, wherein e 0 An error threshold is manually set.
CN202010894524.5A 2020-08-31 2020-08-31 Power distribution network high-resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism Active CN112016473B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010894524.5A CN112016473B (en) 2020-08-31 2020-08-31 Power distribution network high-resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010894524.5A CN112016473B (en) 2020-08-31 2020-08-31 Power distribution network high-resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism

Publications (2)

Publication Number Publication Date
CN112016473A CN112016473A (en) 2020-12-01
CN112016473B true CN112016473B (en) 2022-09-13

Family

ID=73503249

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010894524.5A Active CN112016473B (en) 2020-08-31 2020-08-31 Power distribution network high-resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism

Country Status (1)

Country Link
CN (1) CN112016473B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112929380B (en) * 2021-02-22 2022-04-15 中国科学院信息工程研究所 Trojan horse communication detection method and system combining meta-learning and spatiotemporal feature fusion
CN113080907B (en) * 2021-04-14 2022-10-25 贵州省人民医院 Pulse wave signal processing method and device
CN113466630B (en) * 2021-07-19 2023-10-24 广东电网有限责任公司广州供电局 Power distribution network fault cause classification method based on RSSPN model
CN113447766A (en) * 2021-08-17 2021-09-28 广东电网有限责任公司东莞供电局 Method, device, equipment and storage medium for detecting high-resistance ground fault
CN114089218A (en) * 2021-10-19 2022-02-25 广东电网有限责任公司东莞供电局 Power distribution network high-resistance grounding fault identification method, device, terminal and medium
CN114169249A (en) * 2021-12-16 2022-03-11 福州大学 Artificial intelligence identification method for high-resistance grounding fault of power distribution network
CN114781458B (en) * 2022-05-05 2024-02-20 合肥工业大学 Power distribution network initial fault identification method based on adaptive wavelet and attention mechanism time-frequency memory neural network

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104655423A (en) * 2013-11-19 2015-05-27 北京交通大学 Rolling bearing fault diagnosis method based on time-frequency domain multidimensional vibration feature fusion
CN109726524A (en) * 2019-03-01 2019-05-07 哈尔滨理工大学 A kind of rolling bearing remaining life prediction technique based on CNN and LSTM
CN110307982A (en) * 2019-06-19 2019-10-08 电子科技大学 Bearing fault classification method based on CNN and Adaboost
CN110689069A (en) * 2019-09-25 2020-01-14 贵州电网有限责任公司 Transformer fault type diagnosis method based on semi-supervised BP network
CN111523509A (en) * 2020-05-08 2020-08-11 江苏迪赛司自动化工程有限公司 Equipment fault diagnosis and health monitoring method integrating physical and deep expression characteristics

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108416103A (en) * 2018-02-05 2018-08-17 武汉大学 A kind of method for diagnosing faults of electric automobile of series hybrid powder AC/DC convertor
US20200133254A1 (en) * 2018-05-07 2020-04-30 Strong Force Iot Portfolio 2016, Llc Methods and systems for data collection, learning, and streaming of machine signals for part identification and operating characteristics determination using the industrial internet of things

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104655423A (en) * 2013-11-19 2015-05-27 北京交通大学 Rolling bearing fault diagnosis method based on time-frequency domain multidimensional vibration feature fusion
CN109726524A (en) * 2019-03-01 2019-05-07 哈尔滨理工大学 A kind of rolling bearing remaining life prediction technique based on CNN and LSTM
CN110307982A (en) * 2019-06-19 2019-10-08 电子科技大学 Bearing fault classification method based on CNN and Adaboost
CN110689069A (en) * 2019-09-25 2020-01-14 贵州电网有限责任公司 Transformer fault type diagnosis method based on semi-supervised BP network
CN111523509A (en) * 2020-05-08 2020-08-11 江苏迪赛司自动化工程有限公司 Equipment fault diagnosis and health monitoring method integrating physical and deep expression characteristics

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Learning attention representation with a multi-scale CNN for gear fault diagnosis under different working conditions;yong yao et al.;《sensors》;20200224;第1-21页 *
基于时频融合和注意力机制的深度学习行星齿轮箱故障诊断方法;孔子迁等;《仪器仪表学报》;20190630;第40卷(第6期);第221-227页 *

Also Published As

Publication number Publication date
CN112016473A (en) 2020-12-01

Similar Documents

Publication Publication Date Title
CN112016473B (en) Power distribution network high-resistance grounding fault diagnosis method based on semi-supervised learning and attention mechanism
CN110082640B (en) Distribution network single-phase earth fault identification method based on long-time memory network
CN108573225B (en) Partial discharge signal pattern recognition method and system
Zhang et al. WPD and DE/BBO-RBFNN for solution of rolling bearing fault diagnosis
CN103942749B (en) A kind of based on revising cluster hypothesis and the EO-1 hyperion terrain classification method of semi-supervised very fast learning machine
CN104966161B (en) A kind of power quality recorder data calculation and analysis methods based on gauss hybrid models
CN109884419A (en) A kind of wisdom grid power quality on-line fault diagnosis method
CN109918649B (en) Suicide risk identification method based on microblog text
CN109583635A (en) A kind of short-term load forecasting modeling method towards operational reliability
CN112255500A (en) Power distribution network weak characteristic fault identification method based on transfer learning
CN112215279B (en) Power grid fault diagnosis method based on immune RBF neural network
CN114781458B (en) Power distribution network initial fault identification method based on adaptive wavelet and attention mechanism time-frequency memory neural network
CN114818579A (en) Analog circuit fault diagnosis method based on one-dimensional convolution long-short term memory network
Thomas et al. Neural architecture search algorithm to optimize deep transformer model for fault detection in electrical power distribution systems
CN114328663A (en) High-dimensional theater data dimension reduction visualization processing method based on data mining
Liu et al. CNN-based fault phase identification method of double circuit transmission lines
CN113780346B (en) Priori constraint classifier adjustment method, system and readable storage medium
CN115128402A (en) Power distribution network fault type identification method and system based on data driving
Zhang et al. Deep convolutional neural network for fog detection
CN112163613A (en) Rapid identification method for power quality disturbance
Dunbar et al. Ensemble-based experimental design for targeted high-resolution simulations to inform climate models
Zhang et al. Fault detection, classification, and location for active distribution network based on neural network and phase angle analysis
Zhengfeng Accurate recognition method of continuous sports action based on deep learning algorithm
CN115656714A (en) Power distribution network single-phase earth fault identification method based on deep neural network
CN109740636B (en) Stratum oil-water layer identification method based on AdaBoost combined with GRU

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant