CN113971416A - Cable early fault identification method - Google Patents

Cable early fault identification method Download PDF

Info

Publication number
CN113971416A
CN113971416A CN202111111740.9A CN202111111740A CN113971416A CN 113971416 A CN113971416 A CN 113971416A CN 202111111740 A CN202111111740 A CN 202111111740A CN 113971416 A CN113971416 A CN 113971416A
Authority
CN
China
Prior art keywords
original signal
training
dbn
model
layer
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.)
Pending
Application number
CN202111111740.9A
Other languages
Chinese (zh)
Inventor
黄华峰
王秀茹
迟鹏
刘刚
梁睿
赖勇
赵航宇
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.)
State Grid Jiangsu Electric Power Co ltd Suqian Power Supply Branch
China University of Mining and Technology CUMT
Original Assignee
State Grid Jiangsu Electric Power Co ltd Suqian Power Supply Branch
China University of Mining and Technology CUMT
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 State Grid Jiangsu Electric Power Co ltd Suqian Power Supply Branch, China University of Mining and Technology CUMT filed Critical State Grid Jiangsu Electric Power Co ltd Suqian Power Supply Branch
Priority to CN202111111740.9A priority Critical patent/CN113971416A/en
Publication of CN113971416A publication Critical patent/CN113971416A/en
Pending legal-status Critical Current

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
    • G06F2218/12—Classification; Matching
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/50—Testing of electric apparatus, lines, cables or components for short-circuits, continuity, leakage current or incorrect line connections
    • G01R31/58—Testing of lines, cables or conductors
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • 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
    • 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
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Signal Processing (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Cable Transmission Systems, Equalization Of Radio And Reduction Of Echo (AREA)

Abstract

The invention provides a cable early fault identification method, which comprises the following steps: step S1: selecting cable fault phase current as an original signal, and preprocessing the original signal; the preprocessing comprises the steps of carrying out signal decomposition and noise reduction on the original signal by utilizing the VMD, and reducing Gaussian white noise in the original signal; step S2: selecting time domain features of the preprocessed signals and the first five scale values of refined multi-scale dispersion entropy to construct a feature vector F, normalizing the extracted feature vector F to construct an identification model input feature vector, and enabling the identification model input feature vector to be distributed in [ -1,1] and serve as a feature vector input to the depth confidence network; step S3: and initializing parameters of the DBN model, training the DBN model through pre-training, and optimizing the deep confidence network structure by utilizing a PSO algorithm. The invention improves the robustness and stability of the identification model and improves the efficiency and accuracy of fault identification.

Description

Cable early fault identification method
Technical Field
The invention relates to a cable early fault identification method, and belongs to the field of power system fault diagnosis.
Background
The cable is an indispensable component part in the distribution network, however because the cable work scene is various, and operational environment is abominable, and its running state receives multiple factor influence, causes its operation monitoring data to present multisource, mixed characteristic, and this brings very big challenge for accurate fault identification of cable.
The early fault of the cable is generated at the initial stage of fault formation, the fault characteristics of the cable are difficult to distinguish from transient disturbance, the early fault is identified in time, and a countermeasure is taken for the cable with damaged insulation, so that the situation that the early fault of the cable is further developed into a permanent fault to influence the safe and stable operation of a power system can be avoided; therefore, it is an urgent research task to realize accurate identification of early cable faults.
At present, cable early fault identification is mainly divided into feature extraction and classification identification, and methods used for feature extraction include Wavelet Transform (WT) and Empirical Mode Decomposition (EMD); WT has the problems that a mother wavelet function and the number of decomposition layers are difficult to select and are easily influenced by noise, and EMD has the problems of mode aliasing, end point effect and the like in the decomposition process; the classification identification comprises a Support Vector Machine (SVM), a Deep Belief Network (DBN) and the like, the SVM has difficulty in solving multi-classification problems and training large-scale samples, the learning process of the DBN model is slow, and inappropriate parameter selection can cause the learning to converge to local optimum; therefore, it is very important to provide a method combining feature extraction and classification identification to realize early cable fault identification.
Disclosure of Invention
The invention provides a cable early fault identification method, and aims to solve the problem of judging whether a cable has an early fault or not from the aspect of fault phase current characteristics.
The technical solution of the invention is as follows: a method of cable early fault identification, the method comprising:
step S1: selecting cable fault phase current as an original signal, and preprocessing the original signal; the preprocessing comprises the steps of carrying out signal decomposition and noise reduction on the original signal by utilizing the VMD, and reducing Gaussian white noise in the original signal;
step S2: selecting time domain features of the preprocessed signals and the first five scale values of refined multi-scale dispersion entropy to construct a feature vector F, normalizing the extracted feature vector F to construct an identification model input feature vector, and enabling the identification model input feature vector to be distributed in [ -1,1] and serve as a feature vector input to the depth confidence network;
step S3: initializing DBN model training parameters, training the DBN model through pre-training, and optimizing a deep belief network structure by utilizing a PSO algorithm.
Further, the method for identifying the early fault of the cable further comprises the following steps:
step S4: determining other relevant parameters of the DBN model, and finishing model training to obtain a PSO-DBN model; the other related parameters comprise a loss function, RBM learning rate, RBM iteration times and network layer number.
Further, the signal decomposition and noise reduction of the original signal by using the VMD specifically includes decomposing the original signal and reducing white gaussian noise in the signal.
Further, the signal decomposition of the original signal by using the VMD includes a reasonable setting of the number K of VMD decomposition layers, specifically: decomposing a noisy signal in an original signal by adopting a VMD (virtual matrix display), wherein the number K of decomposition layers is gradually increased from 2, and the central frequency of each modal component is recorded after each decomposition; according to the idea of the center frequency method, when the center frequency is closest to the center frequency when K is 2, the signal is considered to be over-decomposed, and the K value when the center frequency is closest to the center frequency when K is 2 is used as the decomposition layer number of the VMD algorithm.
Further, the constructing the identification model input feature vector specifically includes: normalizing a feature vector F formed by time domain features of the decomposed and denoised original signal and the first five scale values of refined multi-scale dispersion entropy (RCMDE), so that the feature vector is distributed in [ -1,1] and is used as the input of a deep confidence network model; the time domain characteristics of the decomposed and noise-reduced original signal comprise a maximum value (Max), a minimum value (Min), a peak-to-peak value (Ptp), an average value (Mean), a variance (Var), a kurtosis (Kv), a skewness (Sk) and a peak factor (Par) of the decomposed and noise-reduced original signal.
Further, the method for calculating each relevant parameter of the time domain feature is as follows:
peak-to-peak (Ptp): x is the number ofptp＝xmax-xmin；
In the formula: x is the number ofptpIs the peak-to-peak value, xmaxIs the maximum value, x, of the original signal after decomposition and noise reductionminThe minimum value of the original signal after decomposition and noise reduction processing is obtained;
mean (Mean):
Figure BDA0003274137620000031
in the formula:
Figure BDA0003274137620000032
is an arithmetic mean, x is the original signal, and N is the signal sequence length;
variance (Var):
Figure BDA0003274137620000033
in the formula: x is the number ofvarIs the variance;
kurtosis (Kv):
Figure BDA0003274137620000041
in the formula: x is the number ofkvIs the kurtosis;
skewness (Sk):
Figure BDA0003274137620000042
in the formula: is xskSkewness, xstdIs the standard deviation of the measured data to be measured,
Figure BDA0003274137620000043
crest factor (Par):
Figure BDA0003274137620000044
in the formula: x is the number ofparIs the crest factor, xrmsIs the root mean square (rms) value,
Figure BDA0003274137620000045
further, the calculation method of the refined multi-scale dispersion entropy (RCMDE) is as follows:
Figure BDA0003274137620000046
in the formula: eRCMD(x, m, c, d, τ) is the embedding dimension, m is the embedding dimension, c is the number class, d is the time delay, τ is the scale,
Figure BDA0003274137620000047
is the average value of the probability of the scattering pattern corresponding to the coarse grained sequence,
Figure BDA0003274137620000048
is the probability of the spreading pattern corresponding to the kth coarse-grained sequence at the scale τ.
Further, the optimizing the deep belief network structure by using a Particle Swarm Optimization (PSO) specifically includes: determining the number of hidden layer layers and determining the number of hidden layer neurons.
Further, the determining of the number of hidden layers specifically includes determining the number of hidden layers by using a control variable method, setting the number of neurons in each layer to be the same, adjusting the number of hidden layers, and determining a proper number of hidden layers through an error after the DBN model training.
Further, the determining the number of hidden layer neurons specifically includes: and selecting the number of neurons of the optimal hidden layer through a PSO algorithm, wherein the fitness function of the PSO algorithm is set as the identification accuracy rate after the DBN model is trained.
Further, training of the DBN model: unauthorized pre-training and fine-tuning layer by layer; in the pre-training stage, a training sample is used as the input of a DBN model, a Restricted Boltzmann Machine (RBM) is trained by a layer-by-layer greedy method, each RBM is trained by maximizing the probability of input data, and parameters are updated by using a Contrast Divergence (CD) algorithm;
after pre-training, adding a classification layer to the last hidden layer, further fine-tuning the DBN by minimizing the error between an estimated output value and a label, fine-tuning the DBN model reversely by a BP algorithm in a supervision mode, updating parameters of the whole network, and inputting the extracted feature vectors into the PSO-DBN model for classification and identification after training.
The invention provides a cable early fault identification method based on Variational Modal Decomposition (VMD) and PSO algorithm optimized Deep Belief network (PSO-DBN), belonging to the technical field of power system fault identification; firstly, selecting cable fault phase current as basic data of cable early fault identification, and eliminating Gaussian white noise of an original signal by utilizing a VMD (virtual machine format); then, extracting time domain characteristics and fine Composite Multiscale Dispersion Entropy (RCMDE) characteristics, and performing normalization processing on the characteristics to obtain a characteristic vector input by the deep belief network; finally, optimizing the network structure by utilizing a Particle Swarm Optimization (PSO), selecting the number of neurons of the optimal hidden layer, and realizing efficient and accurate identification of early faults of the cable; by the VMD algorithm, data preprocessing is carried out on the original signal, time domain characteristics and RCMDE characteristics are extracted, and the DBN model structure is optimized by the PSO, so that the robustness and stability of the identification model can be further improved, and the efficiency and accuracy of fault identification are improved.
The invention has the beneficial effects that:
the VMD can self-adaptively solve the constructed constraint variational equation to realize effective decomposition of signal components, and has the optimal decomposition effect; the PSO algorithm can optimize the DBN model structure, reduce the training time and has good optimization effect;
2. the method is suitable for identifying the early faults of the cable, and compared with ELM, GRNN and PNN, the method has higher accuracy when the DBN model is used as an identification classification network; compared with an unoptimized DBN model, the PSO-DBN (Particle Swarm Optimization-Deep Belief Networks) reduces the model training time, improves the identification efficiency and has higher identification accuracy.
Drawings
FIG. 1 is a flow chart of cable early failure identification based on VMD and PSO-DBN.
FIG. 2 is an IEEE-13 node test feeder diagram.
FIG. 3 is a graph of SNR for different denoising algorithms.
FIG. 4 is a diagram of mean square error for different denoising algorithms.
FIG. 5 is a similar coefficient diagram for different denoising algorithms.
FIG. 6 is a graph of RCMDE values for each state.
FIG. 7 is a test set accuracy graph for different hidden layers.
Figure 8 is a fitness curve.
FIG. 9 is a graph of mean square error and failure diagnosis accuracy as a function of iteration number.
FIG. 10 is a graph of performance of different identification methods under noise interference.
Detailed Description
A method of cable early fault identification, the method comprising:
step S1: selecting cable fault phase current as an original signal, and preprocessing the original signal; the preprocessing comprises the steps of carrying out signal decomposition and noise reduction on the original signal by utilizing the VMD, and reducing Gaussian white noise in the original signal;
step S2: selecting time domain features of the preprocessed signals and the first five scale values of refined multi-scale dispersion entropy (RCMDE) to construct a feature vector F, normalizing the extracted feature vector F to construct an identification model input feature vector, and distributing the identification model input feature vector in [ -1,1] to serve as a feature vector input to the depth confidence network;
step S3: initializing DBN model training parameters, training a DBN model through pre-training, and optimizing a deep belief network structure by utilizing a Particle Swarm Optimization (PSO); the deep belief network structure is optimized by utilizing a Particle Swarm Optimization (PSO) algorithm.
Step S4: determining other relevant parameters of the DBN model, and finishing model training to obtain a PSO-DBN model; the PSO-DBN model after model training can be used for cable early fault identification; the other related parameters comprise a loss function, RBM learning rate, RBM iteration times and network layer number.
The invention discloses a cable early fault identification method based on a VMD (virtual vehicle model) and a PSO-DBN (power system on-board-bus) model, which solves the problem of judging whether a cable has an early fault from the aspect of fault phase current characteristics, considers the noise influence possibly occurring in an actual field, and brings the condition that learning converges to a local optimal solution due to the fact that the learning process of the DBN model is slow and inappropriate in parameter selection, and provides the cable early fault identification method based on the VMD and the PSO-DBN model.
The signal decomposition and noise reduction of the original signal by using the VMD specifically includes decomposing the original signal and reducing white Gaussian noise in the signal.
The signal decomposition of the original signal by utilizing the VMD comprises the construction and the solution of a variational problem, namely the reasonable setting of the VMD decomposition layer number K, which specifically comprises the following steps: decomposing a noisy signal in an original signal by adopting a VMD (virtual matrix display), wherein the number K of decomposition layers is gradually increased from 2, and the central frequency of each modal component is recorded after each decomposition; according to the central frequency method, when the central frequency is closest to the central frequency when K is 2, the signal is considered to be over-decomposed, and therefore the K value when the central frequency is closest is used as the decomposition layer number of the VMD algorithm.
The method for constructing the input feature vector of the identification model specifically comprises the following steps: normalizing a feature vector F formed by time domain features of the decomposed and noise-reduced original signal and first five scale values of refined multi-scale dispersion entropy (RCMDE), so that the feature vector is distributed in [ -1,1] and is used as the input of a DBN model; the time domain characteristics of the decomposed and noise-reduced original signal comprise a maximum value (Max), a minimum value (Min), a peak-to-peak value (Ptp), an average value (Mean), a variance (Var), a kurtosis (Kv), a skewness (Sk) and a peak factor (Par) of the decomposed and noise-reduced original signal.
The method for calculating each relevant parameter of the time domain feature comprises the following steps:
peak-to-peak (Ptp): x is the number ofptp＝xmax-xmin；
In the formula: x is the number ofptpIs the peak-to-peak value, xmaxIs the maximum value, x, of the original signal after decomposition and noise reductionminThe minimum value of the original signal after decomposition and noise reduction processing is obtained;
mean (Mean):
Figure BDA0003274137620000081
in the formula:
Figure BDA0003274137620000082
is an arithmetic mean, x is the original signal, and N is the signal sequence length;
variance (Var):
Figure BDA0003274137620000091
in the formula: x is the number ofvarIs the variance;
kurtosis (Kv):
Figure BDA0003274137620000092
in the formula: x is the number ofkvIs the kurtosis;
skewness (Sk):
Figure BDA0003274137620000093
in the formula: is xskSkewness, xstdIs the standard deviation of the measured data to be measured,
Figure BDA0003274137620000094
crest factor (Par):
Figure BDA0003274137620000095
in the formula: x is the number ofparIs the crest factor, xrmsIs the root mean square (rms) value,
Figure BDA0003274137620000096
the calculation method of the refined multiscale dispersion entropy (RCMDE) is as follows:
Figure BDA0003274137620000097
in the formula: eRCMD(x, m, c, d, τ) is the embedding dimension, m is the embedding dimension, c is the number class, d is the time delay, τ is the scale,
Figure BDA0003274137620000098
is the average value of the probability of the scattering pattern corresponding to the coarse grained sequence,
Figure BDA0003274137620000099
is the probability of the spreading pattern corresponding to the kth coarse-grained sequence at the scale τ.
The optimizing the deep confidence network structure by using a Particle Swarm Optimization (PSO) algorithm specifically comprises the following steps: determining the number of hidden layer layers and determining the number of neurons of the hidden layers; the determining of the number of the hidden layers specifically comprises the steps of determining the number of the hidden layers by adopting a control variable method, namely setting the number of neurons in each layer to be the same, adjusting the number of the hidden layers, and determining the proper number of the hidden layers through errors after the DBN model is trained; the determining of the number of hidden layer neurons specifically comprises optimizing the number of hidden layer neurons of a DBN (deep Belief networks) model by a PSO (particle Swarm optimization) algorithm; and optimizing the deep belief network structure to determine the optimal number of hidden layers and the optimal number of neurons of the DBN model.
The determining the number of hidden layer neurons specifically comprises: and selecting the number of neurons of the optimal hidden layer through a PSO algorithm, wherein the fitness function of the PSO algorithm is set as the identification accuracy rate after the DBN model is trained.
Training of the DBN model: unauthorized pre-training and fine-tuning layer by layer; in the pre-training stage, a training sample is used as the input of a DBN model, a Restricted Boltzmann Machine (RBM) is trained by a layer-by-layer greedy method, each RBM is trained by maximizing the probability of input data, and parameters are updated by using a Contrast Divergence (CD) algorithm.
After pre-training, adding a classification layer to the last hidden layer, further fine-tuning the DBN by minimizing the error between an estimated output value and a label, fine-tuning the DBN model reversely by a BP algorithm in a supervision mode, updating parameters of the whole network, and inputting the extracted characteristic quantity into the PSO-DBN model for classification and identification after training.
The invention provides a cable early fault identification method based on Variational Modal Decomposition (VMD) and PSO algorithm optimized Deep Belief network (PSO-DBN), belonging to the technical field of power system fault identification; firstly, selecting cable fault phase current as basic data of cable early fault identification, and eliminating Gaussian white noise of an original signal by utilizing a VMD (virtual machine format); then, extracting time domain characteristics and fine Composite Multiscale Dispersion Entropy (RCMDE) characteristics, and performing normalization processing on the characteristics to obtain a characteristic vector input by the deep belief network; finally, optimizing the network structure by utilizing a Particle Swarm Optimization (PSO), selecting the number of neurons of the optimal hidden layer, and realizing efficient and accurate identification of early faults of the cable; by the VMD algorithm, data preprocessing is carried out on the original signal, time domain characteristics and RCMDE characteristics are extracted, and the DBN model structure is optimized by the PSO, so that the robustness and stability of the identification model can be further improved, and the efficiency and accuracy of fault identification are improved.
Example 1
An exemplary embodiment of the present invention will be described in detail below with reference to fig. 1.
A method of cable early fault identification, the method comprising: 1. preprocessing data; 2. optimizing a DBN network; 3. pre-training and fine-tuning.
The data preprocessing comprises the following steps: 1-1, selecting a cable fault phase current as an original signal; 1-2, decomposing VMD signals and reducing noise; 1-3, extracting time domain features and RCMDE values; 1-4, setting a training set and a testing set.
The DBN network optimization comprises the following steps: 2-1, determining a DBN topological structure; 2-2, initializing DBN parameters; 2-3, initializing particle swarm parameters; 2-4, selecting a proper fitness value; 2-5, searching individual extremum and group extremum; 2-6, updating the position and the speed of the particles; 2-7, calculating a particle fitness value; 2-8, updating individual extremum and group extremum; and 2-9, judging whether the termination condition is met.
The pre-training and fine-tuning comprises: 3-1, initializing DBN model training parameters; 3-2, training RBMs layer by layer to realize pre-training; 3-3, utilizing a BP algorithm to finely adjust training parameters; 3-4, completing DBN model training; and 3-5, outputting the identification result.
The 1-2 VMD signal Decomposition and noise reduction comprises the step of performing signal Decomposition and noise reduction on an original signal by utilizing VMD (variable Mode Decomposition), and reducing Gaussian white noise in the signal.
And 1-3, extracting time domain features and RCMDE values, wherein the time domain features and RCMDE values after decomposition and noise reduction are selected, the first five scale values of refined multi-scale dispersion entropy (RCMDE) are selected to construct a feature vector F, and the extracted feature vector F is normalized and distributed in [ -1,1] to serve as the input of the deep confidence network.
1-4, setting a training set and a test set, selecting a certain proportion of data samples in various state type samples to form the training set, and forming the rest samples into the test set; preferably, 80% of the data samples in the various state type samples form a training set, and 20% of the data samples form a testing set; the various state type samples include time domain features of the decomposed noise-reduced signal of the original signal and the first five scale values of the refined multi-scale dispersion entropy (RCMDE).
And 2-1, determining a DBN topological structure, and determining the input and output dimensions of the DBN according to the input feature vector and the output fault type.
And 2-2, initializing DBN parameters, randomly assigning initial weights of the DBN models, and determining the number of hidden layers by a control variable method.
And 2-3, initializing particle swarm parameters, and randomly initializing the position and the velocity vector of each particle.
And 2-4, selecting a proper fitness value, and taking the identification accuracy after the DBN training as the fitness value.
2-5, searching individual extremum and group extremum, wherein the individual extremum is pbestiThe optimal position that the ith particle has experienced is expressed as
Figure BDA0003274137620000131
The update formula for pbest is:
Figure BDA0003274137620000132
wherein f (x) is a fitness function,
Figure BDA0003274137620000133
n is the total number of particles, and m is the dimension of the particles; p is a radical ofijA reconstructed value of a jth dimension in an ith particle; t is tijThe actual value of the j dimension data in the ith particle is taken as the actual value of the j dimension data in the ith particle;
group extremum gbesttFor the optimal positions that i particles have experienced, i.e. the global optimal positions, the expression is:
Figure BDA0003274137620000134
and 2-6, updating the position and the speed of the particle, wherein the position and the speed of the particle are updated according to the following formula:
Figure BDA0003274137620000135
Figure BDA0003274137620000136
in the formula: n is n-dimensional search space, current position X of ith particleiIs composed of
Figure BDA0003274137620000137
The flight velocity of the ith particle is expressed as
Figure BDA0003274137620000138
Recording the fitness value of the current position every time the particle i passes through one position,
Figure BDA0003274137620000139
the flight speed of the nth dimension of the ith particle when the ith particle is iterated to the t generation;
Figure BDA00032741376200001310
the position of the ith particle in the nth dimension when iterating to the t generation, and w represents weight for adjusting the search range; c. C1,c2As an acceleration factor, r1,r2Is the interval [0,1]The random number in (c).
2-9, judging whether the optimization termination condition is met, and if the optimization termination condition is met, outputting a global optimal value, namely the number of neurons of each hidden layer; if the termination condition is not met, the particle position and velocity are updated, a new particle swarm is generated, and the process goes to 2-6.
And 3-1, initializing DBN model training parameters, and training the optimal individual extremum in the DBN model, namely substituting the optimal hidden layer neuron number into the DBN model for training.
3-2, training RBMs layer by layer to realize pre-training, training a Restricted Boltzmann Machine (RBM) by adopting a layer-by-layer greedy method, training each RBM by maximizing the probability of input data thereof, and updating parameters by using a Contrast Divergence (CD) algorithm.
And 3-3, fine-tuning the training parameters by using a BP algorithm, adding a classification layer to the last hidden layer after pre-training, further fine-tuning the DBN model by minimizing the error between an estimated output value and a label, fine-tuning the DBN model reversely by using the BP algorithm in a supervision mode, and updating the parameters of the whole network.
And 3-4, finishing the training of the DBN model and determining the final parameters of the DBN model.
And 3-5, outputting an identification result, inputting a characteristic vector obtained by preprocessing and normalizing a group of original signal data as a DBN model, and outputting a fault type.
Example 2
The technical scheme of the invention is further explained in detail by combining the attached drawings:
the flow of the cable early failure identification method based on the VMD and the PSO-DBN is shown in FIG. 1, and the steps are as follows:
step S1: selecting fault phase current as an original signal, and performing signal decomposition and noise reduction on all samples by using the VMD to reduce white Gaussian noise in the signals; the effectiveness of the method is verified by utilizing a 10kV power distribution network hybrid model of a cable and an overhead line of an IEEE13 node constructed by PSCAD (power system computer aided design), as shown in FIG. 2; six conditions of multi-cycle arc fault, half-cycle arc fault, fixed impedance fault, capacitor switching, motor starting and load switching are simulated in a simulation mode, different fault conditions such as fault positions, fault resistances and fault angles are set, fault phase current is collected, noise influence in an actual field is considered, and Gaussian white noise with different signal-to-noise ratios is added to serve as an original signal.
Decomposing and denoising an original signal by utilizing the VMD to realize data preprocessing, wherein the signal decomposition comprises the construction and the solution of a variational problem, namely, the reasonable setting of the VMD decomposition layer number K: decomposing a noisy signal by adopting a VMD (virtual machine tool) with the number K of decomposition layers gradually increased from 2, and recording the central frequency of each modal component after each decomposition; according to the thought of a central frequency method, namely when the central frequencies are close, the signals are considered to be over-decomposed, so that the K values when the central frequencies are close are selected as the decomposition layer number of the VMD algorithm; when K is 5, the central frequency 3269Hz is closer to the central frequency 3492Hz when K is 2, and according to the idea of the central frequency method, when the central frequencies are close, the signals are considered to be over-decomposed; therefore, the K value when the center frequencies are close is taken as the decomposition layer number of the VMD algorithm.
TABLE 1 center frequencies for different K values
Figure BDA0003274137620000151
In order to verify the applicability of the VMD algorithm, gaussian white noise with different Signal-to-noise ratios is added to the original simulation Signal, under the conditions that the Signal-to-noise ratios are-10, -8, -6, -4, -2, 0, 2, 4, 6, 8, 10dB, respectively, the EMD, the Wavelet transform-Hard threshold (WT-HT), the Wavelet Soft threshold (WT-ST), and the VMD method are adopted to perform denoising, the denoising result is shown in fig. 3, the abscissa is noise with different Signal-to-noise ratios added, and the ordinate is the denoised Signal-to-noise ratio (SNR), the Root Mean Square Error (RMSE), and the waveform similarity coefficient (NCC); for early failure signals with different signal-to-noise ratios, the SNR, RMSE, NCC effect values after VMD denoising are all superior to those of the other three algorithms, and it can be seen that the VMD denoising effect is the best.
Step S2: firstly, a maximum value (Max), a minimum value (Min), a peak-peak value (Ptp), an average value (Mean), a variance (Var), a kurtosis (Kv), a skewness (Sk), a peak factor (Par) and the first five scale values of a refined multi-scale dispersion entropy (RCMDE) of a signal after decomposition and noise reduction are selected to construct a characteristic vector F, and then the characteristic vector F is subjected to normalization processing to be distributed in [ -1,1] and used as the input of a DBN model.
The calculation method of the time domain features and the related parameters in the refined multi-scale dispersion entropy (RCMDE) is as follows:
peak-to-peak (Ptp): x is the number ofptp＝xmax-xmin；
In the formula: x is the number ofptpIs the peak-to-peak value, xmaxFor resolving the maximum, x, of the noise-reduced signalminIs the minimum value of the decomposed noise-reduced signal;
mean (Mean):
Figure BDA0003274137620000161
in the formula:
Figure BDA0003274137620000162
is an arithmetic mean, x is the original signal, and N is the signal sequence length;
variance (Var):
Figure BDA0003274137620000163
in the formula: x is the number ofvarIs the variance;
kurtosis (Kv):
Figure BDA0003274137620000171
in the formula: x is the number ofkvIs the kurtosis;
skewness (Sk):
Figure BDA0003274137620000172
in the formula: is xskSkewness, xstdIs the standard deviation of the measured data to be measured,
Figure BDA0003274137620000173
crest factor (Par):
Figure BDA0003274137620000174
in the formula: x is the number ofparIs the crest factor, xrmsIs the root mean square (rms) value,
Figure BDA0003274137620000175
refined multiscale dispersion entropy (RCMDE):
Figure BDA0003274137620000176
in the formula: eRCMD(x, m, c, d, τ) is the embedding dimension, m is the embedding dimension, c is the number class, d is the time delay, τ is the scale,
Figure BDA0003274137620000177
is the average value of the probability of the scattering pattern corresponding to the coarse grained sequence,
Figure BDA0003274137620000178
is the probability of the spreading pattern corresponding to the kth coarse-grained sequence at the scale τ; in fig. 6, the abscissa is a scale factor, and the ordinate is an RCMDE value, the RCMDE value of each state after VMD denoising becomes smooth, and the curves do not cross each other any more, and can be used in subsequent research.
Step S3: initializing parameters of a DBN model, determining the number of hidden layers of the DBN, optimizing the number of hidden layer neurons of the DBN model through a PSO algorithm, determining the optimal structure of the DBN, and determining the number of the hidden layers and the number of the hidden layer neurons: determining the number of hidden layers by adopting a control variable method, namely setting the number of neurons in each layer to be the same value, adjusting the number of the hidden layers, and determining the proper number of the hidden layers through the error after model training; aiming at the setting of the number of the neurons of the hidden layer, the optimal number of the neurons of the hidden layer is selected through a PSO algorithm; and setting the fitness function of the PSO algorithm as the identification accuracy rate after the DBN model is trained.
Referring to fig. 7, the abscissa is the number of hidden layers, the ordinate is the accuracy, the number of neurons in each layer is set to 200, the number of iterations is set to 500, the experiment is performed by using the same training set and test set, and when the number of hidden layers is 3, the accuracy is the highest, and the accuracy decreases as the number of hidden layers increases, so that the number of hidden layers of the DBN model of the present invention is 3.
With reference to fig. 8, the abscissa is the iteration number, the ordinate is the fitness value, the number of hidden layer neurons of the DBN model is optimized by the PSO algorithm, the iteration number is set to 20, and the fitness value tends to be stable and substantially unchanged after the iteration number reaches 12, so that the final optimal number of hidden layer neurons is: 210-201-193.
Step S4: determining relevant parameters of the DBN model, including a loss function, RBM learning rate, RBM iteration times, network layer number and the like, and finishing model training; the training of the DBN comprises two steps of unauthorized pre-training and fine-tuning layer by layer; in the stage before training, training the RBMs by adopting a layer-by-layer greedy method, and after the training of the previous RBM is finished, taking a hidden layer of the previous RBM as a visible layer of the next RBM; thus, the RBMs can be trained one by one until the last RBM is trained; each RBM is trained by maximizing the probability of its input data, updating parameters with a contrast-divergence (CD) algorithm; after pre-training, add the classification layer to the last hidden layer and further fine-tune the DBN by minimizing the error between the estimated output value and the label; the BP algorithm is adopted to gradually transfer errors from the last layer to the bottom input layer, so that the parameters of the whole network can be updated.
In combination with fig. 9, the abscissa is the number of iterations, the ordinate is the mean square error and the accuracy, the mean square error obtained from the graph gradually decreases with the increase of the number of iterations, and the accuracy of the training set and the accuracy of the test set both gradually increase with the increase of the number of iterations; therefore, to obtain higher fault identification accuracy, the iteration times need to be increased, but the iteration time is prolonged; when the iteration times reach 400 times, the mean square error and the accuracy fluctuation range of the training set and the test set are small and tend to be stable, and in order to shorten the network training time, the DBN iteration time is selected to be 400.
The DBN model with optimized parameters is adopted for performance testing, the anti-noise performance of the model is tested through sample data with different signal-to-noise ratios, and the effectiveness and superiority of the method are verified through comparison with other identification models; selecting an original signal to perform a comparison experiment in order to verify the effectiveness of the selected characteristic; selecting an original signal to perform a comparison experiment for verifying the effectiveness of the selected time domain feature and the RCMDE feature; in order to verify the robustness and accuracy of the VMD + DBN method, Gaussian white noise with different signal-to-noise ratios is added to an original signal, and an Extreme Learning Machine (ELM), a Probabilistic Neural Network (PNN) and a Generalized Regression Neural Network (GRNN) are selected to perform a comparison experiment; in order to verify the optimization effect of PSO, an unoptimized DBN model is selected for comparison experiments.
The DBN model adopts the optimal parameters, the original data are directly imported into the DBN model without any pretreatment and characteristic extraction for comparative analysis, and the experimental results are shown in table 2.
TABLE 2 comparison of VMD feature extraction to raw signal
Figure BDA0003274137620000191
With reference to fig. 10, the abscissa represents the gaussian white noise added to the original signal with different signal-to-noise ratios, and the ordinate represents the identification accuracy of the Extreme Learning Machine (ELM), the Probabilistic Neural Network (PNN), and the Generalized Regression Neural Network (GRNN) under different signal-to-noise ratios; the same training set and test set are adopted for all the four networks, the training set and the test set are randomly ordered, and the identification results of the four networks are shown in table 3.
TABLE 3 comparison of the accuracy of the different methods
Figure BDA0003274137620000201
The DBN model adopts random parameters and optimal parameters to classify and identify the VMD denoised data, and the experimental result is shown in Table 4.
TABLE 4 VMD + PSO-DBN vs. VMD + DBN
Figure BDA0003274137620000202
In summary, the invention provides a cable early fault identification method based on VMD and PSO-DBN, by comparing original signal + DBN with VMD extraction feature + DBN, VMD extraction feature + ELM, VMD extraction feature + GRNN and VMD extraction feature + PNN, and comparing and analyzing VMD + DBN and VMD + PSO-DBN, the technical scheme has the following advantages:
the VMD can adaptively solve the constructed constraint variational equation to realize effective decomposition of the signal components, wherein when the number of decomposition modal layers K is 4, the optimal decomposition effect is achieved; the PSO algorithm can optimize the DBN model structure and reduce the training time, wherein when the number of hidden layer layers is 3 and the number of optimal hidden layer neurons is 210-.
2. The VMD and PSO-DBN method provided by the invention is suitable for identifying early faults of cables, and compared with ELM, GRNN and PNN, the accuracy of the DBN as an identification classification network is higher; compared with the non-optimized DBN, the PSO-DBN reduces the time of model training, improves the identification efficiency and has higher identification accuracy.

Claims (10)

1. A cable early fault identification method is characterized by comprising the following steps:
step S1: selecting cable fault phase current as an original signal, and preprocessing the original signal; the preprocessing comprises the steps of carrying out signal decomposition and noise reduction on the original signal by utilizing the VMD, and reducing Gaussian white noise in the original signal;
step S2: selecting time domain features of the preprocessed signals and the first five scale values of refined multi-scale dispersion entropy to construct a feature vector F, normalizing the extracted feature vector F to construct an identification model input feature vector, and enabling the identification model input feature vector to be distributed in [ -1,1] and serve as a feature vector input to the depth confidence network;
step S3: initializing DBN model training parameters, training the DBN model through pre-training, and optimizing a deep belief network structure by utilizing a PSO algorithm.
2. The method for identifying an early failure of a cable according to claim 1, further comprising:
step S4: determining other relevant parameters of the DBN model, and finishing model training to obtain a PSO-DBN model; the other related parameters comprise a loss function, RBM learning rate, RBM iteration times and network layer number.
3. The method as claimed in claim 1 or 2, wherein the VMD is used to decompose and reduce noise of the original signal, specifically including decomposing the original signal to reduce white gaussian noise in the signal.
4. The method as claimed in claim 3, wherein the signal decomposition of the original signal by the VMD comprises a reasonable setting of the number of VMD decomposition layers K, specifically: decomposing a noisy signal in an original signal by adopting a VMD (virtual matrix display), wherein the number K of decomposition layers is gradually increased from 2, and the central frequency of each modal component is recorded after each decomposition; according to the idea of the center frequency method, when the center frequency is closest to the center frequency when K is 2, the signal is considered to be over-decomposed, and the K value when the center frequency is closest to the center frequency when K is 2 is used as the decomposition layer number of the VMD algorithm.
5. The method for identifying early cable faults as claimed in claim 1 or 2, wherein the building identification model input feature vectors are specifically: normalizing a feature vector F formed by time domain features of the decomposed and denoised original signal and the first five scale values of the refined multi-scale dispersion entropy to enable the feature vector to be distributed in [ -1,1] and serve as the input of a deep confidence network model; the time domain characteristics of the decomposed and noise-reduced original signal comprise the maximum value, the minimum value, the peak-peak value, the average value, the variance, the kurtosis, the skewness and the peak value factor of the decomposed and noise-reduced original signal.
6. A method as claimed in claim 1 or 2, wherein the time domain characteristics are calculated by the following method for each relevant parameter:
peak-to-peak: x is the number ofptp＝xmax-xmin；
In the formula: x is the number ofptpIs the peak-to-peak value, xmaxIs the maximum value, x, of the original signal after decomposition and noise reductionminThe minimum value of the original signal after decomposition and noise reduction processing is obtained;
average value:
Figure FDA0003274137610000021
in the formula:
Figure FDA0003274137610000022
is an arithmetic mean, x is the original signal, and N is the signal sequence length;
variance:
Figure FDA0003274137610000023
in the formula: x is the number ofvarIs the variance;
kurtosis:
Figure FDA0003274137610000031
in the formula: x is the number ofkvIs the kurtosis;
skewness:
Figure FDA0003274137610000032
in the formula: is xskSkewness, xstdIs the standard deviation of the measured data to be measured,
Figure FDA0003274137610000033
crest factor:
Figure FDA0003274137610000034
in the formula: x is the number ofparIs the crest factor, xrmsIs the root mean square (rms) value,
Figure FDA0003274137610000035
7. a method as claimed in claim 1 or 2, wherein the refined multi-scale entropy distribution is calculated as follows:
Figure FDA0003274137610000036
in the formula: eRCMD(x, m, c, d, τ) is the embedding dimension, m is the embedding dimension, c is the number class, d is the time delay, τ is the scale,
Figure FDA0003274137610000037
is the average value of the probability of the scattering pattern corresponding to the coarse grained sequence,
Figure FDA0003274137610000038
Figure FDA0003274137610000039
is the probability of the spreading pattern corresponding to the kth coarse-grained sequence at the scale τ.
8. The method for identifying the early failure of the cable according to claim 1 or 2, wherein the optimizing the deep belief network structure by using the particle swarm optimization algorithm specifically comprises: determining the number of hidden layer layers and determining the number of neurons of the hidden layers; and determining the number of the hidden layers specifically comprises determining the number of the hidden layers by adopting a control variable method, setting the same value for the number of neurons in each layer, adjusting the number of the hidden layers, and determining the proper number of the hidden layers according to the error after the DBN model training.
9. The method for identifying early cable faults as claimed in claim 8, wherein the determining the number of hidden layer neurons specifically includes: and selecting the number of neurons of the optimal hidden layer through a PSO algorithm, wherein the fitness function of the PSO algorithm is set as the identification accuracy rate after the DBN model is trained.
10. A method for early cable fault identification as claimed in claim 1 or 2 wherein training of the DBN model: unauthorized pre-training and fine-tuning layer by layer; in the pre-training stage, a training sample is used as the input of a DBN model, a layer-by-layer greedy method is adopted to train a limited Boltzmann machine, each RBM is trained by maximizing the probability of input data, and a contrast divergence algorithm is utilized to update parameters;
after pre-training, adding a classification layer to the last hidden layer, further fine-tuning the DBN by minimizing the error between an estimated output value and a label, fine-tuning the DBN model reversely by a BP algorithm in a supervision mode, updating parameters of the whole network, and inputting the extracted feature vectors into the PSO-DBN model for classification and identification after training.
CN202111111740.9A 2021-09-23 2021-09-23 Cable early fault identification method Pending CN113971416A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111111740.9A CN113971416A (en) 2021-09-23 2021-09-23 Cable early fault identification method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111111740.9A CN113971416A (en) 2021-09-23 2021-09-23 Cable early fault identification method

Publications (1)

Publication Number Publication Date
CN113971416A true CN113971416A (en) 2022-01-25

Family

ID=79586780

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111111740.9A Pending CN113971416A (en) 2021-09-23 2021-09-23 Cable early fault identification method

Country Status (1)

Country Link
CN (1) CN113971416A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115481916A (en) * 2022-09-27 2022-12-16 呼伦贝尔安泰热电有限责任公司海拉尔热电厂 Method and system for identifying heating area heat index in central heating in alpine region
CN116256592A (en) * 2022-11-28 2023-06-13 国网山东省电力公司德州供电公司 Medium-voltage distribution cable latent fault detection method and system
CN117092206A (en) * 2023-08-09 2023-11-21 国网四川省电力公司电力科学研究院 Defect detection method for cable lead sealing area, computer equipment and storage medium
CN117195097A (en) * 2023-08-09 2023-12-08 南通大学 Cloud-to-ground flash identification method based on lightning electric field signals

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115481916A (en) * 2022-09-27 2022-12-16 呼伦贝尔安泰热电有限责任公司海拉尔热电厂 Method and system for identifying heating area heat index in central heating in alpine region
CN115481916B (en) * 2022-09-27 2023-09-19 呼伦贝尔安泰热电有限责任公司海拉尔热电厂 Method and system for identifying heat index of heating area in central heating in alpine region
CN116256592A (en) * 2022-11-28 2023-06-13 国网山东省电力公司德州供电公司 Medium-voltage distribution cable latent fault detection method and system
CN116256592B (en) * 2022-11-28 2023-09-26 国网山东省电力公司德州供电公司 Medium-voltage distribution cable latent fault detection method and system
CN117092206A (en) * 2023-08-09 2023-11-21 国网四川省电力公司电力科学研究院 Defect detection method for cable lead sealing area, computer equipment and storage medium
CN117195097A (en) * 2023-08-09 2023-12-08 南通大学 Cloud-to-ground flash identification method based on lightning electric field signals

Similar Documents

Publication Publication Date Title
CN113971416A (en) Cable early fault identification method
CN110728360B (en) Micro-energy device energy identification method based on BP neural network
CN109597043B (en) Radar signal identification method based on quantum particle swarm convolutional neural network
CN109597401B (en) Equipment fault diagnosis method based on data driving
CN105678343B (en) Hydropower Unit noise abnormality diagnostic method based on adaptive weighted group of sparse expression
CN112861066B (en) Machine learning and FFT (fast Fourier transform) -based blind source separation information source number parallel estimation method
CN109344751B (en) Reconstruction method of noise signal in vehicle
CN111985528B (en) PDGAN-based cable partial discharge data enhancement method
CN113378652B (en) Disturbance classification method based on EWT-MPE-PSO-BP
CN112613493A (en) Rolling bearing fault diagnosis method based on multi-feature extraction and WOA-ELM
CN113076920B (en) Intelligent fault diagnosis method based on asymmetric domain confrontation self-adaptive model
CN112200038A (en) CNN-based rapid identification method for oscillation type of power system
CN112434743A (en) Fault identification method based on GIL metal particle partial discharge time domain waveform image
CN111709350B (en) Low-frequency oscillation modal parameter identification method and system based on FCM clustering
CN113111786A (en) Underwater target identification method based on small sample training image convolutional network
CN117590173A (en) Cable partial discharge pattern recognition method based on convolutional neural network
CN112946600A (en) Method for constructing radar HRRP database based on WGAN-GP
Wang et al. An Automatic Identification Framework for Complex Power Quality Disturbances Based on Ensemble CNN
Zheng et al. Performance comparison of full-batch BP and mini-batch BP algorithm on spark framework
CN117516588A (en) Laser gyroscope fault diagnosis method based on optimized kernel extreme learning machine
CN112329535B (en) CNN-based quick identification method for low-frequency oscillation modal characteristics of power system
CN110610203A (en) Electric energy quality disturbance classification method based on DWT and extreme learning machine
CN115879412A (en) Layout level circuit diagram size parameter optimization method based on transfer learning
CN113391172A (en) Partial discharge diagnosis method and system based on time sequence integration and used for multi-source ultrasonic detection
Eltanashi et al. Proposed speaker recognition model using optimized feed forward neural network and hybrid time-mel speech feature

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