CN116540025A - Fault detection method based on transfer learning and residual error network - Google Patents
Fault detection method based on transfer learning and residual error network Download PDFInfo
- Publication number
- CN116540025A CN116540025A CN202310565793.0A CN202310565793A CN116540025A CN 116540025 A CN116540025 A CN 116540025A CN 202310565793 A CN202310565793 A CN 202310565793A CN 116540025 A CN116540025 A CN 116540025A
- Authority
- CN
- China
- Prior art keywords
- fault
- time
- network
- zero sequence
- sequence voltage
- 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
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 23
- 238000013526 transfer learning Methods 0.000 title claims abstract description 20
- 238000010586 diagram Methods 0.000 claims abstract description 21
- 238000012549 training Methods 0.000 claims abstract description 15
- 230000009466 transformation Effects 0.000 claims abstract description 15
- 238000004458 analytical method Methods 0.000 claims abstract description 11
- 239000003990 capacitor Substances 0.000 claims abstract description 11
- 238000005070 sampling Methods 0.000 claims abstract description 10
- 238000013507 mapping Methods 0.000 claims description 24
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 13
- 238000004422 calculation algorithm Methods 0.000 claims description 12
- 238000000034 method Methods 0.000 claims description 12
- 230000007935 neutral effect Effects 0.000 claims description 12
- 238000013528 artificial neural network Methods 0.000 claims description 11
- 238000013508 migration Methods 0.000 claims description 10
- 230000005012 migration Effects 0.000 claims description 10
- 230000000737 periodic effect Effects 0.000 claims description 9
- 238000000605 extraction Methods 0.000 claims description 7
- 238000013519 translation Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 5
- 230000008901 benefit Effects 0.000 claims description 4
- 230000008859 change Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 238000004880 explosion Methods 0.000 claims description 3
- 230000008014 freezing Effects 0.000 claims description 3
- 238000007710 freezing Methods 0.000 claims description 3
- 238000001228 spectrum Methods 0.000 claims description 3
- 238000012360 testing method Methods 0.000 claims description 3
- 238000011426 transformation method Methods 0.000 claims description 3
- 230000008034 disappearance Effects 0.000 claims 1
- 238000012546 transfer Methods 0.000 claims 1
- 238000003745 diagnosis Methods 0.000 abstract description 4
- 238000013461 design Methods 0.000 abstract description 2
- 238000010891 electric arc Methods 0.000 description 6
- 230000006872 improvement Effects 0.000 description 6
- 238000011160 research Methods 0.000 description 5
- 238000004088 simulation Methods 0.000 description 5
- 238000000354 decomposition reaction Methods 0.000 description 4
- 230000016507 interphase Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000007637 random forest analysis Methods 0.000 description 3
- 230000007704 transition Effects 0.000 description 3
- 235000010643 Leucaena leucocephala Nutrition 0.000 description 2
- 230000015556 catabolic process Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000006731 degradation reaction Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000036039 immunity Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 244000020998 Acacia farnesiana Species 0.000 description 1
- 206010000369 Accident Diseases 0.000 description 1
- 240000007472 Leucaena leucocephala Species 0.000 description 1
- 239000010426 asphalt Substances 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008033 biological extinction Effects 0.000 description 1
- 238000010219 correlation analysis Methods 0.000 description 1
- 239000013078 crystal Substances 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000009413 insulation Methods 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000010183 spectrum analysis Methods 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
- 238000011144 upstream manufacturing Methods 0.000 description 1
Classifications
-
- 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
- G01R31/08—Locating faults in cables, transmission lines, or networks
- G01R31/081—Locating faults in cables, transmission lines, or networks according to type of conductors
- G01R31/086—Locating faults in cables, transmission lines, or networks according to type of conductors in power transmission or distribution networks, i.e. with interconnected conductors
-
- 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
- G01R31/08—Locating faults in cables, transmission lines, or networks
- G01R31/088—Aspects of digital computing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
- Y04S10/52—Outage or fault management, e.g. fault detection or location
Abstract
The invention relates to the technical field of single-phase earth fault identification of power distribution networks, in particular to a fault detection method based on transfer learning and a residual error network. Comprising the following steps: detecting the zero sequence voltage of the power distribution network, and if the effective value of the zero sequence voltage exceeds 0.15Um, considering that a fault occurs in the power distribution network; sampling time length is zero sequence voltage signals of two periods, and performing time-frequency analysis on the zero sequence voltage; adopting continuous wavelet transformation to obtain a two-dimensional time-frequency diagram of zero sequence voltage; inputting the time-frequency diagram into a depth residual error network to further judge the fault type; the possibility of load and capacitor switching interference is eliminated, and then the arc light high-resistance grounding fault can be judged. The invention designs that the zero-rest phenomenon always appears periodically, can continuously carry out wavelet transformation and extract time-frequency information as fault characteristics; the zero sequence voltage waveform is a strong time sequence, resNet18 models can be effectively distinguished by training on an image data set of the ImageNet, fault diagnosis accuracy is high, and model training time is short.
Description
Technical Field
The invention relates to the technical field of single-phase earth fault identification of power distribution networks, in particular to a fault detection method based on transfer learning and a residual error network.
Background
At present, single-phase earth faults in a power distribution network occupy a relatively high proportion, and accurate online identification of faults occurring in the power distribution network is an important ring for rapidly processing the faults. If the fault type cannot be timely determined, the fault can be transmitted upstream or downstream along the power distribution network, so that the fault can be further developed into various possibilities such as interphase faults, and the safe operation of the power distribution network is seriously threatened.
One important challenge in single-phase earth fault type identification comes from the variety of types. From the perspective of the transition resistance, single-phase ground faults can be divided into metallic ground, low-resistance ground and high-resistance ground. There is no clear limit for the resistance value, and it is considered that a high-resistance ground fault occurs when the transition resistance reaches 500 Ω or more, depending on the cause of the fault.
According to the fault cause, tree barriers, wire falling and the like are called high-resistance grounding, and experimental data show that the high-resistance ground resistance is generally above 500-1.000 ohms, and even can reach tens of thousands ohms for dry asphalt pavement, the resistance can reach above 3kΩ during tree flash discharge, and the human body electric shock resistance can reach above 1.5-2.5 kΩ.
Because of the difficulty in detection of high-resistance faults, the distribution line cannot always remove faults in time after the high-resistance ground faults occur, interphase insulation damage is caused by continuous operation, interphase faults are even caused, and safe operation of a power grid is seriously threatened. The high-resistance ground fault is accompanied by the generation of an electric arc, the change of the electric arc resistance can be greatly changed in a short time along with the reburning and extinguishing of the electric arc, and even the misoperation of protection devices of adjacent lines and equipment can be caused, so that the electric shock and fire accidents are caused.
The high-resistance grounding fault needs to be accurately detected, and arc extinction, overvoltage protection and even fault line or section cutting and the like are carried out. And compared with the metallic grounding and the low-resistance grounding, the high-resistance grounding fault has less obvious electrical quantity characteristics (the zero sequence voltage can be less than 15% of the phase voltage, the fault current can be less than 1A), the fault point is more unstable, and the detection and the processing are more difficult.
It is worth fortunately that the type of the fault is understood and identified according to the rich fault information contained in the fault transient state electric quantity of the power distribution network, namely, the fault is classified and identified after the feature is extracted from the fault signal. For example, in a "accurate positioning method and system for power distribution network faults based on random forest" with publication number CN202211005095.7, bai Hao et al implement feature extraction and feature screening through wavelet transformation and random forest, and the screened fault features are classified by a cooperative training semi-supervised classifier composed of random forest, so that high-resistance ground fault and disturbance events can be reliably distinguished under the condition of a small number of labeled training samples, and certain noise immunity is achieved. In a new method for selecting a power distribution line ground fault of a resonant ground system, which is cited in journal of electric power system protection and control, once crystal et al firstly sample neutral point zero sequence current and then perform generalized S transformation on the neutral point zero sequence current, and after normalization processing is performed on energy corresponding to each row of an obtained time-frequency matrix, the data are input into an extreme learning machine for recognition. In the support vector machine-based distribution line high-resistance ground fault detection method cited in journal of electronic design engineering, zheng Xingjiong et al fourier-transform zero-order current samples and input the energy component eigenvectors of fundamental waves, harmonics and inter-harmonics thereof into a support vector machine. When the fault closing angle of the grounding resistor is under a specific condition, misjudgment can occur to the algorithm. In the self-adaptive single-phase earth fault on-line positioning research and application based on multi-information fusion cited in journal of electric measurement and instruments, xu Guang et al apply big data technology theory and fuzzy C-means clustering analysis to extract the threshold value of fault characteristic quantity, thereby improving single-phase earth fault accuracy research. However, research on the current algorithm still has difficulty in overcoming the following challenges:
spectral analysis is susceptible to failure from disturbances (e.g., switching operations, load switching, capacitor investment, etc.) similar to the characteristics of high-resistance faults;
the algorithm is susceptible to noise, and the noise immunity is weak, so that the sensitivity of the algorithm is low.
In order to solve the above problems, the present invention provides a fault detection method based on transfer learning and a residual error network.
Disclosure of Invention
The invention aims to provide a fault detection method based on transfer learning and a residual error network, so as to solve the problems in the background technology.
In order to solve the above technical problems, the present invention provides a fault detection method based on transfer learning and residual error network, comprising the following steps:
s1, detecting zero sequence voltage of a power distribution network, and if the effective value of the zero sequence voltage exceeds 0.15Um, considering that a fault occurs in the power distribution network;
s2, performing time-frequency analysis on the zero sequence voltage by using a zero sequence voltage signal with sampling time length of two periods; adopting continuous wavelet transformation to obtain a two-dimensional time-frequency diagram of zero sequence voltage;
s3, inputting the time-frequency diagram into a depth residual error network to further judge the fault type; through the collection analysis of a plurality of rounds, after eliminating the possibility of switching interference of the load and the capacitor, the arc light high-resistance grounding fault can be judged.
Among them, the deep learning method is widely used to solve the classification problem of time series, such as a multi-level wavelet decomposition network (Multilevel Wavelet Decomposition Network, mWDN), a time series attention prototype network (Time Series Attentional Prototype Network, tapNet), and the like. By converting the time series data into a two-dimensional time series image, a model of matured and stable visual field can be applied. Considering that the zero sequence voltage waveform is taken as a strong time sequence, the front and rear inputs of the zero sequence voltage waveform have strong relevance, the depth residual error network is adopted to identify the potential time relevance in the time sequence image, and the potential time relevance in the time sequence image is learned.
As a further improvement of the technical scheme, in S2, the time-frequency analysis is performed on the zero-sequence voltage, so as to determine whether a high-resistance ground fault occurs in the power distribution network, which is based on the following steps:
when the high-resistance ground fault HIF occurs in the power distribution network, the resistance of a ground medium is larger, the fault current is much smaller than the load current, the fault characteristic is not obvious, but the zero-break phenomenon always occurs periodically when the HIF occurs, the frequency domain characteristic cannot characterize the periodic occurrence rule of the zero-break phenomenon, continuous wavelet transformation is introduced, the frequency domain characteristic of the high-resistance ground fault is extracted by a time-frequency domain transformation method, the time-frequency domain characteristic of the high-resistance ground fault is extracted, and a two-dimensional time-frequency diagram is obtained to judge whether the high-resistance ground fault occurs.
As a further improvement of the technical scheme, in the step S2, when the feature extraction is performed on the time domain signal, a one-dimensional voltage signal feature extraction algorithm is provided based on wavelet transformation, and the one-dimensional signal is converted into a two-dimensional time-frequency signal; wherein, the continuous wavelet transform of W (x) shown in the core formula (1) is defined as:
wherein ,
wherein S is a scale factor, and S is more than 0; τ is a translation factor;as a function of scale factors and translation factors.
The sampling frequency of the scheme is selected to be 1kHZ. The minimum and maximum dimensions are automatically determined based on the energy spread of the wavelet in frequency and time. Compared with the weak zero-break phenomenon in the time domain signal, the periodic variation of the high-frequency harmonic content in the time-frequency diagram is more obvious. By comparing the zero sequence voltage time-frequency diagram, the odd harmonic content is obvious when the arc light high-resistance grounding fault occurs in the neutral point ungrounded system, so that the zero sequence voltage waveform is distorted.
As a further improvement of the technical scheme, in the step S3, a zero sequence voltage waveform is observed as a strong time sequence, and the principle of strong relevance is input before and after the zero sequence voltage waveform, so that the identification network is constructed by adopting a distribution network ground fault type identification algorithm based on transfer learning and a residual network, and potential time relevance in a time sequence image is learned; wherein the model selects ResNet18 model trained on the ImageNet image dataset.
As a further improvement of the technical scheme, in the S3, in a ResNet18 model, the low-layer characteristics have strong migration capability, the characteristics of the high-layer convolution layers are abstract characteristics related to specific tasks, and more high-layer convolution layer parameters need to be trained and updated for data sets with larger differences;
at this time, shallow fault characteristics are extracted by adopting a migration learning method, and deep fault characteristics are further learned; the trained network architecture can realize small sample migration through freezing parameters, and fine tuning parameters are tested under different topological structures;
the expression capacity of the neural network can be enhanced by increasing the number of network layers, but the gradient vanishing and gradient explosion problems can be caused according to the neural network counter-propagation principle, and the deep residual error network is used for fundamentally solving the degradation problem not only by introducing BatchNorm, but also by residual error learning.
As a further improvement of the technical solution, in S3, compared with the conventional convolutional network, the residual structure diagram in the res net18 model introduces a shortcut connection in the residual block. The method has the function of realizing the identity mapping of the input, and leading the input x of the neural network to be directly connected to the output x of the parameter layer through an identity mapping I, wherein x-x is the same as x;
the mapping relation corresponding to the connection of the input part and the output part can be respectively expressed as F (x) and x, and the whole mapping of the residual block is F (x) +x;
if the expected mapping of the residual block is H (x), the expected mapping relation of the convolution layer in the residual block is F (x) =h (x) -x, namely, the residual between the expected mapping and the identity mapping;
meanwhile, a normalization layer is introduced to improve the model training speed. Research shows that the structure reduces the learning difficulty and can build a deep neural network.
As a further improvement of the present technical solution, in S3, in the res net18 model:
when load and capacitor switching occur, the current can also have fluctuation of the same order of magnitude, and each feeder line can generate interference similar to the frequency spectrum characteristic of the zero sequence current when high-resistance faults occur;
the high-resistance grounding fault has the advantages that the fault current characteristics change under different grounding media and neutral point grounding modes, and the zero-break phenomenon can be weak or offset and other conditions occur; this makes the high resistance ground fault not a fixed feature, difficult to detect;
because the switching interference of the load and the capacitor is usually only in a quite short period of time, the zero sequence voltage signal of two periods is sampled only, and whether the signal has arc high-resistance grounding fault cannot be judged; the sampling frequency of the scheme is 1MHz, and arc high resistance grounding is judged when the zero sequence voltage has periodic fault characteristics. Because the time of the collected signals is short, the acquisition is performed for fifty times as one round, and the execution time of the algorithm is not more than 2s.
Compared with the prior art, the invention has the beneficial effects that:
1. in the fault detection method based on the transfer learning and residual error network, continuous wavelet transformation can be performed for the periodic occurrence of the zero-rest phenomenon, and time-frequency information is extracted as fault characteristics;
2. in the fault detection method based on the transfer learning and residual error network, the zero sequence voltage waveform is a strong time sequence, the ResNet18 model can be effectively distinguished by training on the ImageNet image data set, the fault diagnosis accuracy is high, and the model training time is short.
Drawings
FIG. 1 is a flow chart illustrating exemplary single phase ground fault type identification in accordance with the present invention;
FIG. 2 is a schematic diagram of an exemplary ResNet 18-based ground fault type identification model architecture in accordance with the present invention;
FIG. 3 is a diagram of the residual block in an exemplary ResNet18 of the present invention;
FIG. 4 is a schematic diagram of an exemplary neutral ungrounded system single-phase ground fault in accordance with the present invention;
FIG. 5 is a simulation of a single phase ground fault in accordance with an exemplary embodiment of the present invention;
fig. 6 is a graph of accuracy of an exemplary model of the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Example 1
As shown in fig. 1-6, the present embodiment provides a fault detection method based on migration learning and residual error network, which includes the following steps.
S1, detecting zero sequence voltage of a power distribution network, and if the effective value of the zero sequence voltage exceeds 0.15Um, considering that a fault occurs in the power distribution network;
in the step, the basis for judging whether the power distribution network has faults or not by detecting the zero sequence voltage of the power distribution network is as follows:
zero sequence voltage is not generally existed in a normal power distribution network, when a grounding short circuit occurs in a neutral point direct grounding system or when a single-phase grounding occurs in a non-direct grounding system, zero sequence voltage is generated, and asymmetric operation and single-phase operation are main reasons for generating zero sequence current;
when the system fails, the three phases become asymmetric, and at the moment, the negative sequence component and the zero sequence component or one of the negative sequence component and the zero sequence component with amplitude values can be decomposed, and the system can be judged to be failed by detecting the two components which are not supposed to normally appear, particularly the zero sequence component when the single phase grounding occurs.
S2, performing time-frequency analysis on the zero sequence voltage by using a zero sequence voltage signal with sampling time length of two periods; adopting continuous wavelet transformation to obtain a two-dimensional time-frequency diagram of zero sequence voltage;
in this step, time-frequency analysis is performed on the zero-sequence voltage, so as to determine whether a high-resistance ground fault occurs in the power distribution network, where the basis is:
when the high-resistance ground fault HIF occurs in the power distribution network, the resistance of a ground medium is larger, the fault current is much smaller than the load current, the fault characteristic is not obvious, but the zero-break phenomenon always occurs periodically when the HIF occurs, the frequency domain characteristic cannot characterize the periodic occurrence rule of the zero-break phenomenon, continuous wavelet transformation (continue wavelet transform, CWT) is introduced, the frequency domain characteristic is extracted by a time-frequency domain transformation method, the time-frequency domain characteristic is extracted, and a two-dimensional time-frequency diagram is obtained to judge whether the high-resistance ground fault occurs.
Specifically, when the feature extraction is carried out on the time domain signal, a one-dimensional voltage signal feature extraction algorithm is firstly provided based on wavelet transformation, and the one-dimensional signal is converted into a two-dimensional time-frequency signal; wherein, the continuous wavelet transform of W (x) shown in the core formula (1) is defined as:
wherein ,
wherein S is a scale factor, and S is more than 0; τ is a translation factor;as a function of scale factors and translation factors.
In this embodiment, the sampling frequency is selected to be 1kHZ. The minimum and maximum dimensions are automatically determined based on the energy spread of the wavelet in frequency and time. Compared with the weak zero-break phenomenon in the time domain signal, the periodic variation of the high-frequency harmonic content in the time-frequency diagram is more obvious. By comparing the zero sequence voltage time-frequency diagram, the odd harmonic content is obvious when the arc light high-resistance grounding fault occurs in the neutral point ungrounded system, so that the zero sequence voltage waveform is distorted.
S3, inputting the time-frequency diagram into a depth residual error network to further judge the fault type; and collecting 50 rounds in total, and judging that the arc light has high-resistance grounding faults if the possibility of load and capacitor switching interference is eliminated.
In this step, the deep learning method is widely used to solve the classification problem of time series, such as a multi-level wavelet decomposition network (Multilevel Wavelet Decomposition Network, mWDN), a time series attention prototype network (Time Series Attentional Prototype Network, tapNet), and the like. By converting the time series data into a two-dimensional time series image, a model of matured and stable visual field can be applied. Considering that the zero sequence voltage waveform is taken as a strong time sequence, the front and rear inputs of the zero sequence voltage waveform have strong relevance, the depth residual error network is adopted to identify the potential time relevance in the time sequence image, and the potential time relevance in the time sequence image is learned.
In the embodiment, the zero sequence voltage waveform is observed to be taken as a strong time sequence, and the front and rear inputs of the zero sequence voltage waveform have strong relevance, so that the identification network is constructed by adopting a distribution network ground fault type identification algorithm based on transfer learning and a residual network, and potential time relevance in a time sequence image is learned; wherein, as shown in fig. 2, the model chooses a res net18 model trained on an ImageNet image dataset.
In the ResNet18 model, the low-level features have strong migration capability, while the high-level convolution layer features are abstract features related to specific tasks, and more high-level convolution layer parameters need to be trained and updated for data sets with larger differences;
at this time, shallow fault characteristics are extracted by adopting a migration learning method, and deep fault characteristics are further learned; the trained network architecture can realize small sample migration through freezing parameters, and fine tuning parameters are tested under different topological structures;
the expression capacity of the neural network can be enhanced by increasing the number of network layers, but the gradient vanishing and gradient explosion problems can be caused according to the neural network counter-propagation principle, and the deep residual error network is used for fundamentally solving the degradation problem not only by introducing BatchNorm, but also by residual error learning.
Further, the residual structure in the ResNet18 model introduces a shortcut connection in the residual block compared to the conventional convolutional network. The method has the function of realizing the identity mapping of the input, and leading the input x of the neural network to be directly connected to the output x of the parameter layer through an identity mapping I, wherein x-x is the same as x;
the mapping relation corresponding to the connection of the input part and the output part can be respectively expressed as F (x) and x, and the whole mapping of the residual block is F (x) +x;
if the expected mapping of the residual block is H (x), the expected mapping relation of the convolution layer in the residual block is F (x) =h (x) -x, namely, the residual between the expected mapping and the identity mapping;
the introduction of the normalization layer increases the speed of model training. Research shows that the structure reduces the learning difficulty and can build a deep neural network.
Further, when load and capacitor switching occur, the current can also fluctuate by the same order of magnitude, and each feeder line can generate interference similar to the frequency spectrum characteristic of the zero sequence current in the high-resistance fault;
the high-resistance grounding fault has the advantages that the fault current characteristics change under different grounding media and neutral point grounding modes, and the zero-break phenomenon can be weak or offset and other conditions occur; this makes the high resistance ground fault not a fixed feature, difficult to detect;
because the switching interference of the load and the capacitor is usually only in a quite short period of time, the zero sequence voltage signal of two periods is sampled only, and whether the signal has arc high-resistance grounding fault cannot be judged; in this embodiment, the sampling frequency is 1MHZ, and the arc high resistance grounding is determined only when the zero sequence voltage has periodic fault characteristics. Because the time of the collected signals is short, fifty times of acquisition are preferable as one round, and the execution time of the algorithm is not more than 2s.
In addition, a single-phase ground fault schematic diagram of the neutral point ungrounded system is shown in fig. 4, in which The system three-phase power supply potentials are respectively given, and N is the system neutral point. A phase A single-phase earth fault occurs at the K position, wherein the arc high-resistance earth fault module is formed by connecting an arc model and a fault resistor in series and is expressed as a transition resistor of a fault point. When single-phase earth fault occurs, the fault resistor generates partial voltage and reflects the partial voltage into fault phase voltage, so that the partial voltage can be reflected through zero sequence voltage.
In order to further study the single-phase earth fault characteristics, a simulation model of a 10KV single-phase earth fault is built, as shown in figure 5. The circuit can simulate the normal operation of a simulated circuit, single-phase low-resistance ground faults, high-resistance ground faults and single-phase arc ground conditions. The total simulation time of the system is 0.2s, and the sampling frequency is 1MHz.
The high-resistance grounding fault is accompanied with intermittent electric arc generation, an electric arc model is built through a self-defined assembly, and a low-resistance electric arc grounding sampling Cassie model is built. The cassie arc model is mainly suitable for a large-current arcing period before current zero crossing and is more suitable for a mathematical model of low-resistance arc.
Due to the size of the single phase ground fault dataset, it is much smaller compared to the ImageNet dataset that trains the original res net. Thus, in order to quickly identify the type of failure, the network initialization parameter in this embodiment is a parameter of the initial pre-trained network. According to the analysis, we build a depth residual network framework. The input image size is converted into 224×224×3, and the picture is input to the residual module through the convolution layer. Finally, 3 categories are designated to judge the fault type through an average pooling layer, a full connection layer, a softmax layer and a classification layer. Designating the solver as sgdm, training the initial learning rate of the parameter as 0.01, and training the batch size as 64 for 5 rounds.
The scheme is completed on a Windows11 operating system, and the simulation software version is Matlab R2022b.
Firstly, randomly selecting 80% of data as a training set and 20% of data as a test set; the simulation test data comprises an arc high-resistance grounding fault, a low-resistance grounding fault sample and an arc grounding fault sample.
The operation effect is shown in fig. 6, and 100% accuracy can be achieved through 2 epochs. By introducing the residual structure, the training time can be greatly shortened, and the fault diagnosis accuracy can be improved.
In view of the above, the proposed fault detection method based on migration learning and residual error network aims at the problems of unobvious fault phenomenon, difficult detection and the like of arc high-resistance grounding of the power distribution network, and firstly, zero-sequence voltage fault characteristics are extracted. Then, a depth residual error network is built to identify the fault type. According to parameters of a power distribution network in a certain place, a power distribution network model is built for simulation. And extracting the time-frequency domain distribution characteristics of the zero sequence voltage of the distribution line after the occurrence of high-resistance grounding faults, low-resistance grounding faults and arc grounding by collecting fault data of the zero sequence voltage. Experiments show that the power distribution network fault type detection method based on the transfer learning and depth residual error network time-frequency domain correlation analysis can effectively judge that the fault type has higher effectiveness and effectiveness aiming at a neutral point ungrounded system, and can effectively solve two challenge problems in the field mentioned in the background art.
Specifically, the zero-break phenomenon can appear periodically, continuous wavelet transformation can be performed, and time-frequency information is extracted as fault characteristics; meanwhile, the method is a strong time sequence aiming at the zero sequence voltage waveform, the ResNet18 model can be effectively distinguished by training on the ImageNet image data set, the fault diagnosis accuracy is high, and the model training time is short.
Those of ordinary skill in the art will appreciate that the processes implementing all or a portion of the steps of the above embodiments may be implemented by hardware or may be implemented by a program to instruct the associated hardware.
The foregoing has shown and described the basic principles, principal features and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the above-described embodiments, and that the above-described embodiments and descriptions are only preferred embodiments of the present invention, and are not intended to limit the invention, and that various changes and modifications may be made therein without departing from the spirit and scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (7)
1. The fault detection method based on the transfer learning and residual error network is characterized by comprising the following steps:
s1, detecting zero sequence voltage of a power distribution network, and if the effective value of the zero sequence voltage exceeds 0.15Um, considering that a fault occurs in the power distribution network;
s2, performing time-frequency analysis on the zero sequence voltage by using a zero sequence voltage signal with sampling time length of two periods; adopting continuous wavelet transformation to obtain a two-dimensional time-frequency diagram of zero sequence voltage;
s3, inputting the time-frequency diagram into a depth residual error network to further judge the fault type; through the collection analysis of a plurality of rounds, after eliminating the possibility of switching interference of the load and the capacitor, the arc light high-resistance grounding fault can be judged.
2. The fault detection method based on the transfer learning and residual error network according to claim 1, wherein in S2, the time-frequency analysis is performed on the zero-sequence voltage, so as to determine whether a high-resistance ground fault occurs in the power distribution network, and the basis is that:
when the high-resistance ground fault HIF occurs in the power distribution network, the resistance of a ground medium is larger, the fault current is much smaller than the load current, the fault characteristic is not obvious, but the zero-break phenomenon always occurs periodically when the HIF occurs, the frequency domain characteristic cannot characterize the periodic occurrence rule of the zero-break phenomenon, continuous wavelet transformation is introduced, the frequency domain characteristic of the high-resistance ground fault is extracted by a time-frequency domain transformation method, the time-frequency domain characteristic of the high-resistance ground fault is extracted, and a two-dimensional time-frequency diagram is obtained to judge whether the high-resistance ground fault occurs.
3. The fault detection method based on the transfer learning and residual error network according to claim 2, wherein in the step S2, when the feature extraction is performed on the time domain signal, a one-dimensional voltage signal feature extraction algorithm is provided based on wavelet transformation, and the one-dimensional signal is converted into a two-dimensional time-frequency signal; wherein, the continuous wavelet transform of W (x) shown in the core formula (1) is defined as:
wherein ,
wherein S is a scale factor, and S is more than 0; τ is a translation factor;as a function of scale factors and translation factors.
4. The fault detection method based on the transfer learning and residual error network according to claim 1, wherein in the step S3, the recognition network is constructed by adopting a distribution network ground fault type recognition algorithm based on the transfer learning and residual error network based on the principle that the zero sequence voltage waveform is taken as a strong time sequence and has strong correlation in front and back inputs, so that potential time correlation in a time sequence image can be learned; wherein the model selects ResNet18 model trained on the ImageNet image dataset.
5. The fault detection method based on the transfer learning and residual error network according to claim 4, wherein in the step S3, in the ResNet18 model, the low-level features have strong transfer capability, the features of the high-level convolution layers are abstract features related to specific tasks, and more high-level convolution layer parameters need to be trained and updated for data sets with larger differences;
at the moment, shallow fault characteristics are extracted by adopting a migration learning method, and deep fault characteristics can be further learned; the trained network architecture can realize small sample migration through freezing parameters, and then test under different topological structures through fine tuning parameters;
the expression capacity of the neural network can be enhanced aiming at the increase of the network layer number, but the problems of gradient disappearance and gradient explosion can be caused according to the neural network counter-propagation principle, and the deep residual error network is learned by introducing BatchNorm and residual error.
6. The fault detection method based on the transfer learning and residual network according to claim 5, wherein in the step S3, compared with the conventional convolutional network, the residual structure diagram in the res net18 model introduces a shortcut connection in the residual block for realizing the identity mapping of the input, so that the input x of the neural network is directly connected to the output x of the parametric layer through an identity mapping I x→x;
the mapping relation corresponding to the connection of the input part and the output part can be respectively expressed as F (x) and x, and the whole mapping of the residual block is F (x) +x;
if the expected mapping of the residual block is H (x), the expected mapping relation of the convolution layer in the residual block is F (x) =h (x) -x, namely, the residual between the expected mapping and the identity mapping;
meanwhile, a normalization layer is introduced to improve the speed of model training.
7. The method for fault detection based on transfer learning and residual network of claim 6 wherein in S3, in the res net18 model:
when load and capacitor switching occur, the current can also have fluctuation of the same order of magnitude, and each feeder line can generate interference similar to the frequency spectrum characteristic of the zero sequence current when high-resistance faults occur;
the high-resistance grounding fault has the advantages that the fault current characteristics change under different grounding media and neutral point grounding modes, and the zero-break phenomenon can be weak or offset and other conditions occur;
because the switching interference of the load and the capacitor is usually only in a quite short period of time, the zero sequence voltage signal of two periods is sampled only, and whether the signal has arc high-resistance grounding fault cannot be judged; and determining that the arc high resistance is grounded when the zero sequence voltage has periodic fault characteristics.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310565793.0A CN116540025A (en) | 2023-05-19 | 2023-05-19 | Fault detection method based on transfer learning and residual error network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310565793.0A CN116540025A (en) | 2023-05-19 | 2023-05-19 | Fault detection method based on transfer learning and residual error network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116540025A true CN116540025A (en) | 2023-08-04 |
Family
ID=87455860
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310565793.0A Pending CN116540025A (en) | 2023-05-19 | 2023-05-19 | Fault detection method based on transfer learning and residual error network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116540025A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117826019A (en) * | 2024-03-06 | 2024-04-05 | 国网吉林省电力有限公司长春供电公司 | Line single-phase grounding fault area and type detection method of neutral point ungrounded system |
| CN117909909A (en) * | 2024-03-19 | 2024-04-19 | 青岛鼎信通讯股份有限公司 | Arc grounding fault identification method for power distribution network |
-
2023
- 2023-05-19 CN CN202310565793.0A patent/CN116540025A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117826019A (en) * | 2024-03-06 | 2024-04-05 | 国网吉林省电力有限公司长春供电公司 | Line single-phase grounding fault area and type detection method of neutral point ungrounded system |
| CN117909909A (en) * | 2024-03-19 | 2024-04-19 | 青岛鼎信通讯股份有限公司 | Arc grounding fault identification method for power distribution network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | ArcNet: Series AC arc fault detection based on raw current and convolutional neural network | |
| CN116540025A (en) | Fault detection method based on transfer learning and residual error network | |
| Xiong et al. | Incipient fault identification in power distribution systems via human-level concept learning | |
| Baqui et al. | High impedance fault detection methodology using wavelet transform and artificial neural networks | |
| CN100546140C (en) | High impedance fault detects | |
| Michalik et al. | New ANN-based algorithms for detecting HIFs in multigrounded MV networks | |
| US11656263B2 (en) | Effective feature set-based high impedance fault detection | |
| Sekar et al. | A fuzzy rule base approach for High Impedance Fault detection in distribution system using Morphology Gradient filter | |
| Ferreira et al. | Probabilistic transmission line fault diagnosis using autonomous neural models | |
| Sekar et al. | Combined mathematical morphology and data mining based high impedance fault detection | |
| CN114414942A (en) | Power transmission line fault identification classifier, identification method and system based on transient waveform image identification | |
| CN116008731B (en) | Power distribution network high-resistance fault identification method and device and electronic equipment | |
| Wang et al. | Application of extension neural network algorithm and chaos synchronization detection method to partial discharge diagnosis of power capacitor | |
| CN113447766A (en) | Method, device, equipment and storage medium for detecting high-resistance ground fault | |
| Rong et al. | An approach for detecting fault lines in a small current grounding system using fuzzy reasoning spiking neural p systems | |
| Narasimhulu et al. | LWT based ANN with ant lion optimizer for detection and classification of high impedance faults in distribution system | |
| Li et al. | A fault pattern and convolutional neural network based single-phase earth fault identification method for distribution network | |
| Yuan et al. | Faulty feeder detection based on image recognition of voltage-current waveforms in non-effectively grounded distribution networks | |
| Ekici et al. | Comparison of different regression models to estimate fault location on hybrid power systems | |
| Sahoo et al. | A method to detect high impedance faults in distribution feeders | |
| CN116466267A (en) | Distribution network ground fault type identification method based on cyclic neural network | |
| Vahidi et al. | An approach to detection of high impedance fault using discrete wavelet transform and artificial neural networks | |
| CN112710923B (en) | Data-driven single-phase earth fault line selection method based on post-fault steady-state information | |
| Hafidz et al. | Neuro wavelet algortihm for detecting high impedance faults in extra high voltage transmission systems | |
| Alshumayri et al. | Distribution grid fault diagnostic employing Hilbert-Huang transform and neural networks |
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 |