CN116304822A - Transformer state prediction method based on multi-element perception and feature extraction - Google Patents
Transformer state prediction method based on multi-element perception and feature extraction Download PDFInfo
- Publication number
- CN116304822A CN116304822A CN202310267211.0A CN202310267211A CN116304822A CN 116304822 A CN116304822 A CN 116304822A CN 202310267211 A CN202310267211 A CN 202310267211A CN 116304822 A CN116304822 A CN 116304822A
- Authority
- CN
- China
- Prior art keywords
- transformer
- fault
- winding
- characteristic
- state
- 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
- 238000000034 method Methods 0.000 title claims abstract description 66
- 238000000605 extraction Methods 0.000 title claims abstract description 39
- 230000008447 perception Effects 0.000 title claims abstract description 17
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 22
- 230000015654 memory Effects 0.000 claims abstract description 11
- 238000007781 pre-processing Methods 0.000 claims abstract description 9
- 238000004804 winding Methods 0.000 claims description 91
- 239000011159 matrix material Substances 0.000 claims description 20
- 230000008569 process Effects 0.000 claims description 16
- 239000002184 metal Substances 0.000 claims description 15
- 229910052751 metal Inorganic materials 0.000 claims description 15
- 239000013598 vector Substances 0.000 claims description 12
- 238000001514 detection method Methods 0.000 claims description 10
- 230000015556 catabolic process Effects 0.000 claims description 5
- 238000006073 displacement reaction Methods 0.000 claims description 4
- 230000005484 gravity Effects 0.000 claims description 4
- 238000003825 pressing Methods 0.000 claims description 4
- 230000003595 spectral effect Effects 0.000 claims description 4
- 238000012549 training Methods 0.000 description 23
- 239000003921 oil Substances 0.000 description 9
- 230000008901 benefit Effects 0.000 description 7
- XEEYBQQBJWHFJM-UHFFFAOYSA-N Iron Chemical group [Fe] XEEYBQQBJWHFJM-UHFFFAOYSA-N 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 230000006870 function Effects 0.000 description 5
- 238000012795 verification Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 4
- 238000012544 monitoring process Methods 0.000 description 4
- 238000001228 spectrum Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 230000001133 acceleration Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 229910000976 Electrical steel Inorganic materials 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000001816 cooling Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000004868 gas analysis Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000002427 irreversible effect Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000011235 metanalysis Methods 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000012806 monitoring device Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000016273 neuron death Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000004445 quantitative analysis Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 208000024891 symptom Diseases 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
- G01H17/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves, not provided for in the preceding groups
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/761—Proximity, similarity or dissimilarity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Economics (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Medical Informatics (AREA)
- Tourism & Hospitality (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Operations Research (AREA)
- Entrepreneurship & Innovation (AREA)
- Quality & Reliability (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Primary Health Care (AREA)
- Protection Of Transformers (AREA)
Abstract
The invention discloses a transformer state prediction method based on multi-element perception and feature extraction, which comprises the steps of collecting vibration signals on the surface of an oil tank of a transformer in a normal state and a fault state; preprocessing vibration signals and visualizing the vibration signals into gray feature images, storing a part of gray feature images as a fault feature gallery, using the rest gray feature images as a model data set, extracting features based on an algorithm of combining a convolutional neural network and a long-short-time memory network by using the model data set, encoding and decoding original gray feature images by using a 3D convolutional neural network, selecting the feature images of the vibration signals in a fault state as a benchmark, selecting matched decoders from all time steps to decode the output feature images, calculating the spatial similarity between the decoder and the fault feature gallery, classifying the features by using a similarity IoU index, and considering the two to belong to the same feature state when the similarity is larger than a preset threshold value, and outputting a tag of the feature state in the fault feature gallery.
Description
Technical Field
The invention belongs to the technical field of transformer monitoring, and particularly relates to a transformer state prediction method based on multi-element sensing and feature extraction.
Background
The power transformer is an important power device for realizing long-distance transmission and electric energy distribution in the power grid, and the stability of the power transformer is critical to the safety of the power grid. The faults and the running state of the power transformer are affected in various aspects, and the stable running of the transformer is affected by natural environment factors, human factors, defects of the transformer and the matched equipment thereof, and the like. The transformer vibration is caused by vibration of a transformer body (collectively referred to as an iron core, a winding, and the like) and vibration of a cooling device. The vibration of the iron core is mainly caused by the magnetostriction phenomenon of the silicon steel sheet. Vibration of the windings is caused by the existence of a leakage magnetic field when a load current flows through the windings, and dynamic electromagnetic forces are generated among the windings, the wire cakes and the wire turns. The long-term vibration effect causes mechanical defects such as deformation of the internal mechanical structure of the transformer, the state changes are irreversible, and the vibration form correspondingly changes, so that the vibration signal of the transformer contains abundant mechanical fault information. The existing transformer winding mechanical state identification methods comprise a dissolved gas analysis method and a fault sound analysis method, the methods based on electric quantity are relatively mature, but the transformer is required to be taken out of operation, the requirement of on-line monitoring cannot be met, and quantitative analysis of symptoms of faults is difficult. The fault sound analysis method includes an ultrasonic detection method, a noise detection method, a vibration detection method, and the like. The state diagnosis method based on the non-electric quantity is more flexible to use, can monitor signal data under the condition of not influencing the working of the transformer, but the technology lacks theoretical research, has single data type and incomplete state fault data at present, and has great difficulty in practical field application.
The above information disclosed in the background section is only for enhancement of understanding of the background of the invention and therefore may contain information that does not form the prior art that is already known to a person of ordinary skill in the art.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a transformer state prediction method based on multi-element sensing and feature extraction, which can monitor signal data without influencing the working condition of a transformer and has high monitoring precision.
The invention aims at realizing the following technical proposal, the transformer state prediction method based on multi-element perception and feature extraction comprises the following steps,
collecting vibration signals on the surface of an oil tank of the transformer in a normal state and a fault state;
preprocessing the vibration signal and visualizing the vibration signal into gray characteristic images, storing a part of the gray characteristic images as a fault characteristic image library, and using the rest of the gray characteristic images as a model data set, wherein in the preprocessing, the vibration signal is read, characteristic values are calculated and stored as a characteristic value matrix, Z-Score standardization is carried out based on the characteristic value matrix, and the formula (1):
wherein T' is a characteristic value, mu is a sample mean value, sigma is a sample standard deviation, T is a characteristic value after standardized treatment, the standard deviation is subjected to normal distribution with the mean value of 0 and the standard deviation of 1, and gray characteristic images M are constructed by using the maximum value ratio of each channel element and the absolute value and the maximum value ratio of adjacent difference values, and the formula (2):
wherein T is a matrix form of the normalized eigenvalues, the matrix size is L1, L is the number of eigenvalues contained in one image, i, j, a, b ranges from 0 to L, T i For the value of the ith row of the sequence, |T| max Absolute value of maximum value, T i -T j Is the difference between adjacent elements of the sequence, |T a -T b | max The size of the obtained gray image matrix M is L x L, and the characteristic value range is [0,1] for the maximum value of the difference value of any two elements in the sequence, namely the difference between the maximum value and the minimum value]。
Extracting features based on an algorithm of combining a convolutional neural network and a long-short-time memory network by a model data set, encoding and decoding an original gray feature image by using a 3D convolutional neural network, encoding by using an attention module and a dynamic semantic vector sequence instead of a fixed semantic vector as input information in an encoder part, automatically selecting a subset of vectors in the decoding process of a decoder, and quantitatively comparing a feature image decoded and output by the decoder with the gray feature image by using a peak signal-to-noise ratio PSNR and a structural similarity SSIM;
and selecting the matched decoder to decode the output characteristic image from all time steps by taking the characteristic image of the vibration signal in the fault state as a reference, calculating the spatial similarity between the characteristic image and the fault characteristic gallery, classifying the characteristics by using a similarity IoU index, and outputting the label of the characteristic state in the fault characteristic gallery when the characteristic image is larger than a preset threshold value and the characteristic image is considered to belong to the same characteristic state.
In the transformer state prediction method based on multi-element perception and feature extraction, load current of a transformer vibration noise online detection system is respectively adjusted under a normal state and a fault state, 4A load current is used as a starting point, 0.8A is used as a step length, the load current is sequentially adjusted to 8A load current, and vibration signal acquisition is respectively carried out on the surface of an oil tank under the normal state and the fault state to be used as a sample for state prediction.
In the transformer state prediction method based on multi-element sensing and feature extraction, the fault state comprises a loosening fault of an iron core and a winding clamp, a winding warping fault, a winding dislocation fault, a winding bulge fault and a partial discharge fault.
In the transformer state prediction method based on multi-element sensing and feature extraction, winding loosening faults are set for a transformer, a screw rod for fixing the winding is loosened by 3 gears respectively 4 N.m, 8 N.m and 12 N.m by using a torque wrench, and a time domain vibration signal is obtained to be used as a vibration signal under the winding clamping piece loosening faults.
In the transformer state prediction method based on multi-element sensing and feature extraction, winding warping faults are set for a transformer, the winding is deformed by applying pressure to a tap in the winding process of the winding, and a time domain vibration signal is obtained as a vibration signal under the winding warping faults.
In the transformer state prediction method based on multi-element sensing and feature extraction, winding dislocation faults are set for a transformer, and a time domain vibration signal under rated current is obtained as a vibration signal under the winding dislocation faults.
In the transformer state prediction method based on multi-element sensing and feature extraction, winding bulge faults are arranged on a transformer, a medium is added between a high-voltage winding and a low-voltage winding to enable the high-voltage winding to bulge outwards and the low-voltage winding to sink inwards, and meanwhile, the high-voltage winding generates axial displacement to obtain a time domain vibration signal as a vibration signal under the winding bulge faults.
In the transformer state prediction method based on multi-element sensing and feature extraction, partial discharge faults are set for a transformer, metal spikes are set on the high-voltage outlet side of a dry-type transformer, the partial discharge faults of the dry-type transformer are simulated, step-up operation is carried out on the dry-type transformer, and time domain vibration signals from partial discharge to breakdown between the metal spikes and metal clamping pieces are measured in real time to serve as vibration signals under the partial discharge faults.
In the transformer state prediction method based on multi-component perception and feature extraction, the feature values comprise fundamental frequency amplitude, fundamental frequency specific gravity, spectrum energy and spectrum complexity.
In the transformer state prediction method based on multi-element sensing and feature extraction, the formula (3) for classifying the features by using the similarity IoU index is as follows:
wherein, the intersection is the intersection of two image elements, the unit is the union of the two, and IoU is the similarity.
Compared with the prior art, the invention has the following advantages: the transformer state prediction method based on the multi-element sensing and feature extraction reads a plurality of groups of different feature values in the data set to carry out vectorization processing, so as to generate multi-dimensional data. The data is visualized as a gray scale feature image and fed into a convolutional neural network to extract features. The convolution neural network is combined with the long-short-term memory network, the hidden state of the circulation network is expanded to be three-dimensional, and the matrix multiplication operation of the traditional LSTM is replaced by convolution, so that space characteristics can flow among nodes of the circulation network in a three-dimensional tensor form, and the multi-element perception enables ConvLSTM to have strong prediction capability. And then comparing the predicted characteristic image with a fault characteristic gallery in a dividing way, and outputting the expected future state of the transformer by the model. The whole fault prediction flow is convenient and accurate in judgment, on one hand, the state of the transformer equipment is comprehensively depicted from the angle of an image by utilizing a characteristic extraction method-a plurality of groups of characteristic value imaging, the singleness that only a certain characteristic value is analyzed by the traditional method is compensated, and the image output by the model has stronger intuitiveness; on the other hand, the image characteristics are processed through the combined network of the two kinds of neural networks, so that characteristic values of a large number of different types of reaction transformer equipment states can be analyzed, and references are provided for transformer fault pre-judgment and data depth mining of power grid main equipment.
Drawings
Various other advantages and benefits of the present invention will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the invention. It is evident that the figures described below are only some embodiments of the invention, from which other figures can be obtained without inventive effort for a person skilled in the art. Also, like reference numerals are used to designate like parts throughout the figures.
In the drawings:
FIG. 1 is a schematic diagram of a transformer vibration signal acquisition system based on a method for predicting transformer states based on multivariate perception and feature extraction in accordance with one embodiment of the invention;
FIG. 2 is a schematic diagram of a gray feature image of a feature to be extracted of a transformer state prediction method based on multivariate perception and feature extraction according to one embodiment of the invention;
fig. 3 is an illustration of an Encoder-Decoder model based on the Attention mechanism for a transformer state prediction method based on multivariate perception and feature extraction in accordance with one embodiment of the present invention.
The invention is further explained below with reference to the drawings and examples.
Detailed Description
Specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.
It should be noted that certain terms are used throughout the description and claims to refer to particular components. Those of skill in the art will understand that a person may refer to the same component by different names. The specification and claims do not identify differences in terms of components, but rather differences in terms of the functionality of the components. As used throughout the specification and claims, the terms "include" and "comprise" are used in an open-ended fashion, and thus should be interpreted to mean "include, but not limited to. The description hereinafter sets forth a preferred embodiment for practicing the invention, but is not intended to limit the scope of the invention, as the description proceeds with reference to the general principles of the description. The scope of the invention is defined by the appended claims.
For the purpose of facilitating an understanding of the embodiments of the present invention, reference will now be made to the drawings, by way of example, and specific examples of which are illustrated in the accompanying drawings.
For better understanding, as shown in fig. 1 to 3, the transformer state prediction method based on multivariate perception and feature extraction includes:
collecting vibration signals on the surface of an oil tank of the transformer in a normal state and a fault state;
preprocessing the vibration signal and visualizing the vibration signal into gray characteristic images, storing a part of the gray characteristic images as a fault characteristic image library, and using the rest of the gray characteristic images as a model data set, wherein in the preprocessing, the vibration signal is read, characteristic values are calculated and stored as a characteristic value matrix, Z-Score standardization is carried out based on the characteristic value matrix, and the formula (1):
wherein T' is a characteristic value, mu is a sample mean value, sigma is a sample standard deviation, T is a characteristic value after standardized treatment, the standard deviation is subjected to normal distribution with the mean value of 0 and the standard deviation of 1, and gray characteristic images M are constructed by using the maximum value ratio of each channel element and the absolute value and the maximum value ratio of adjacent difference values, and the formula (2):
wherein T is a matrix form of the normalized eigenvalues, namely a complete eigenvalue, the matrix size is L1, L is the number of eigenvalues contained in one image, and i, j, a, b ranges from 0 to L, T i For the value of the ith row of the sequence, |T| max Absolute value of maximum value, T i -T j Is the difference between adjacent elements of the sequence, |T a -T b | max The size of the obtained gray image matrix M is L x L, and the characteristic value range is [0,1] for the maximum value of the difference value of any two elements in the sequence, namely the difference between the maximum value and the minimum value]。
Extracting features based on an algorithm of combining a convolutional neural network and a long-short-time memory network by a model data set, encoding and decoding an original gray feature image by using a 3D convolutional neural network, encoding by using an attention module and a dynamic semantic vector sequence instead of a fixed semantic vector as input information in an encoder part, automatically selecting a subset of vectors in the decoding process of a decoder, and quantitatively comparing a feature image decoded and output by the decoder with the gray feature image by using a peak signal-to-noise ratio PSNR and a structural similarity SSIM;
and selecting the matched decoder to decode the output characteristic image from all time steps by taking the characteristic image of the vibration signal in the fault state as a reference, calculating the spatial similarity between the characteristic image and the fault characteristic gallery, classifying the characteristics by using a similarity IoU index, and outputting the label of the characteristic state in the fault characteristic gallery when the characteristic image is larger than a preset threshold value and the characteristic image is considered to belong to the same characteristic state.
The combination of the convolutional neural network (CNN, convolutional Neural Network) and the long-short-time memory network (LSTM, long Short Term Memory) -ConvLSTM not only maintains the advantages of a fully-connected long-short-time memory network (FC-LSTM), but also has the time sequence modeling capability of LSTM, and is very suitable for predicting spatiotemporal data because of the inherent convolutional structure of the convolutional neural network. The ConvLSTM model is introduced into a coding prediction structure, and a transformer state pre-judging model is established. The algorithm has no fixed requirement on the data quantity, and has the advantages of high learning freedom degree, good robustness, high training speed and high precision.
In a preferred embodiment of the transformer state prediction method based on multi-element sensing and feature extraction, the load current of the transformer vibration noise on-line detection system is respectively adjusted under a normal state and a fault state, the load current of 4A is taken as a starting point, 0.8A is taken as a step length, the load current is sequentially adjusted to 8A, and vibration signal acquisition is respectively carried out on the surface of the oil tank under the normal state and the fault state to be used as a sample for state prediction.
In a preferred embodiment of the transformer state prediction method based on multi-element sensing and feature extraction, the fault state comprises a loosening fault of an iron core and a winding clamp, a winding warping fault, a winding dislocation fault, a winding bulge fault and a partial discharge fault.
In a preferred implementation mode of the transformer state prediction method based on multi-element sensing and feature extraction, winding loosening faults are set for a transformer, a torque wrench is utilized to loosen a screw rod for fixing windings by 3 gears, namely 4 N.m, 8 N.m and 12 N.m, and a time domain vibration signal is obtained to be used as a vibration signal under the winding clamping piece loosening faults.
In the preferred implementation mode of the transformer state prediction method based on multi-element sensing and feature extraction, winding warping faults are set for the transformer, the winding is deformed by applying pressure to a tap in the winding finger process of the winding, and a time domain vibration signal is obtained as a vibration signal under the winding warping faults.
In a preferred embodiment of the transformer state prediction method based on multi-element sensing and feature extraction, winding dislocation faults are set for a transformer, and a time domain vibration signal under rated current is obtained as a vibration signal under the winding dislocation faults.
In the preferred implementation mode of the transformer state prediction method based on multi-element sensing and feature extraction, winding bulge faults are arranged on a transformer, a medium is added between a high-voltage winding and a low-voltage winding to enable the high-voltage winding to bulge outwards and the low-voltage winding to bulge inwards, meanwhile, the high-voltage winding generates axial displacement, and a time domain vibration signal is obtained to serve as a vibration signal under the winding bulge faults.
In the preferred implementation mode of the transformer state prediction method based on multi-element sensing and feature extraction, a partial discharge fault is set for a transformer, a metal spike is set on a high-voltage outlet side of a dry-type transformer, the partial discharge fault of the dry-type transformer is simulated, a boosting operation is carried out on the dry-type transformer, and a time domain vibration signal from partial discharge to breakdown between the metal spike and a metal clamp is measured in real time and is used as a vibration signal under the partial discharge fault.
In a preferred embodiment of the transformer state prediction method based on multivariate sensing and feature extraction, the feature values include fundamental frequency amplitude, fundamental frequency specific gravity, spectral energy and spectral complexity.
In a preferred embodiment of the transformer state prediction method based on multivariate sensing and feature extraction, the formula (3) for classifying the features by using the similarity IoU index is as follows:
wherein, the intersection is the intersection of two image elements, the unit is the union of the two, and IoU is the similarity.
In one embodiment, the transformer vibration noise on-line detection system is used for respectively collecting signals on the surface of the transformer oil tank in a normal state and a fault state, preprocessing the collected typical fault and normal vibration signals of the transformer, processing to obtain a plurality of groups of characteristic value data, visualizing the characteristic value data into a gray characteristic texture image, storing a part of characteristic fault data into a typical fault recognition library, and using the part of characteristic fault data as a model data set mostly, wherein the data ratio for training, verification and testing is 7:2:1. After training, verifying and testing the data set of the previous 16 days, comparing the predicted characteristic image of the model with a gallery, outputting the prejudgment of the state of the transformer on the 17 th day, and taking corresponding maintenance measures by a power supply worker in advance.
Complete signature sequenceThe total 500 sequential characteristic images are provided, the first 16 images in each 17 characteristic sequential images are used as the length of the input characteristic sequence, the last image is the output sequence, and the step length is set to be 8. All the data of training have labels, and then the sequence characteristic image of the model training by the first input sequence is { x } 0 ，x 1 ，...，x 15 The label of the corresponding output characteristic image sequence is { x } 16 Second input sequence feature image { x } 9 ，x 10 ，...，x 24 The label of the corresponding output characteristic image sequence is { x } 25 And so on.
The signal preprocessing is to read the original vibration signal and calculate and store the characteristic value. The characteristic values are selected from fundamental frequency amplitude, fundamental frequency specific gravity, spectrum energy and spectrum complexity which can represent the mechanical state of the transformer.
The characteristic value matrix is subjected to Z-Score standardization, and compared with a min-max normalization method, the method not only can remove dimensions, but also can enable variables of all dimensions to be identical, because each dimension obeys normal distribution with a mean value of 0 and a variance of 1, the data of each dimension plays the same role when the difference value is calculated finally, and the huge influence of selection of different dimensions on distance calculation is avoided, wherein a specific formula is as follows:
wherein T' is the original data, mu is the sample mean value, sigma is the sample standard deviation, T is the characteristic value after standardized treatment, and obeys the normal distribution with the mean value of 0 and the standard deviation of 1. And constructing a characteristic image M by using the maximum value ratio of each channel element and the absolute value and the maximum value ratio of adjacent difference values, for example, (2):
wherein T is a complete feature sequence, the matrix size is l×1, and L is the number of feature values contained in one image. The element value ranges of the obtained image matrix M are all 0 and 1, so that the convergence rate of the neural network can be increased and the occurrence of the over-fitting problem can be avoided while the image can keep as much transformer characteristic information as possible.
In one embodiment, each frame of image is resized to 128 x 128 pixels. The prediction method comprises the following steps:
1) The on-line detection system for the vibration noise of the transformer comprises a vibration acceleration sensor and an acoustic sensor, wherein the vibration acceleration sensor and the acoustic sensor are connected with a monitoring device, and a monitoring terminal is transmitted through remote data transmission.
2) Respectively adjusting load current in a normal state and a winding fault state, taking 4A as a starting point, taking 0.8A as a step length, sequentially adjusting to 8A, respectively acquiring vibration signals on the surface of an oil tank in the above states, and taking the vibration signals as a sample for state prejudgment;
3) The transformer is provided with winding loosening, warping, dislocation and bulge faults, and vibration signals are acquired on the surface of the oil tank under the states;
4) The high-voltage outlet side of the transformer is provided with a metal spike, and the partial discharge fault of the transformer is simulated. Step-up operation is carried out on the transformer, and vibration signals from partial discharge to breakdown between the metal spike and the metal clamp are measured in real time;
5) The collected typical fault and normal vibration signals of the transformer are preprocessed and then visualized as gray characteristic images, as shown in fig. 2.
6) And extracting features by utilizing an algorithm combining the convolutional neural network and the long-short-time memory network, and encoding and decoding the original gray feature image by utilizing the combination of the convolutional neural network based on multi-element perception and the special long-short-time memory network. An Encoder-Decoder model diagram is shown in FIG. 3.
7) The feature images output by decoding are quantitatively compared with the actual feature images, and the performance of the model training is quantitatively evaluated by the evaluation indexes on the data set: peak signal-to-noise ratio (PSNR) and Structural Similarity (SSIM).
8) Using typical fault signature images derived from a large amount of raw data as a reference, matching decoded signature images are selected from all time steps and their spatial similarity to a fault signature gallery is calculated, the similarity IoU index is used to classify the features, a calculation formula such as (3), and the calculation result can indicate which type of state the sample belongs to.
Wherein, the intersection is the intersection of two image elements, and the unit is the union of the two. Setting a threshold value, when the threshold value is larger than the threshold value, the two are considered to belong to the same characteristic state, and the model outputs reference judgment on the state of the transformer on the 17 th day of each time step: normal operation state, loosening fault of iron core and winding clamp, winding dislocation fault, winding warping fault, winding bulge fault and partial discharge fault.
In one embodiment, the transformer vibration noise on-line detection system is used for respectively collecting original vibration signals on the surface of the transformer oil tank in a normal state and a typical fault state, the preprocessed signals are converted into gray feature images, the ConvLSTM algorithm based on multi-element perception is used for extracting texture features of the images, the mean square error loss value and the structural similarity are used for evaluating the judging result, the similarity index is used for classifying the feature images, and the classifying result is output through the model so as to achieve the effect of prejudging the internal mechanical state of the transformer, thereby achieving the purpose of fault prediction, and the specific embodiment comprises the following steps:
s1: under the condition that the internal mechanical state of the transformer is normal, the load current is adjusted to take 0.8A as a step length, the time domain vibration signals of the transformer are respectively obtained under the states of 4A, 4.8A, 5.6A, 6.4A, 7.2A and 8A, and the horizontal distance between the microphone and the surface of the transformer oil tank is 1m;
s2: setting winding loosening faults for the transformer, loosening a screw rod of a fixed winding by using a torque wrench for 3 gears, wherein the gears are respectively 4 N.m, 8 N.m and 12 N.m, and obtaining a time domain vibration signal corresponding to the state;
s3: setting winding warping faults for the transformer, and applying pressure to the tap positions in the winding process of the winding to deform the winding so as to obtain a time domain vibration signal corresponding to the state;
s4: setting winding dislocation faults for the transformer, and obtaining time domain vibration signals of the state under rated current;
s5: arranging winding bulge faults on the transformer, and adding a medium between a high-voltage winding and a low-voltage winding to enable the high-voltage winding to bulge outwards and the low-voltage winding to sink inwards, and simultaneously enabling the high-voltage winding to generate certain axial displacement to acquire a time domain vibration signal corresponding to the state;
s6: the transformer is provided with partial discharge faults, and the high-voltage outlet side of the dry-type transformer is provided with a metal spike to simulate the partial discharge faults of the dry-type transformer. And (3) performing boosting operation on the dry-type transformer, and measuring time domain vibration signals from partial discharge to breakdown between the metal spike and the metal clamp in real time.
S7: the acquired typical fault and normal vibration signals of the transformer are preprocessed by MATLAB, a plurality of groups of characteristic value data are obtained through processing, the characteristic value data are visualized into gray characteristic images, the specific visualization modes refer to (1) and (2), the matrix dimension of the complete characteristic sequence image is 48 x 48, and the quantity L of characteristic values contained in one image is 48 because the time interval of the original time domain data is half an hour, so that the generated image contains characteristics within one day. The resolution of the greyscaled feature map is changed to 128 x 128. And a part of the model data set is stored as a typical fault recognition library for feature comparison with the predicted result of the subsequent model, and the rest of the feature images are correspondingly stored in a training, verifying and testing folder to be used as a model data set so as to generate a file path, and a training, verifying and testing model is generated.
S8: and extracting features by utilizing an algorithm combining the convolutional neural network and the long-short-time memory network, and encoding and decoding the original gray feature image by using the 3D convolutional neural network. In the encoder part, an attention module is used, a dynamic semantic vector sequence is used as input information for encoding instead of a fixed semantic vector, and a subset of vectors is automatically selected during decoding by a decoder. This way it is achieved that different weight coefficients are assigned due to the difference in the contribution of the input sequence data to the output predicted sequence, which is the main idea of the Attention mechanism.
S9: the feature images output by decoding are quantitatively compared with the actual feature images by using peak signal to noise ratio (PSNR) and Structural Similarity (SSIM), the PSNR evaluates the pixel difference between two frames of images like MSE, and the higher the PSNR and the SSIM, the better the prediction effect of the model is, and the more accurate the result is compared with the actual feature images.
S10: the method takes typical fault characteristic images derived from a large amount of original data as a reference, selects matched decoding characteristic images from all time steps, calculates the spatial similarity between the decoding characteristic images and a fault characteristic gallery, classifies the characteristics by using a similarity IoU index, and can represent which type of state the sample belongs to. When the threshold is set to be 0.5, the model outputs a label of the state in the gallery, namely, the reference judgment of the state of the transformer on the 17 th day of each time step: normal operation state, loosening fault of iron core and winding clamp, winding dislocation fault, winding warping fault, winding bulge fault and partial discharge fault.
In one embodiment, the ConvLSTM network combines a Convolutional Neural Network (CNN) and a long and short time memory network (LSTM), and the main structure of the ConvLSTM network comprises an input layer, a convolutional layer, a pooling layer, a full connection layer and an output layer. The convolutional neural network part of the model uses a two-dimensional convolution method, and an input layer image is an RGB image, so that in channels (the number of input data channels) in the first convolution is 3, out channels (the number of output data channels) is 32, the convolution kernel size is 4, the step size is 2, the padding value is 1 (the number of zero padding layers around the image in the convolution process), and the output data size is 64×64×32. The out channels of the first convolution are in channels of the second convolution, the out channels of the second convolution are 64, and the remaining parameters are unchanged, so that the output size of the second convolution is 32×32×64. And so on, undergoing 7 two-dimensional convolutions, the final output data size is 1 x 1024. The output data has the height and width of 1, so that the additional pooling is not needed to be used as characteristic dimension reduction. The activation function used in the whole process is a leak ReLU function, which can solve the problem of neuronal death, and compared with the ReLU activation function, the leak ReLU is different in that the input is less than 0, the value is negative, and the gradient is tiny. The data processed by the convolutional neural network is input into the LSTM network, the forgetting gate, the input gate and the output gate in the LSTM network adjust the parameters of the data through continuous iterative training of a large amount of data, so that the data can learn the time fitting relation among the data in the data information extracted from the convolutional neural network, the input and output data of the predicted time sequence are effectively and dynamically modeled, finally the trained data are fitted through the ConvLSTM network, the predicted image is output through the neurons of the full-connection layer, and the network can circulate for 16 times and then output a preset result because SEQ_SIZE is 16. The whole prediction process firstly inputs data to carry out model training, and determines model parameters. Training and verifying: initializing a model structure and an optimizer, if a GPU exists, accelerating loading of the model by using cuda, selecting a MSE loss function as a loss function, and reading training data sets into input and tag variables one by one. And (3) training round by round according to the set value of the EPOCH (iterative times of training), inputting input data into a model to be transmitted forward, performing error calculation, gradient zero clearing, and updating parameters after backward transmitting errors. And (3) performing verification once every five times of training, reading the verification data set into the input and label variables one by one in the verification process, performing the rest training processes, and finally storing the model. The epoch and loss values are printed row by row in the training and verifying process, the change of the loss value along with the increase of the epoch in each cycle can be seen in real time, the value is allowed to fluctuate in a small range, and the current training and verifying process is considered to be effective as the current training and verifying process is basically and continuously reduced. The initial iteration number EPOCH is set to 300, the learning rate is set to one thousandth, after training and verification are completed, a loss curve graph is opened, when the curve tends to be stable in the iteration number, if the loss values of the training set and the test set are very close, the current parameters are proper, and adjustment is not needed. If the training set performs far better than the test set, such overfitting, then data enhancement, adding data sets, etc. can be considered for suppression. Otherwise, the under fitting phenomenon occurs, and the increase of the network depth and the number of units of the hidden layer can be considered. As for the parameters which are continuously accumulated, updated and cleared in the training process in other models, the parameters can be saved along with the saving of the models, and manual adjustment is not needed.
The prediction method adopts a ConvLSTM algorithm based on multi-element perception to extract the characteristics of gray images, evaluates the prediction result by using a mean square error loss value and structural similarity, and evaluates the classification result by using IoU. The method has the advantages that the original vibration signals are obtained by the transformer vibration noise on-line detection system, the preprocessed signals are converted into gray feature images, the multi-element perceived ConvLSTM model is utilized to extract texture features of the images, the mean square error loss value and the structural similarity are utilized to evaluate judging results, then the similarity index is utilized to classify the feature images, and the classifying results are output through the model so as to achieve the effect of prejudging the internal mechanical state of the transformer. Further, the prediction algorithm adopts DDPAE, DFN, CDNA, FRNN, VPN Baseline, predRNN, predRNN +, MIM, BP-Net, E3D-LSTM and other prediction algorithms.
Although the embodiments of the present invention have been described above with reference to the accompanying drawings, the present invention is not limited to the above-described specific embodiments and application fields, and the above-described specific embodiments are merely illustrative, and not restrictive. Those skilled in the art, having the benefit of this disclosure, may effect numerous forms of the invention without departing from the scope of the invention as claimed.
Claims (10)
1. A transformer state prediction method based on multi-element perception and feature extraction is characterized by comprising the following steps of,
collecting vibration signals on the surface of an oil tank of the transformer in a normal state and a fault state;
preprocessing the vibration signal and visualizing the vibration signal into gray characteristic images, storing a part of the gray characteristic images as a fault characteristic image library, and using the rest of the gray characteristic images as a model data set, wherein in the preprocessing, the vibration signal is read, characteristic values are calculated and stored as a characteristic value matrix, Z-Score standardization is carried out based on the characteristic value matrix, and the formula (1):
wherein T' is a characteristic value, mu is a sample mean value, sigma is a sample standard deviation, T is a characteristic value after standardized treatment, the standard deviation is subjected to normal distribution with the mean value of 0 and the standard deviation of 1, and gray characteristic images M are constructed by using the maximum value ratio of each channel element and the absolute value and the maximum value ratio of adjacent difference values, and the formula (2):
wherein T is a matrix form of the normalized eigenvalues, the matrix size is L1, L is the number of eigenvalues contained in one image, i, j, a, b ranges from 0 to L, T i For the value of the ith row of the sequence, |T| max Absolute value of maximum value, T i -T j Is the difference between adjacent elements of the sequence, |T a -T b | max The size of the obtained gray image matrix M is L x L, and the characteristic value range is [0,1] for the maximum value of the difference value of any two elements in the sequence, namely the difference between the maximum value and the minimum value]；
Extracting features based on an algorithm of combining a convolutional neural network and a long-short-time memory network by a model data set, encoding and decoding gray feature images by using a 3D convolutional neural network, encoding by using an attention module and a dynamic semantic vector sequence instead of a fixed semantic vector as input information in an encoder part, automatically selecting a subset of vectors in the decoding process of a decoder, and quantitatively comparing the feature images decoded and output by the decoder with the gray feature images by using peak signal-to-noise ratio PSNR and structural similarity SSIM;
and selecting the matched decoder to decode the output characteristic image from all time steps by taking the characteristic image of the vibration signal in the fault state as a reference, calculating the spatial similarity between the characteristic image and the fault characteristic gallery, classifying the characteristics by using a similarity IoU index, and outputting the label of the characteristic state in the fault characteristic gallery when the characteristic image is larger than a preset threshold value and the characteristic image is considered to belong to the same characteristic state.
2. The transformer state prediction method based on multi-component sensing and feature extraction according to claim 1, wherein preferably, the load current of the transformer vibration noise on-line detection system is adjusted under the normal state and the fault state respectively, the load current of 4A is taken as a starting point, 0.8A is taken as a step length, the load current is sequentially adjusted to 8A, and vibration signal acquisition is carried out on the surface of the oil tank under the normal state and the fault state respectively to be taken as a sample for state prediction.
3. The method of claim 1, wherein the fault conditions include core and winding clip loosening faults, winding buckling faults, winding dislocation faults, winding bulge faults, and partial discharge faults.
4. The transformer state prediction method based on multi-component sensing and feature extraction according to claim 3, wherein winding loosening faults are set for the transformer, a torque wrench is utilized to loosen a screw rod of a fixed winding by 3 gears, namely 4N-m, 8N-m and 12N-m, respectively, and a time domain vibration signal is obtained as a vibration signal under the winding clamping piece loosening faults.
5. The transformer state prediction method based on multi-component sensing and feature extraction according to claim 3, wherein a winding warping fault is set for the transformer, the winding is deformed by applying pressure to a tap in the winding finger process of the winding, and a time domain vibration signal is obtained as a vibration signal under the winding warping fault.
6. The method for predicting the state of a transformer based on multivariate sensing and feature extraction according to claim 3, wherein a winding dislocation fault is set for the transformer, and a time domain vibration signal under rated current is obtained as the vibration signal under the winding dislocation fault.
7. The transformer state prediction method based on multi-component sensing and feature extraction according to claim 3, wherein a winding bulge fault is set for the transformer, a medium is added between a high-voltage winding and a low-voltage winding to enable the high-voltage winding to bulge outwards and the low-voltage winding to bulge inwards, and meanwhile the high-voltage winding generates axial displacement, so that a time domain vibration signal is obtained as a vibration signal under the winding bulge fault.
8. The transformer state prediction method based on multi-element sensing and feature extraction according to claim 3, wherein a partial discharge fault is set for the transformer, a metal spike is set on a high-voltage outlet side of the dry-type transformer, the partial discharge fault of the dry-type transformer is simulated, a step-up operation is performed on the dry-type transformer, and a time domain vibration signal from partial discharge to breakdown between the metal spike and a metal clamping piece is measured in real time as a vibration signal under the partial discharge fault.
9. The transformer state prediction method based on multivariate perception and feature extraction of claim 1, wherein the feature values comprise fundamental frequency amplitude, fundamental frequency specific gravity, spectral energy, and spectral complexity.
10. The transformer state prediction method based on multivariate perception and feature extraction according to claim 1, wherein the formula (3) for classifying the features using the similarity IoU index is:
wherein, the intersection is the intersection of two image elements, the unit is the union of the two, and IoU is the similarity.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310267211.0A CN116304822A (en) | 2023-03-17 | 2023-03-17 | Transformer state prediction method based on multi-element perception and feature extraction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310267211.0A CN116304822A (en) | 2023-03-17 | 2023-03-17 | Transformer state prediction method based on multi-element perception and feature extraction |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116304822A true CN116304822A (en) | 2023-06-23 |
Family
ID=86779336
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310267211.0A Pending CN116304822A (en) | 2023-03-17 | 2023-03-17 | Transformer state prediction method based on multi-element perception and feature extraction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116304822A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116402584A (en) * | 2023-04-23 | 2023-07-07 | 中航信移动科技有限公司 | Event generation method based on multiple data sources, storage medium and electronic equipment |
-
2023
- 2023-03-17 CN CN202310267211.0A patent/CN116304822A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116402584A (en) * | 2023-04-23 | 2023-07-07 | 中航信移动科技有限公司 | Event generation method based on multiple data sources, storage medium and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111751133B (en) | Intelligent fault diagnosis method of deep convolutional neural network model based on non-local mean embedding | |
| CN109884419B (en) | Smart power grid power quality online fault diagnosis method | |
| KR102094320B1 (en) | Method for improving image using reinforcement learning | |
| CN109389171B (en) | Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology | |
| CN111879397B (en) | Fault diagnosis method for energy storage mechanism of high-voltage circuit breaker | |
| CN112818764B (en) | Low-resolution image facial expression recognition method based on feature reconstruction model | |
| CN114048769A (en) | Multi-source multi-domain information entropy fusion and model self-optimization method for bearing fault diagnosis | |
| CN113865868B (en) | Rolling bearing fault diagnosis method based on time-frequency domain expression | |
| CN116304822A (en) | Transformer state prediction method based on multi-element perception and feature extraction | |
| CN113076920B (en) | Intelligent fault diagnosis method based on asymmetric domain confrontation self-adaptive model | |
| CN114821155A (en) | Multi-label classification method and system based on deformable NTS-NET neural network | |
| CN112305379A (en) | Mode identification method and system for GIS insulation defect | |
| CN110726898A (en) | Power distribution network fault type identification method | |
| CN111915101A (en) | Complex equipment fault prediction method and system based on LPP-HMM method | |
| CN111179235A (en) | Image detection model generation method and device, and application method and device | |
| CN117034003A (en) | Full life cycle self-adaptive fault diagnosis method, system, equipment and medium for aerospace major product manufacturing equipment | |
| CN113884844A (en) | Transformer partial discharge type identification method and system | |
| CN115290326A (en) | Rolling bearing fault intelligent diagnosis method | |
| CN115859142A (en) | Small sample rolling bearing fault diagnosis method based on convolution transformer generation countermeasure network | |
| CN117009916A (en) | Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning | |
| CN113295413B (en) | Traction motor bearing fault diagnosis method based on indirect signals | |
| CN114676733A (en) | Fault diagnosis method for complex supply and delivery mechanism based on sparse self-coding assisted classification generation type countermeasure network | |
| CN114639000A (en) | Small sample learning method and device based on cross-sample attention aggregation | |
| CN117219124A (en) | Switch cabinet voiceprint fault detection method based on deep neural network | |
| CN111025100A (en) | Transformer ultrahigh frequency partial discharge signal mode identification method and device |
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 |