CN112763943A - Diagnosis system and method for transformer winding fault classification and positioning - Google Patents
Diagnosis system and method for transformer winding fault classification and positioning Download PDFInfo
- Publication number
- CN112763943A CN112763943A CN202011429110.1A CN202011429110A CN112763943A CN 112763943 A CN112763943 A CN 112763943A CN 202011429110 A CN202011429110 A CN 202011429110A CN 112763943 A CN112763943 A CN 112763943A
- Authority
- CN
- China
- Prior art keywords
- data
- transformer winding
- fault
- frequency
- model
- 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.)
- Granted
Links
Images
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
- G01R31/50—Testing of electric apparatus, lines, cables or components for short-circuits, continuity, leakage current or incorrect line connections
- G01R31/72—Testing of electric windings
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/50—Testing of electric apparatus, lines, cables or components for short-circuits, continuity, leakage current or incorrect line connections
- G01R31/52—Testing for short-circuits, leakage current or ground faults
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/50—Testing of electric apparatus, lines, cables or components for short-circuits, continuity, leakage current or incorrect line connections
- G01R31/62—Testing of transformers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Power Engineering (AREA)
- Locating Faults (AREA)
Abstract
变压器绕组故障分类与定位的诊断系统及方法，该系统包括扫频仪、变压器绕组、检测终端、电感L1、分压电阻R1、分压电阻R2、电容C。扫频仪输出扫频信号，经过电感L1、分压电阻R1从接地点N输入到变压器绕组上，流经变压器绕组内部的高压绕组A并输出，通过电容C并被分压电阻R2分压后，被检测终端采集扫频数据；检测终端通过构建的深度学习卷积神经网络模型诊断出变压器绕组的故障类型与位置。本发明在实现故障类型诊断高精度的基础上，解决了故障定位问题，具有一定工程意义。在本发明中构建的仿真样本上效果良好，故障类型诊断精度高达100%，故障定位正确率高达83.33%。
A diagnostic system and method for classifying and locating transformer winding faults, the system includes a frequency sweeper, transformer windings, a detection terminal, an inductor L1, a voltage dividing resistor R1, a voltage dividing resistor R2, and a capacitor C. The frequency sweeper outputs the sweep frequency signal, which is input from the grounding point N to the transformer winding through the inductor L1 and the voltage dividing resistor R1, flows through the high-voltage winding A inside the transformer winding and outputs, and passes through the capacitor C and is divided by the voltage dividing resistor R2. , the detected terminal collects the frequency sweep data; the detection terminal diagnoses the fault type and location of the transformer winding through the constructed deep learning convolutional neural network model. The invention solves the problem of fault location on the basis of realizing high-precision fault type diagnosis, and has certain engineering significance. The simulation sample constructed in the present invention has a good effect, the fault type diagnosis accuracy is as high as 100%, and the fault location accuracy rate is as high as 83.33%.
Description
Technical Field
The invention relates to the field of fault diagnosis of electrical equipment, in particular to a transformer winding fault classification and positioning diagnosis system and method.
Background
The transformer winding fault is an internal fault and belongs to an accumulation type fault. The frequency response method has a high sensitivity characteristic to winding faults. Compared with other detection means, the method can find the fault more timely and effectively, prevent undetected micro deformation from gradually developing into the fault with larger deformation degree, and threaten the operation of the power grid, thereby gaining wide attention. In the existing diagnostic methods, a large amount of research has been conducted on analyzing a single-fault frequency sweep curve and changes in impedance values by using a frequency sweep impedance method. However, the impedance value does not vary monotonically with the fault location. The fault point cannot be judged more accurately only by taking the impedance value as a criterion. The research on the positioning of the transformer winding based on the sweep frequency impedance method fails to provide a specific diagnosis method based on an intelligent algorithm and expands the diagnosis method to the positioning of various faults.
Most of the existing frequency response intelligent diagnosis methods extract characteristic information features based on data or images. Most of the features extracted from the data are based on statistical methods, and the types of the features are few, generally less than 10. And the image extraction method is more. And visualizing the frequency, amplitude and phase information data, and extracting computer graphic features by using rectangular coordinate images, three-dimensional curve images or polar coordinate images of phase-amplitude values. In the image-based feature extraction method, the feature amount is often large. Due to the fact that data dimensionality is rich, the winding condition can be represented by means of frequency response data, and good representation is achieved on a proper intelligent algorithm model.
The artificial intelligence algorithm is various, the current common machine learning algorithm for fault type diagnosis comprises a decision tree model, a support vector machine, a random forest and the like, and the current emerging integrated algorithm such as XGboost, Adaboost and the like has excellent performance in winding diagnosis. The research on fault degree detection and fault location relative to fault type diagnosis is slightly insufficient, and especially the fault location problem has high dependence on a detection method. The detection and fault location of the fault degree is deeper information mining after fault category diagnosis, and is limited by the complexity of a feature extraction method and hyper-parameter adjustment, and the phenomena of overfitting and low precision are presented in a detection result.
At present, a fault degree detection method basically depends on an algorithm model optimized by an optimization algorithm, such as a support vector machine optimized by a particle swarm algorithm, a random forest optimized by an ant colony algorithm and the like, however, the model has limited learning capability and insufficient feature extraction, so that the result difference is strong. Therefore, a good quantification method and a deep learning algorithm with a stronger learning ability are needed to solve the problem of detecting and locating the fault degree.
Disclosure of Invention
Aiming at the technical problem, the invention provides a transformer winding fault classification and positioning diagnosis system and a method thereof, wherein the diagnosis system enables a transformer winding model to be asymmetric relative to a sweep generator and acquisition equipment by adding an external circuit element, so that fault positions can be distinguished; the diagnosis method destroys the symmetry of the transformer winding model to the detection end, so that more accurate fault location can be realized, and the model of fault diagnosis and fault location adopts a deep learning convolution neural network model. The invention has good fault classification and positioning problem diagnosis effects.
The technical scheme adopted by the invention is as follows:
a diagnostic system for transformer winding fault classification and localization, the system comprising:
the device comprises a sweep generator, a transformer winding and a detection terminal;
the sweep generator is connected with one end of an inductor L1, the other end of the inductor L1 is connected with one end of a divider resistor R1, and the other end of the divider resistor R1 is connected with a grounding point N of a transformer winding;
one end of a high-voltage winding A of the transformer winding is connected with a grounding point N, the other end of the high-voltage winding A of the transformer winding is connected with one end of a capacitor C, the other end of the capacitor C is respectively connected with one end of a divider resistor R2 and a detection terminal, and the other end of the divider resistor R2 is grounded.
A diagnosis method for classifying and positioning faults of a transformer winding is characterized in that a sweep frequency signal is output by a sweep frequency instrument, is input to the transformer winding from a grounding point N through an inductor L1 and a divider resistor R1, flows through a high-voltage winding A in the transformer winding and is output, and sweep frequency data are acquired by a detected terminal after being divided by a capacitor C and a divider resistor R2; and the detection terminal diagnoses the fault type and position of the transformer winding through the constructed deep learning convolutional neural network model.
The transformer winding fault classification and positioning diagnosis method comprises the following steps:
s1, the detection terminal subtracts the collected sweep frequency data from the frequency response data stored in the detection terminal under the normal working condition;
and S2, recombining the frequency data, the amplitude data and the phase data slices processed in the S1, cutting off head and tail data, and keeping the length-width ratio of a recombined data matrix as close to 1 as possible: 1;
s3, reshaping the size of the data matrix by utilizing an interpolation up-sampling or interpolation down-sampling method for the recombined data matrix of S2;
and S4, putting the processed data in the S3 into a trained fault classification model and a fault positioning model to obtain a diagnosis result.
The invention relates to a diagnosis system and a diagnosis method for transformer winding fault classification and positioning, which have the following technical effects:
1) the external symmetrical structure of the transformer winding is destroyed by adding the capacitance and inductance elements on the external circuit, so that a frequency response method based means for realizing fault location is provided while fault type diagnosis is not hindered. Compared with the traditional method, the obtained amplitude data has a slight difference on the curve, but the general trend is consistent.
2) The invention adopts a special method to extract the frequency response data information. Normal data is subtracted from fault data to reduce data complexity, by slicing and regrouping the data. The size of the matrix is reshaped by an interpolation up-sampling method and a interpolation down-sampling method, and the data of three characteristics of frequency, amplitude and phase are converted into a square matrix. During model training, the data accuracy after complexity reduction is found to be improved by about 10% compared with the former.
3) And on the aspects of fault type diagnosis and fault positioning, the invention adopts the convolutional neural network to train data. Compared with a long-short term memory neural network which is good at time sequence data processing, the convolutional neural network has different reading modes. As for the simulation data set of the invention, the accuracy of the convolutional neural network is improved by about 15 percent on the fault positioning problem and the accuracy of the fault type diagnosis problem is improved by about 5 percent by using the longer short-term memory neural network.
4) The invention solves the problem of fault location on the basis of realizing high precision of fault type diagnosis, and has certain engineering significance. The simulation sample constructed in the invention has good effect, the fault type diagnosis precision is as high as 100%, and the fault positioning accuracy is as high as 83.33%. The improved two models have stable effect.
Drawings
Fig. 1 is a schematic diagram of the fault diagnosis actual measurement of a transformer winding.
FIG. 2(1) is a graph comparing the amplitude-frequency curves of the conventional and improved external circuits;
FIG. 2(2) is a graph comparing the phase-frequency curves of the conventional circuit and the improved external circuit.
FIG. 3(1) is a radial inner concave amplitude-frequency curve diagram of the conventional head and tail symmetric points;
FIG. 3(1-1) is a partial schematic enlarged view I of FIG. 3 (1);
fig. 3(1-2) is a partially enlarged schematic view of fig. 3 (1).
FIG. 3(2) is a conventional radial-concave phase-frequency curve diagram of the symmetrical points at the head and tail portions;
FIG. 3(2-1) is a partial enlarged view of the first part of FIG. 3 (2);
fig. 3(2-2) is a partial enlarged schematic view of fig. 3 (2).
FIG. 3(3) is a radial inner concave amplitude-frequency curve diagram of the improved head and tail symmetrical points;
FIG. 3(3-1) is a partial enlarged view of the first part of FIG. 3 (3);
fig. 3(3-2) is a partial enlarged schematic view of fig. 3 (3).
FIG. 3(4) is a radial inner concave amplitude-frequency curve diagram of the symmetrical points of the head and tail parts after improvement.
FIG. 3(4-1) is a partial enlarged view of the first part of FIG. 3 (4);
fig. 3(4-2) is a partial enlarged schematic view of fig. 3 (4).
FIG. 4 is a fault classification model learning graph.
Fig. 5 is a fault localization model learning graph.
FIG. 6 is a schematic representation of data slice recombination.
Fig. 7 is a schematic diagram of the up-sampling principle.
Fig. 8(1) is a matrix diagram obtained by extracting turn-to-turn short circuit fault data from a total sample and performing data processing;
fig. 8(2) is a matrix diagram obtained by extracting coil strand breakage fault data from a total sample and processing the data;
fig. 8(3) is a matrix diagram obtained by extracting metal foreign matter fault data from a total sample and processing the data;
FIG. 8(4) is a matrix diagram of axial distortion fault data extracted from the total sample after data processing;
FIG. 8(5) is a matrix diagram of the radial concave fault data extracted from the total sample after data processing;
fig. 8(6) is a matrix diagram obtained by extracting the radially outer concave fault data from the total sample and processing the data.
Fig. 9(1) is a header frequency response data diagram of the nine-stage winding model after being processed;
fig. 9(2) is a middle frequency response data diagram of the nine-stage winding model after being processed;
fig. 9(3) is a data diagram of the tail frequency response of the nine-stage winding model after processing.
Detailed Description
As shown in fig. 1, in the transformer winding fault classification and positioning diagnostic system, functions of creating, triggering, timing, reading, writing and the like of a virtual channel of an NI-DAQ toolbox in labview software are called to realize the construction of a data acquisition system. The system comprises the following module parts:
the device comprises a sweep generator 1, a transformer winding 2 and a detection terminal 3;
the sweep generator 1 is connected with one end of an inductor L1, the other end of the inductor L1 is connected with one end of a voltage dividing resistor R1, and the other end of the voltage dividing resistor R1 is connected with a grounding point N of a transformer winding.
One end of a high-voltage winding A of the transformer winding 2 is connected with a grounding point N, the other end of the high-voltage winding A of the transformer winding 2 is connected with one end of a capacitor C, the other end of the capacitor C is respectively connected with one end of a divider resistor R2 and the detection terminal 3, and the other end of the divider resistor R2 is grounded.
The detection terminal 3 is a portable computer equipped with an acquisition card.
The frequency scanner 1 adopts a homonymous TH 1312-6060W frequency scanner.
Inductance L1: 36.9 mH.
Voltage-dividing resistance R1: 50 omega.
Voltage-dividing resistance R2: 50 omega.
A capacitance C: 21.8 pF.
The transformer winding 2 comprises a high-voltage winding A, a high-voltage winding B and a high-voltage winding C; the method of the invention is used for sequentially measuring each single winding.
The grounding point N is specifically connected with the high-voltage winding A in a mode that: and in the Y-Y connection mode, one end of the high-voltage winding A is grounded, namely is connected with a grounding point N, and the other end of the high-voltage winding A is connected with an external circuit. The side of the high-voltage winding A connected with an external circuit is connected with a capacitor C.
A diagnosis method for classifying and positioning faults of a transformer winding is characterized in that a sweep frequency signal is output by a sweep frequency instrument 1, is input to a transformer winding 2 from a grounding point N through an inductor L1 and a voltage dividing resistor R1, flows through a high-voltage winding A in the transformer winding 2 and is output, and is subjected to voltage division by a voltage dividing resistor R2 through a capacitor C, and sweep frequency data are collected by a detected terminal 3; and the detection terminal 3 diagnoses the fault type and position of the transformer winding 2 through the constructed deep learning convolutional neural network model.
The transformer winding fault classification and positioning diagnosis method comprises the following steps:
and S1, the detection terminal 3 subtracts the collected sweep frequency data from the frequency response data stored in the detection terminal under the normal working condition. The frequency response data subtraction processing refers to correspondingly subtracting frequency, amplitude and phase data in the sweep frequency data from data stored in the detection terminal 2 under a normal working condition.
And S2, recombining the frequency data, the amplitude data and the phase data slices processed in the S1, cutting off head and tail data, and keeping the length-width ratio of a recombined data matrix as close to 1 as possible: 1.
as shown in fig. 6, the slicing process removes the most front and the most rear data in the frequency response data, so that the dimension of the matrix can be set more conveniently in the subsequent recombination process. Taking the data of the present invention as an example, three sampling points at the end of the 3 × 2003 data are removed to obtain 3 × 2000 data. Recombination procedure the sliced data were aligned to approximate 1: the data matrix is resized in the manner of 1 gauge. In this patent, the setting is for a size of 80 × 75, and when the data ratio is close to 1: 1, the data is not easily distorted during the interpolation sampling process.
And S3, reshaping the size of the data matrix by utilizing an interpolation up-sampling or interpolation down-sampling method for the recombined data matrix of S2. The method has the specific significance that the data after subtraction shows the data change trend and greatly reduces the complexity of the data, a large amount of data with the value of 0 or smaller and data with larger amplitude of the changed salient part form sharp contrast, and the mode is reducedDifficulty in training. Slice recombination keeps equal scale scaling during interpolation up and down sampling as much as possible, and distortion of matrix data during sampling is reduced. Upsampling principle as shown in fig. 7, the left graph of fig. 7 is upsampled as the right graph: m, n, p, q are calculated by averaging adjacent top, bottom, left and right known points to calculate m as, for example, the formula:
the calculation method corresponding to x is as the formula:
the specific operation method can be according to an API in the cv2 library of python or a matlab function: resize () implementation. The invention reshapes the 80 x 75 data matrix into 128 x 128 dimensions using an interpolation upsampling method.
And S4, putting the processed data in the S3 into a trained fault classification model and a fault positioning model to obtain a diagnosis result. The fault classification model and the fault positioning model are both realized by adopting a convolutional neural network model in a deep learning algorithm. The input to the model is single-channel matrix data. The model is adjusted and the generalization capability of the model is enhanced by changing the number and the size of convolution kernels, the types of activation functions, the pooling mode, the regularization mode and the coefficient, the type of an optimizer and the like.
The improvement method of the external circuit comprises the following steps: an inductor L1 is connected in series at the signal injection end of the detection external circuit, and a capacitor C is connected in series at the output end. By changing the symmetrical structure of the transformer winding 2 to the sweep frequency signal output end in this way, the difference of frequency response curves can be amplified, and faults of symmetrical points can be distinguished.
Taking the normal working condition of the winding as an example:
the amplitude-frequency and phase-frequency curves before and after the improvement are shown in fig. 2(1) and fig. 2 (2). The improved detection method is basically consistent with the trend before improvement on an amplitude-frequency curve and a phase-frequency curve, and the amplitude of the low-frequency band in the amplitude-frequency curve is slightly reduced.
Taking radial inward recess of the winding as an example:
the amplitude-frequency and phase-frequency curves before modification are shown in fig. 3(1) and fig. 3(2), and the amplitude-frequency and phase-frequency curves after modification are shown in fig. 3(3) and fig. 3 (3).
The enlarged partial views of the two end ranges of fig. 3(1) and 3(2) are shown in fig. 3(1-1), 3(1-2), 3(2-1) and 3 (2-2). The conventional methods of fig. 3(1), 3(2) have a problem that, for a fault at a symmetrical point, data are highly overlapped and the fault position cannot be identified no matter on the amplitude frequency curve or the phase frequency curve.
The improved method in fig. 3(3), 3(3) and 3(3) has a shift in the low frequency band on the amplitude frequency and phase frequency curves, and the data are no longer highly overlapped. Although the fault curves on the head and tail symmetrical points cannot be completely overlapped due to the process problem in practice, the method can still amplify the data difference and solve the diagnosis problem of faults of more tiny degree.
In the transformer winding classification and positioning diagnosis method, training set and test set fault samples are constructed through a winding lumped parameter model, and 108 groups are obtained in total, so that the method has sufficiency. The sample balance is considered in the division of the training set and the test set, namely the number of the fault type samples in the training set and the test set is equal, the number of the fault position samples in the training set and the test set is also equal, and the training set and the test set respectively account for half of the total data set. The frequency of the sweep frequency signal is from 3K to 300KHz, and the amplitude-frequency calculation method is shown as a formula (1). And Uo and Ui are respectively input, output and measured voltage, and amplitude is acquired in a logarithmic mode with the base 10.
The method for analyzing and processing data in the algorithm comprises the following steps: the data of the acquired frequency, amplitude and phase are subtracted from the data under the normal working condition, so that the data complexity can be reduced while the curve change is highlighted, and the model training is facilitated. And (3) recombining the processed data slices: and removing unimportant parts from the beginning and the end of the sampling points, and recombining the frequency, the amplitude and the phase into a matrix close to a square. The matrix has approximate row and column sizes, and the number of columns should be a multiple of 3, which ensures that the data is not distorted too much during the interpolation sampling process and does not generate redundant data deviating from the sample. And finally, remolding the model into a square by utilizing interpolation up-sampling and interpolation down-sampling. Taking the data of the invention as an example, 2003 data of amplitude, frequency and phase are subtracted from the data of normal working condition, and then the data are sliced and recombined into a matrix of 80 × 75. Upsampling by interpolation becomes a 128 x 128 square matrix.
The fault classification is realized by adopting a convolutional neural network model in a deep learning algorithm, a learning curve of the model on a data set is shown in fig. 4, and an overview of the model is shown in table 1.
TABLE 1 overview of the Fault Classification model
Adjustable weight number of output tensor dimensionality of neural network layer
The fault classification model comprises 10 layers which are respectively as follows: the device comprises a convolution layer 1, an activation layer 1, a pooling layer 1, a convolution layer 2, an activation layer 2, a pooling layer 2, a flattening layer, a full-connection layer 1, a full-connection layer 2 and an output layer. The input is 128 x 128 single channel matrix data. Convolutional layers 1, 2 consist of 32, 6 × 6 and 64, 3 × 3 convolutional kernels, respectively; the activation layers 1, 2 use ReLU as the activation function. In a deep network, the activation function can effectively inhibit gradient disappearance. The pooling layers 1 and 2 are maximum pooling and mean pooling, respectively. The 64 images of 31 x 31 size obtained in the pooling layer 2 are flattened by the flattening layer and input into the fully-connected layer 1. The fully-connected layer 1 consists of 64 neurons, overfitting is inhibited by adopting an L1 regularization mode, and the ReLU is still adopted as an activation function. The fully-connected layer 2 is not regularized, and the rest is identical to the fully-connected layer 1. The sample contains 6 types of faults in total, the output layer is composed of 6 neurons with an activation function of Softmax, and each neuron outputs a corresponding fault probability value. From the learning curve of fig. 4, it can be observed that the model training process is stable, and the accuracy of the model on both the training set and the test set is 100%.
The types of faults are: turn-to-turn short circuit, coil strand breakage, metal foreign matter, axial distortion, radial inward concavity and radial outward concavity. As shown in fig. 8(1) - (8) (6), matrix map samples after data processing of each fault data are extracted from the total samples, the 128 × 128 matrix information is put into the constructed model for training, and for the purpose of fault classification, the activation function of the last layer of the neural network needs to be set to softmax, so that the model can output the diagnosis probability value of each fault.
The fault location model is also realized by adopting a convolutional neural network, the learning curve of the model on a data set is shown in fig. 5, and the overview of the model is shown in table 2.
TABLE 2 overview of fault location model
Adjustable weight number of output tensor dimensionality of neural network layer
The general framework is basically consistent with the convolutional neural network in fault classification, and it is slightly modified that 6 × 6 convolution kernels are changed to 3 × 3 in convolutional layer 1, and the number of extended convolution kernels is 64. A full connection layer is expanded, and the number of the neurons of the full connection layers 1-3 is 128, 64 and 32 respectively. The model position is divided into a head end, a middle end and a tail end 3, the output layer is 3 neurons, and the output result is the probability value of the head end, the middle end and the tail end. And setting a self-adaptive callback function in the training process, so that the model can self-adaptively adjust the learning rate of the optimizer and store the model with the highest precision in the training process. The model with the best effect on the learning curve in fig. 5 is selected by using the callback function, and the accuracies of the model in the training set and the model in the verification set are 98.15% and 83.33%, respectively.
And the detection terminal 3 processes the frequency response data and then puts the processed frequency response data into a previously trained fault positioning model for intelligent diagnosis, and the model judges the probability value of the fault at each position according to the data and takes the maximum probability value as a diagnosis basis.
For example: the first three sections of the nine-level winding model are divided into a head part, the middle three sections are divided into a middle part, and the rear three sections are divided into a tail part. Fig. 9(1), fig. 9(2), and fig. 9(3) are frequency response data of three positions after processing, respectively. The activation function of the last layer is also set to softmax in order for the model to output a probability value for diagnosing the location of the fault.
The invention provides a diagnostic system and a diagnostic method for fault classification and positioning of a transformer winding, which are used for destroying the head-tail symmetrical structure of a transformer winding model to a detection end in a mode of adding an external circuit element. The frequency response data of the head fault position and the tail fault position under the same fault can be distinguished in simulation; the difference of frequency response data of different fault positions can be increased in actual measurement. And after the training set data is obtained, putting the training set data into two models of fault classification and fault positioning for training respectively, and obtaining a model for classifying and positioning the fault of the transformer winding by adjusting parameters.
After the inductance at the head part and the capacitance at the tail part of the external circuit are increased, the whole model is not symmetrical relative to a detection instrument and a sweep generator. Therefore, when the head and tail symmetrical points have the same fault, the frequency response data are not consistent any more, the curves are not overlapped any more, and therefore the symmetrical fault points are distinguished.
In the feature extraction method, the features are not extracted by adopting the traditional data or image-based method, but the dimension slices of the frequency, the amplitude and the phase in the data are recombined after the difference data is obtained by subtracting the normal data from the fault sample data, and the data dimension is reshaped by utilizing an interpolation up-sampling or down-sampling method. The result of subtracting the fault sample data from the normal data improves the precision of the fault positioning model by about 5 percent compared with the direct fault sample data when the fault sample data is put into training. The method for extracting the characteristics can overcome the defect that high-frequency band information is neglected under an amplitude-phase polar coordinate scatter diagram. Compared with the traditional curve graph, the fault information characteristic can be embodied, and the training of the model is facilitated.
The model for fault diagnosis and fault location adopts a deep learning convolution neural network model. The training set acquired in claim 3 is processed into a single-channel image and put into training, and the model effect is adjusted by changing the number and size of convolution kernels, the types of activation functions, the pooling mode, the regularization mode and coefficients, the types of optimizers and the like. On the verification set, the diagnosis precision of the training set and the verification set in the fault classification model is 100%; the precision of a training set in the fault positioning model is as high as 98.15%, and the precision of a verification set is 83.33%. From the results, the method realizes fault type diagnosis and fault positioning, can amplify the difference of curves or data in practical application, and has stronger engineering significance.
Claims (5)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011429110.1A CN112763943B (en) | 2020-12-09 | 2020-12-09 | Diagnosis system and method for classification and location of transformer winding faults |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011429110.1A CN112763943B (en) | 2020-12-09 | 2020-12-09 | Diagnosis system and method for classification and location of transformer winding faults |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112763943A true CN112763943A (en) | 2021-05-07 |
| CN112763943B CN112763943B (en) | 2022-04-08 |
Family
ID=75693527
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011429110.1A Active CN112763943B (en) | 2020-12-09 | 2020-12-09 | Diagnosis system and method for classification and location of transformer winding faults |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112763943B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN204347165U (en) * | 2015-01-15 | 2015-05-20 | 国网四川省电力公司电力科学研究院 | Deformation of transformer winding failure simulation device |
| CN204719182U (en) * | 2015-06-15 | 2015-10-21 | 国网山西省电力公司电力科学研究院 | Inside transformer Solid Insulator Breakdown of Electric supersonic sounding device |
| CN111025022A (en) * | 2019-12-27 | 2020-04-17 | 广东电网有限责任公司电力科学研究院 | Transformer short-circuit fault positioning detection method and system |
| CN111273199A (en) * | 2018-12-05 | 2020-06-12 | 国网四川省电力公司眉山供电公司 | Intelligent detection method for transformer winding deformation based on sweep frequency impedance curve identification |
| RU2723911C1 (en) * | 2019-08-09 | 2020-06-18 | федеральное государственное бюджетное образовательное учреждение высшего образования "Тольяттинский государственный университет" | Method of electrodynamic tests of power transformers |
-
2020
- 2020-12-09 CN CN202011429110.1A patent/CN112763943B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN204347165U (en) * | 2015-01-15 | 2015-05-20 | 国网四川省电力公司电力科学研究院 | Deformation of transformer winding failure simulation device |
| CN204719182U (en) * | 2015-06-15 | 2015-10-21 | 国网山西省电力公司电力科学研究院 | Inside transformer Solid Insulator Breakdown of Electric supersonic sounding device |
| CN111273199A (en) * | 2018-12-05 | 2020-06-12 | 国网四川省电力公司眉山供电公司 | Intelligent detection method for transformer winding deformation based on sweep frequency impedance curve identification |
| RU2723911C1 (en) * | 2019-08-09 | 2020-06-18 | федеральное государственное бюджетное образовательное учреждение высшего образования "Тольяттинский государственный университет" | Method of electrodynamic tests of power transformers |
| CN111025022A (en) * | 2019-12-27 | 2020-04-17 | 广东电网有限责任公司电力科学研究院 | Transformer short-circuit fault positioning detection method and system |
Non-Patent Citations (1)
| Title |
|---|
| 王旭红等: "采用RWNN的电机定子绕组匝间故障诊断方法", 《高电压技术》 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112763943B (en) | 2022-04-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Sun et al. | Fault diagnosis of rolling bearing based on empirical mode decomposition and improved manhattan distance in symmetrized dot pattern image | |
| CN111751763B (en) | Power transformer winding fault diagnosis method based on GSMallat-NIN-CNN network | |
| KR102163217B1 (en) | Method and Apparatus for ECG Arrhythmia Classification using a Deep Convolutional Neural Network | |
| CN107564025B (en) | A Semantic Segmentation Method of Infrared Image of Power Equipment Based on Deep Neural Network | |
| CN112364931B (en) | Few-sample target detection method and network system based on meta-feature and weight adjustment | |
| Mahapatra et al. | Retinal vasculature segmentation using local saliency maps and generative adversarial networks for image super resolution | |
| CN111985543A (en) | Construction method, classification method and system of hyperspectral image classification model | |
| CN109916628A (en) | A fault diagnosis method for rolling bearings based on improved multi-scale amplitude sensing arrangement entropy | |
| CN107808138B (en) | Communication signal identification method based on FasterR-CNN | |
| CN104778692B (en) | A Fabric Defect Detection Method Based on Sparse Representation Coefficient Optimization | |
| CN106339998A (en) | Multi-focus image fusion method based on contrast pyramid transformation | |
| CN114821102B (en) | Dense citrus quantity detection method, equipment, storage medium and device | |
| CN106875380A (en) | A kind of heterogeneous image change detection method based on unsupervised deep neural network | |
| CN113657383A (en) | Defect region detection method and device based on lightweight segmentation model | |
| CN103745453B (en) | Urban residential areas method based on Google Earth remote sensing image | |
| CN117152422B (en) | A UV image anchor-free target detection method, storage medium, and electronic equipment | |
| CN113989652A (en) | Method and system for detection of cultivated land change under hierarchical multiple judgment rules | |
| CN112818920A (en) | Double-temporal hyperspectral image space spectrum joint change detection method | |
| CN113963272A (en) | A UAV image target detection method based on improved yolov3 | |
| CN118072080A (en) | Deep learning-based crop disease severity identification method | |
| CN111639697A (en) | Hyperspectral image classification method based on non-repeated sampling and prototype network | |
| CN106971392A (en) | A kind of combination DT CWT and MRF method for detecting change of remote sensing image and device | |
| CN114820438A (en) | Muffler surface material defect detection method and system | |
| CN115222712A (en) | Anomaly detection method in hyperspectral images using dual denoisers to replace artificial priors | |
| CN112763943A (en) | Diagnosis system and method for transformer winding fault classification and positioning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |