CN-122023916-A - Distribution transformer monitoring method and device

CN122023916ACN 122023916 ACN122023916 ACN 122023916ACN-122023916-A

Abstract

The invention discloses a distribution transformer monitoring method and device, and relates to the technical field of transformer monitoring. The method comprises the steps of collecting core texture signals of a distribution transformer under steady-state operation, constructing a reference texture signal through preprocessing, collecting real-time core texture signals, respectively extracting features of the real-time core texture signals and the reference texture signals through variation modal decomposition and symmetric point mode algorithms to obtain real-time feature images and reference feature images, comparing the real-time feature images with the reference feature images through an improved scale-invariant feature transformation algorithm, identifying suspected fault signal feature images and corresponding real-time core texture signals, inputting the suspected fault signal feature images into an improved depth residual shrinkage network model, outputting fault types of the distribution transformer, carrying out space-time synchronism analysis on the suspected fault signals, and calculating by combining fault type results to obtain fault diagnosis results to realize monitoring of the distribution transformer.

Inventors

ZHOU XIANJUN
WU FAQING
YU MENGTIAN
CHENG ZHONGYU
LIAN WEIZHENG
LIU FEI
LIU TAIQI

Assignees

国网湖北省电力有限公司洪湖市供电公司

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (10)

1. A monitoring method of a distribution transformer is characterized by comprising the following steps: S1, acquiring a core texture signal of a distribution transformer under a steady-state operation condition, and preprocessing the core texture signal to obtain a reference texture signal; S2, acquiring a real-time core texture signal, carrying out feature extraction on the real-time core texture signal through variation modal decomposition and a symmetric point mode algorithm, carrying out feature extraction on a reference texture signal through variation modal decomposition and a symmetric point mode algorithm, and correspondingly obtaining a real-time feature image and a reference feature image; s3, inputting the suspected fault signal characteristic image into an improved depth residual error shrinkage network model, and outputting to obtain a distribution transformer fault type result; And S4, carrying out space-time synchronism analysis on the suspected fault signals to obtain space-time synchronism analysis results, and calculating to obtain fault diagnosis results based on the fault type results and the space-time synchronism analysis results of the distribution transformer to realize monitoring of the distribution transformer.
2. The method for monitoring the distribution transformer according to claim 1, wherein the step of collecting the core texture signal of the distribution transformer under the steady-state operation condition comprises the following specific steps: Collecting core texture signals of the distribution transformer under a steady-state operation condition, wherein the core texture signals comprise voiceprint signals, vibration pattern signals and thermal voiceprint signals, the voiceprint signals are collected through electret capacitor gun-type directional microphones, the microphones are uniformly distributed around a transformer box body according to a symmetrical arrangement principle and are particularly arranged at four key structural parts near an iron core clamping piece, at the end part of a winding, in a radiating fin area and in a box body bearing part; Before starting data acquisition, whether the transformer is in a steady-state operation condition is firstly judged, wherein a steady-state judgment condition is set to be within 24 hours continuously, the load fluctuation amplitude of the transformer is less than or equal to +/-5%, the unbalance degree of three-phase current is less than or equal to 2%, the oil temperature change rate is less than or equal to 0.5 ℃ per hour, meanwhile, no obvious external interference exists, and when the condition is met, the transformer is judged to be in steady-state operation, data acquisition is started, and finally a core texture signal is obtained.
3. The method for monitoring a distribution transformer according to claim 2, wherein the preprocessing the core texture signal to obtain a reference texture signal comprises the following specific steps: For the collected voiceprint signals (T) preprocessing, and based on the preprocessed voiceprint signals, extracting three types of features to construct voiceprint references: The calculation formula of the frequency spectrum complexity is as follows: ; Wherein, the For the power spectral density duty cycle of the voiceprint signal at frequency f, As an upper limit of the effective frequency, Is the spectral complexity; The calculation formula of the dominant frequency is: ; Wherein, the The power duty cycle for the 50Hz doubled component, For the number of times of frequency multiplication, Is the dominant frequency; The calculation formula of the high-low frequency ratio is as follows: ; Wherein, the For the frequency doubling times corresponding to the high-low frequency demarcation, Is a high-low frequency ratio; The three types of characteristics are formed into a voiceprint characteristic matrix according to time sequence Statistical analysis is carried out on the voiceprint feature matrix, the mean value and standard deviation of each feature are calculated, and a voiceprint reference interval is constructed , wherein, Is the mean value of the voiceprint features, Standard deviation of voiceprint features; for the collected vibration pattern signal And (t) preprocessing, namely removing power frequency interference by adopting a 50Hz notch filter, removing environmental vibration noise by wavelet threshold denoising, extracting three types of characteristics and constructing a vibration pattern reference: the calculation formula of the vibration amplitude distribution parameter is as follows: ; Wherein, the Is the peak-to-peak value of the moire signal, Is the effective value of the moire signal, Is a vibration amplitude distribution parameter; The calculation formula of the periodic alternating characteristic parameters is as follows: ; Wherein, the As an autocorrelation function of the moire signal, In order for the delay time to be a time delay, Is a periodic alternating characteristic parameter; the calculation formula of the energy transfer path parameters is as follows: ; Wherein, the Is the first The collecting part and the first Cross-correlation coefficients of vibration ripple signals of the collecting positions, Is the combination number of4 acquisition positions, As a parameter of the energy transmission path, And Indexing the collected parts; The three types of characteristics form a vibration pattern characteristic matrix Statistical analysis is carried out on the vibration pattern characteristic matrix, and a vibration pattern reference interval is constructed , wherein, Is the mean value of the vibration pattern characteristics, Standard deviation of vibration pattern characteristics; For the collected voiceprint signals Preprocessing, wherein the preprocessing process is consistent with the voiceprint signal, and two types of features are extracted to construct a hot voiceprint reference: The fringe density parameter is calculated as: ; Wherein, the As the total number of fringes in the time-frequency spectrum, For the effective frequency range of the thermal voiceprint, For the total number of samples of the signal, As a parameter of the density of the stripes, Correcting the coefficient for the temperature change rate; The calculation formula of the arrangement direction parameter is as follows: ; Wherein, the Is the included angle between the kth stripe and the horizontal direction, Is an arrangement direction parameter; combining the two types of features into a thermal voiceprint feature matrix Statistical analysis is carried out on the thermal voiceprint feature matrix, and a thermal voiceprint reference interval is constructed , wherein, Is the mean value of the vibration pattern characteristics, Standard deviation of vibration pattern characteristics; Integrating the characteristic matrix of voiceprint, vibration pattern and thermal voiceprint with reference interval to construct multi-mode texture reference model Wherein Respectively a reference interval of voiceprint, vibration pattern and thermal voiceprint; and selecting a signal with a characteristic value within +/-5% of the center of the reference interval from the core texture signals, and taking the signal as a reference texture signal.
4. A monitoring method of distribution transformer according to claim 3, wherein the collecting real-time core texture signal, performing feature extraction on the real-time core texture signal through variation mode decomposition and symmetric point mode algorithm, performing feature extraction on reference texture signal through variation mode decomposition and symmetric point mode algorithm, and correspondingly obtaining real-time feature image and reference feature image, comprises the following steps: collecting a real-time core texture signal, wherein the real-time core texture signal comprises a real-time voiceprint signal, a real-time vibration voiceprint signal and a real-time thermal voiceprint signal, and the real-time voiceprint signal is recorded as The real-time vibration signal is recorded as The real-time thermal voiceprint signal is recorded as The acquisition parameters of the three types of signals are consistent with those of the step S1, the real-time core texture signals are subjected to unified preprocessing, the preprocessed real-time core texture signals are respectively subjected to variational modal decomposition and decomposition, and the complete mathematical model comprises an objective function, constraint conditions and an augmented Lagrange function: ; ; Wherein, the Is the kth inherent modal component, K is the modal number, For the center frequency of the kth IMF component, As a dirac pulse function, In units of imaginary numbers, Refers to the pre-processed real-time signal, In order to be a penalty factor, In order to be a lagrange multiplier, For the inner product operation, the method comprises the steps of, Is an augmented lagrangian function; And respectively carrying out symmetric point mode transformation on 5 IMF components and original signals obtained by decomposing the real-time core texture signals, and converting a one-dimensional time domain signal into a two-dimensional polar coordinate characteristic image, wherein the transformation core formula is as follows: ; ; ; Wherein, the Is the first The radius of the individual sample points in polar coordinates, For the i-th sample point amplitude of the IMF component, And The maximum and minimum magnitudes of the IMF components, Is the first The counterclockwise rotation angle of the mirror-symmetrical portions, The number of symmetrical parts is 6, In order to rotate the angle of the clockwise direction, As a hysteresis factor, the value of 20, The angle gain factor is taken as 30 degrees, For the maximum radius of the polar feature image, For the index of the sample point, Index for mirror image unit; Extracting a group of radial arms from the SDP conversion results of the K IMF components and the original signals of the real-time core texture signals respectively, and fusing to obtain a real-time characteristic image, wherein the fusing rule is that the radial arms of the components are sequentially overlapped to corresponding mirror units of polar coordinate images according to the central frequency of the IMF components from low to high; and carrying out feature extraction on the reference texture signal through variation modal decomposition and a symmetric point mode algorithm to obtain a reference feature image.
5. The method for monitoring a distribution transformer according to claim 4, wherein the step of performing feature comparison on the real-time feature image and the reference feature image by using the improved scale invariant feature transform algorithm, identifying to obtain a feature image of a suspected fault signal, and recording a real-time core texture signal corresponding to the feature image of the suspected fault signal as the suspected fault signal comprises the following steps: carrying out Gaussian blur and downsampling on the characteristic images in different scales to generate 6 layers of Gaussian pyramids, wherein each layer comprises 5 Gaussian difference images and scale factors In the Gaussian difference image, comparing each pixel with 8 neighborhood pixels and pixels corresponding to the upper layer and the lower layer, screening out extreme points as candidate feature points, eliminating candidate points with low contrast and edge response, retaining stable feature points, constructing a 16×16 neighborhood window by taking the feature points as the center, dividing the neighborhood window into 4×4 sub-areas, calculating gradient histograms of 8 directions in each sub-area, and generating 128-dimensional feature description vectors; calculating Euclidean distance between the real-time characteristic image characteristic description vector and the reference characteristic image characteristic description vector, adopting the ratio criterion of nearest neighbor to next nearest neighbor for initial matching, and keeping meeting the requirement Wherein As the nearest-neighbor distance is the nearest-neighbor distance, Is the next nearest neighbor distance; Randomly selecting 3 characteristic points A, B, C and corresponding matching points A ', B ', C ' from the initial matching pair to construct a triangle ABC and A ' B ' C ' the matching validity is verified by the following formula: ; Wherein, the And Respectively triangular ABC、 The length of the corresponding side of A ' B ' C ', And Respectively the included angles of the corresponding vertexes, Takes the value of 0.1 for the consistency threshold value of the side length proportion, And In order to be a threshold for angular consistency, The value of the water-based paint is 1, The value is 20; traversing all initial matching pairs, repeating similar triangle verification, and counting voiceprint matching points Number of matching points of vibration patterns Number of hot voiceprint matches ; Calculating the average value of the matching points of the reference characteristic image and the real-time characteristic image Setting a single-mode matching threshold value ; And when the number of the feature image matching points is lower than the corresponding threshold value, marking the feature image as a suspected fault signal feature image, and marking a real-time core texture signal corresponding to the suspected fault signal feature image as a suspected fault signal.
6. The method for monitoring a distribution transformer according to claim 5, wherein the step of inputting the characteristic image of the suspected fault signal into an improved depth residual shrinkage network model and outputting a fault type result of the distribution transformer comprises the following steps: Normalizing the feature images of the suspected fault signals, and performing channel splicing according to the sequence of voiceprint-vibration print-hot voiceprint to obtain a fusion feature matrix For fusion feature matrix Performing principal component analysis and dimension reduction processing to obtain a dimension-reduced core feature matrix And then, a two-dimensional feature map which is remolded into 32 multiplied by 1 is constructed, an improved depth residual shrinkage network model is constructed, a multi-classification cross entropy loss function is adopted as a loss function of the improved depth residual shrinkage network, and a calculation formula is as follows: ; Wherein, the In order to be a number of samples of a batch, Is the first The true labels of the individual samples are then displayed, Predicted for network The probability of a class failure, For multi-class cross entropy loss, For the index of the sample, Indexing for fault category; core feature matrix Inputting the improved depth residual error shrinkage network model after training, and outputting the probability distribution of various faults Corresponding to normal, mechanical looseness, electromagnetic unbalanced load and local overheating; and selecting the category with the largest probability distribution as a distribution transformer fault type result.
7. The method for monitoring distribution transformer according to claim 6, wherein the construction of the improved depth residual shrinkage network model comprises the following specific steps: The improved depth residual error shrinkage network has the overall structure that an input layer, a convolution layer, 4 improved residual error shrinkage units, a BN layer, APReLu layers, a global average pooling layer, a full connection layer and an output layer are formed; The adaptive parameter correction linear unit APReLu is adopted, and the core formula is as follows: ; ; ; Wherein, the Is the output feature vector of APReLu layers, In order to adapt the weight value of the device, The intermediate variables are learned for the weights, As a matrix of weights, the weight matrix, As a result of the bias term, In order to input the feature vector(s), Performing global average pooling on the input feature vector x to obtain an output result; Each improved residual contraction unit IRSBU includes a main path, namely a convolution layer, a BN layer, a APReLu layer and a convolution layer, and a threshold learning path, namely an absolute value layer, a GAP layer, a full connection layer and a Sigmoid layer, wherein the specific formulas are as follows: ; Wherein, the For the feature matrix of the c-th channel, For the c-th channel feature matrix after soft thresholding, Is the adaptive threshold for the c-th channel, Scaling coefficients are independent for the channels.
8. The method for monitoring a distribution transformer according to claim 7, wherein the step of performing the space-time synchronicity analysis on the suspected fault signal to obtain a space-time synchronicity analysis result comprises the following specific steps: the space-time synchronicity analysis result comprises a time synchronicity coefficient and a space synchronicity coefficient, and the calculation formula of the time synchronicity coefficient is as follows: ; Wherein, the Is the first Class modality and No A temporal synchronicity coefficient of the class modality, the closer the absolute value is to 1, the stronger the temporal synchronicity, Is the first A time series of fault signatures of the class modality, For the time-sampling point index, Is the first The mean value of the time series of modal-like fault signatures, Is the first A time series of fault signatures of the class modality, Is the first The mean value of the time series of modal-like fault signatures, And Indexing the modal fault characteristics; Setting | | 0.85 Is a strong synchronization and, In order to be a weak synchronization, If the two or more modes are asynchronous, the compound fault is judged when the two or more modes meet the strong synchronization, and if the fault characteristics of only one mode occur and the two or more modes are asynchronous, the compound fault is judged as a single fault; The spatial synchronicity coefficient is calculated by the following formula: ; Wherein, the Is the first Class modality and No The closer to 1 the spatial synchronicity coefficient of the class mode, the higher the spatial overlap, Is the first The spatial region corresponding to the modal-like fault signature, Is that And (3) with Is defined by the number of intersection area sensor points, Is that And (3) with Is the number of union region sensor sites, Is the first A space region corresponding to the modal-like fault feature; When (when) =1, Indicating that the fault signature is concentrated in the same spatial region, when =0, Which indicates that the fault signature is distributed across different regions, and the composite fault can be subdivided into region composite faults and cross-region composite faults in combination with the time synchronicity coefficient.
9. The method for monitoring a distribution transformer according to claim 8, wherein the fault diagnosis result is calculated based on a distribution transformer fault type result and a space-time synchronicity analysis result, and specifically comprises the following steps: The positioning formula of the mechanical loosening fault is as follows: ; Wherein, the For the mechanical loosening fault location result, For the sensor index to be used, The average amplitude of the low frequency IMF component for the vibration pattern acquired for the first sensor, For the average amplitude of the low-frequency IMF component of the vibration pattern corresponding to the first sensor in the reference model, The absolute value of the deviation of the amplitude of the first region from the reference, Is the deviation rate; The positioning formula of the electromagnetic unbalanced load fault is as follows: ; Wherein, the For the electromagnetic unbalanced load fault positioning result, Spectral distortion of the mid-frequency IMF component of the voiceprint acquired for the first sensor, The local time synchronism coefficient of the voiceprint and the vibration pattern of the first area; The localization formula of the local overheat fault is: ; ; ; Wherein, the For the localized overheat fault localization result, The density of hot voiceprint stripes acquired for the first sensor, Is the mean value of the fringe density in the reference model, In order to increase the density of the material, The total energy of the high frequency IMF component for the thermal voiceprint acquired for the first sensor, As the energy mean value in the reference model, Is the energy increment; the composite fault positioning adopts a region superposition and weight distribution strategy, namely, the region composite fault takes the intersection region of each single fault positioning result, and the cross-region composite fault outputs probability according to the network of each single fault Weight is distributed, and the weight formula is that And finally obtaining a fault diagnosis result, wherein the fault diagnosis result comprises a distribution transformer fault type result and a fault positioning result.
10. A distribution transformer monitoring device comprising a memory, a processor and a computer program stored on the memory, wherein the processor executes the computer program to carry out the steps of the method of any one of claims 1 to 9.

Description

Distribution transformer monitoring method and device Technical Field The invention relates to the technical field of transformers, in particular to a monitoring method and device for a distribution transformer. Background The distribution transformer is used as core equipment of a power system distribution network link and bears key responsibilities of voltage conversion, electric energy distribution and transmission, and the running state of the distribution transformer is directly related to power supply reliability and power consumption safety. In the long-term operation process, the distribution transformer is easily subjected to hidden faults such as mechanical loosening, electromagnetic unbalanced load, local overheating and the like due to the influences of electromagnetic force, mechanical vibration, environmental factors, material aging and the like, and if the faults are not timely identified, the faults can be gradually developed into serious faults due to weak early characteristics and slow evolution. In the traditional distribution transformer fault monitoring, a single scale invariant feature transformation algorithm is often adopted for fault feature identification. According to the method, the characteristic images of the transformer in the running state are collected, scale invariant characteristic points in the images are extracted, the characteristic description vector is constructed, then the characteristic description vector is compared with the preset normal state characteristic vector, whether the equipment has faults or not is judged according to the matching result, and certain application is achieved in a scene with obvious conventional fault characteristics. However, most of fault characteristic signals of the distribution transformer are broadband multi-mode signals, and comprise complex information of multiple frequency bands such as voiceprint, vibration pattern and thermal voiceprint, the conventional scale-invariant characteristic transformation algorithm does not conduct targeted frequency band decomposition and characteristic optimization on original signals, and useful signals and interference noise of different frequency bands cannot be effectively separated, so that fine characteristic differences corresponding to early hidden faults cannot be accurately captured, mismatching and mismatching situations are prone to occurring in a fault identification process finally, and actual requirements of accurate early fault monitoring are difficult to meet. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a method and a device for monitoring a distribution transformer, which are used for solving the problems existing in the background art. In order to achieve the purpose, the invention is realized by the following technical scheme that the method and the device for monitoring the distribution transformer comprise the following steps: S1, acquiring a core texture signal of a distribution transformer under a steady-state operation condition, and preprocessing the core texture signal to obtain a reference texture signal; S2, acquiring a real-time core texture signal, carrying out feature extraction on the real-time core texture signal through variation modal decomposition and a symmetric point mode algorithm, carrying out feature extraction on a reference texture signal through variation modal decomposition and a symmetric point mode algorithm, and correspondingly obtaining a real-time feature image and a reference feature image; s3, inputting the suspected fault signal characteristic image into an improved depth residual error shrinkage network model, and outputting to obtain a distribution transformer fault type result; And S4, carrying out space-time synchronism analysis on the suspected fault signals to obtain space-time synchronism analysis results, and calculating to obtain fault diagnosis results based on the fault type results and the space-time synchronism analysis results of the distribution transformer to realize monitoring of the distribution transformer. Preferably, the collecting the core texture signal of the distribution transformer under the steady-state operation condition includes the following specific steps: The method comprises the steps of collecting core texture signals of a distribution transformer under a steady-state operation condition, wherein the core texture signals comprise voiceprint signals, vibration pattern signals and thermal voiceprint signals, the voiceprint signals are collected through electret capacitor gun-type directional microphones, the microphones are uniformly distributed around a transformer box body according to a symmetrical arrangement principle and are specifically arranged at four key structure parts near an iron core clamping piece, at the winding end part, in a radiating fin area and at a box body bearing part, 1 microphone is arranged at each part, the microphones are vertical to the surfa