CN111597957A

CN111597957A - Transformer winding fault diagnosis method based on morphological image processing

Info

Publication number: CN111597957A
Application number: CN202010398021.9A
Authority: CN
Inventors: 李振华; 张宇杰; 李振兴; 徐艳春; 邾玢鑫; 刘颂凯; 杨楠; 张磊
Original assignee: China Three Gorges University CTGU
Current assignee: China Three Gorges University CTGU
Priority date: 2020-05-12
Filing date: 2020-05-12
Publication date: 2020-08-28
Anticipated expiration: 2040-05-12
Also published as: CN111597957B

Abstract

The transformer winding fault diagnosis method based on morphological image processing comprises the following steps: establishing a lumped parameter model for the transformer winding to obtain amplitude-frequency curve data of the normal transformer winding and various faults of the transformer winding; selecting two frequency segments with obvious changes for the acquired amplitude-frequency curve data, and storing the two frequency segments in the form of images; establishing an image library and a label for the obtained image; and importing images under two frequency bands of the transformer winding fault to be detected, and carrying out preprocessing and morphological methods on the images in the established image library to finally obtain an area value. And sequencing the area values, and diagnosing the transformer winding faults according to the sequencing result. The method has high accuracy, can identify the fault type and present the fault degree of the fault type, and is beneficial to the comprehensive evaluation of the running state of the transformer to be tested by maintainers.

Description

Transformer winding fault diagnosis method based on morphological image processing

Technical Field

The invention relates to the technical field of transformer fault detection, in particular to a transformer winding fault diagnosis method based on morphological image processing, which is used for offline detection of a transformer winding.

Background

The transformer is used as a power grid junction, plays a role in converting voltage levels and transferring energy. The safe and stable operation of the electric network cannot leave the normal working state. Thus, fault detection of the transformer is particularly important. The existing detection methods for transformer faults can be roughly divided into non-electrical quantity detection and electrical quantity detection. The detection of the non-electrical quantity includes DGA analysis, ultrasonic detection, infrared temperature measurement, etc. The electric quantity detection includes a frequency response method, partial discharge detection, and the like. By utilizing a frequency response method and through off-line detection, the difference of the contrast amplitude of the secondary side and the primary side is obtained for the primary detection input sweep frequency sinusoidal signal of the transformer, and the difference of the amplitude-frequency curves under different faults can be better obtained. And according to the change of the analysis curve, the purpose of fault diagnosis is achieved. The method has high accuracy, and compared with the DGA, the method has less possibility of misjudgment and can more accurately grasp the fault position. Compared with infrared temperature measurement, the method depends on curve judgment rather than subjective temperature judgment. Therefore, the frequency response method is more and more widely applied to fault diagnosis of the transformer.

There have been many studies related to the frequency response method aiming at the frequency response method to integrate the influence of various fault changes on the frequency response curve. However, the prior art has insufficient utilization of the frequency response method, which is mainly expressed in that: firstly, the transformer cannot be effectively diagnosed by using partial difference of frequency response curves. Secondly, the frequency response data can not be mined by the existing method, and the existing method is excessively dependent on the thought judgment.

The Chinese patent 'transformer winding deformation detection method based on frequency response impedance method' (CN106338237A) obtains frequency response impedance data by innovating a frequency response method. The method can effectively diagnose whether the transformer is in a fault state by using the curve similarity of the two images as a criterion. However, it is impossible to diagnose what type of fault the fault belongs to, and even under what fault level.

Chinese patent "a transformer fault diagnosis method and apparatus" (CN108267660B), provides a method for judging whether a transformer has a fault by using surge voltage in a heterogeneous manner, and inputs surge voltage at a low voltage side, observes a voltage signal at a high voltage side, and analyzes the voltage signal. However, as in the above, the type of fault still cannot be diagnosed, and only whether the working state of the transformer is abnormal or not can be determined.

Therefore, when the transformer is subjected to fault diagnosis by using the electric quantity, the main problem is insufficient utilization of data. The data obtained by the existing method cannot be processed to obtain more accurate diagnosis results and higher identification accuracy.

Disclosure of Invention

In order to solve the problems, the invention provides a transformer winding fault diagnosis method based on morphological image processing, which can effectively utilize the electric quantity data obtained by a frequency response method and achieve the purpose of more accurately diagnosing faults of a transformer. Compared with the existing detection means, the method has high accuracy, can identify the fault type and present the fault degree of the fault type, and is favorable for the maintainer to comprehensively evaluate the running state of the transformer to be detected.

The technical scheme adopted by the invention is as follows:

the transformer winding fault diagnosis method based on morphological image processing comprises the following steps:

step 1: establishing a lumped parameter model for the transformer winding to obtain amplitude-frequency curve data of the normal transformer winding and various faults of the transformer winding;

step 2: selecting two frequency segments with obvious changes for the amplitude-frequency curve data acquired in the step 1, and storing the two frequency segments in the form of images;

and step 3: establishing an image library and a label for the image obtained in the step 2;

and 4, step 4: and (4) importing images of the transformer winding to be tested in two frequency bands, and carrying out preprocessing and morphological methods on the images in the image library established in the step (3) to finally obtain an area value.

And 5: and 4, sequencing the area values in the step 4, and diagnosing the transformer winding faults according to the sequencing result.

In the step 1, before diagnosing a certain type of transformer winding fault, a model needs to be established and simulated for the transformer winding, a lumped parameter model is established to reflect the electrical characteristics of the transformer winding, and the transformer winding models under different faults are simulated by changing the element parameter values at different positions:

the simulation obtains 46 groups of data, including 45 groups of fault data of which the element parameter values are respectively increased by 20%, 40%, 60%, 80% and 100% under C, K, L fault types at the head part, the middle part and the tail part, and normal data under normal conditions; applying voltage frequency sweep signals to the primary side of the circuit model under different types of faults, and obtaining the ratio of the secondary side output voltage of ten times frequency to the primary frequency sweep signal:

wherein, U1 is a sweep frequency voltage signal inputted at the primary side; u2 is a voltage signal of the secondary side response output; dB is amplitude-frequency data which changes along with the frequency of the sweep.

In step 1, the 46 sets of data include: normal data under normal conditions; fault data of 20%, 40%, 60%, 80% and 100% increase of head inductance, longitudinal capacitance and capacitance to ground; the middle part inductance, the longitudinal capacitance and the ground capacitance are increased by 20%, 40%, 60%, 80% and 100% of fault data; tail inductance, longitudinal capacitance, and capacitance to ground increase 20%, 40%, 60%, 80%, and 100% of fault data.

In the step 2, because the difference of the full-frequency section change curve images is small, the parts with large difference of the amplified amplitude-frequency curves are respectively a frequency section a: the frequency is 0-55 kHz, and the amplitude is-110-10 dB; frequency segment b: 80-200 kHz, and the amplitude is-600 to-210 dB.

In the step 3, 92 images can be obtained by intercepting the images of two frequency bands for the obtained 46 sets of data. And classifying the two images under each group of faults into one type, and labeling the two images with labels, wherein the labels comprise fault positions, fault types and fault grades of the faults.

In step 4, the imported image is the same as the two frequency bands in the image library, i.e. the same frequency range and the same amplitude range.

In the step 4, the image is preprocessed as follows: linearly adding the two frequency bands of the image to be detected and all the corresponding frequency bands in the image library to realize image synthesis; then, performing image negation, and adjusting gray value information of the synthesized image; and finally, filling the image by using a morphological method.

In the step 4, the area value of each synthesized image is obtained, and the two area values under each group of faults are added to obtain the total area value under each type of faults; since the image library contains 46 images, 46 area values are finally obtained, and the area values are placed behind the labels in the image library to be used as a basis for judgment of final diagnosis.

In the step 5, the images to be detected are sorted according to the area values, and according to the label result, which fault type the image to be detected belongs to is determined, specifically: according to the sequencing result, the fault type at the head of the ranking is a diagnosis result, and two unequal fault grades ranked at the top end are the upper limit and the lower limit of the diagnosis fault grade.

The invention relates to a transformer winding fault diagnosis method based on morphological image processing, which has the following technical effects:

(1): the method for introducing the image library fully covers the fault type of the transformer.

(2): and (3) analyzing the amplitude-frequency curve by combining morphology, fully mining the amplitude-frequency curve, and judging the coincidence degree of the amplitude-frequency curve by a contrast quantification method.

(3): the diagnosis method overcomes the bottleneck of insufficient evaluation of the running state of the transformer and provides a more precise and complete diagnosis result. And providing fault information of the fault type and the corresponding fault grade interval on the basis of whether the fault exists or not.

(4): the invention provides a method for analyzing and diagnosing a frequency response method curve of a transformer. The concept of the image library is introduced, and a method, a flow and a rule for constructing the image library are provided. By combining an image processing technology and a morphological method, the difference of the amplitude-frequency curve characteristics of the frequency response method is visually expressed and quantified by utilizing the operations of image combination, gray value information adjustment, image inversion, image filling, area calculation and the like. The frequency section images at two ends of the amplitude-frequency curve of the transformer to be tested are input, and the fault type and the corresponding fault grade interval of the transformer to be tested can be effectively identified through the operation and program diagnosis. And comprehensively and finely evaluating the state of the transformer to be tested.

(5): the verification proves that the invention has accurate fault identification effect and good performance. Compared with the existing transformer winding diagnosis method, the method has the advantages that the recognition accuracy is high, the fault type of the transformer is further pointed out and the possible fault grade interval of the transformer is deduced on the basis of judging whether the transformer is in fault, and the running state of the transformer to be tested is more comprehensively and specifically presented.

Drawings

FIG. 1 is a schematic diagram of a constructed image library.

FIG. 2 is a flowchart of the process of the present invention.

Fig. 3 is a diagram of a fault type simulation verification result.

Fig. 4 is a diagram of a fault level simulation verification result.

Detailed Description

And drawing a frequency response amplitude-frequency curve under different faults according to the acquired data. And intercepting and amplifying two frequency bands which obviously change along with the parameter change wave curve, and storing the frequency bands in an image form. Then, an image library is constructed by all the saved images, and the fault type, the fault position and the fault level of the two images under each fault are identified. After the graph library of the transformer is constructed, the imported amplitude-frequency curve image group to be tested and the image group in the fault library can be compared in a programmed mode through an image processing technology. The fault type of the transformer to be tested and the interval of the corresponding fault grade can be diagnosed.

And the programming part is combined with a morphological image processing technology, and a pre-established image library is utilized to diagnose the amplitude-frequency curve. Image merging to linearly merge the curve images under the corresponding frequency segments and the images in the image library in an equal proportion one by one; adjusting the gray value information of the combined image so as to fill the image and reflect the difference of the amplitude-frequency curve; and comparing the difference between the curve image to be detected and various faults in the image library for quantification, and finding out a group of faults which are most matched with the curve image to be detected from small to large. By analyzing the sequencing result, which type of fault to be detected belongs to and the grade interval of the corresponding fault can be diagnosed.

The diagnosis result comprises: and after the results of the fault types and the fault levels are obtained, the sequencing results in the diagnostic program, the curve images most matched with the fault images to be detected and the curve image difference of the most matched images on the two frequency bands can be checked. In order to more intuitively represent the difference of the curve images, the result is presented by combining the two curves and filling the image result. The frequency response curve obtained by the frequency response method has no sudden change in the whole, namely the same fault type is increased along with the fault grade, the trend is basically unchanged, and the difference of two images under different faults is obvious under two frequency sections. Therefore, after the diagnosis result is obtained, the image can be used as a reference for the examining staff to suspect the diagnosis result.

Example (b):

the method for diagnosing the fault of the transformer winding based on the morphological image processing needs to model and simulate the winding of a certain type of transformer to obtain data frequency response amplitude-frequency data before the transformer leaves a factory. According to the data, a required image library is constructed in advance according to the rules and the method of the invention, and necessary basis is provided for fault diagnosis.

The image library, as described above, requires image information for two frequency bins of the amplitude-frequency curve, as well as the label content. The tag content must include the type of failure, location of failure, and level of failure. Supplementary information for this type of failure may also be added as appropriate. As shown in fig. 1. The two frequency bands are a and b respectively, the content of the two frequency bands is the size of the image, and the image is stored in a matrix form and contains the pixel value information of the image. The subsequent part is label information. The label information includes fault location, fault type, fault level, and supplementary information, fault name. At the fault position, s, z and w respectively represent that the fault position is at the head part, the middle part and the tail part and are represented by the initial pinyin of the fault position. At fault type C, K, L represents the three fault types of the transformer respectively. At fault level, 20, 40, 60, 80, 100 represents the percentage of the fault level under the fault.

The frequency response method belongs to an off-line detection method, and needs to be temporarily removed from a power grid when amplitude-frequency data of a transformer to be detected are acquired. Inputting a sweep frequency voltage signal at the primary side, and according to the method of the invention, responding data at the secondary side of the mobile phone and obtaining final amplitude-frequency data according to a formula (1). In order to be matched with the two frequency sections in the image library, the two frequency sections with the same frequency and amplitude as the image library in the amplitude-frequency curve to be detected are extracted and input into a program in an image form for judgment. It should be noted that: the size of the two input images needs to be consistent with that of the two images in the image library, and if the two input images are not consistent, the operation cannot be performed. When the images are inconsistent, the images can be amplified or reduced through up-sampling or down-sampling so as to ensure that the images are correspondingly compared with the frequency response curves in the image library.

After the images are acquired, the input images and the image library images are processed using image processing techniques, and a flowchart is shown in fig. 2.

(1) Reading in an image group to be detected and an image library;

(2) combining the image group to be detected with the x-th image in the image library and adjusting gray value information;

(3) inverting the new image and filling the image;

(4) calculating the sum of the areas of the filled image groups, and putting the sum into the structural elements;

(5) circularly traversing all elements in the image library, and repeating the operations (2) to (4);

(6) and sorting by taking the area as a criterion.

In the step (1), the image group to be detected is two images of the transformer to be detected under the same two frequency sections and amplitudes as the image library, the size of the image group to be detected is the same as the image specification in the image library, and the image library and the amplitude-frequency curve image to be detected are gray images.

The reason is that: the frequency response curve does not need to consider the change of the color of the frequency response curve, but focuses more on the change information of the amplitude-frequency curve. The effect on the analysis is the same regardless of whether it is a color image or a grayscale image. However, the storage space of the color image is enlarged by 3 times compared with the gray image, which wastes the resources of the storage space, causes unnecessary waste, and is completely unnecessary. The operations in the program have also been performed for the processing of the grayscale image as an example. If the image to be detected is a color image, the color image can be converted into a gray image and then subsequent operation is carried out.

And (3) performing operation in the step (2) mainly by carrying out equal-proportion linear addition on the images, adjusting gray value information and changing part of gray value information. Image equal proportion linear addition:

wherein, Image is the synthesized Image, I₁、I₂Two images before composition.

The reason for the proportional linear addition is: both images are amplitude-frequency curve images, and any one of the images cannot be emphasized more preferentially if the two images are compared, otherwise, the subsequent area calculation result is inaccurate. The coefficients must all be 0.5. The grayscale image pixel value is 255 at maximum. When the sum of the gray values of the two images is larger than 255, the gray value of the two images is reduced by 255, so that a part of important curve image information is lost. However, when the gray-level value is less than 0 when subtracting two images, the pixel value will be the absolute value of the subtraction result. After the linear addition, the gray scale information of the non-overlapping part of the curves becomes half of the original value, that is, the brightness is halved. At this time, the gray scale information is adjusted to be bright. After this operation is completed, for the subsequent image filling to proceed smoothly. The filling is effective when a part of the image information is changed so that the difference portion of the amplitude-frequency curves of the two images becomes a closed curve portion. In order to make this step smooth. The first column of the first frequency band image and the first row of the second frequency band image are assigned 255. The image filling can be smoothly performed on the former part of the first frequency band and the former part of the second frequency band.

In step (3), the last step of image filling, namely image inversion, is still required. The image filling is to fill a part surrounding the white part with white and reverse lines. Whereas the curved segments are black. Therefore, in order to fill in the difference representing the amplitude-frequency curve with an image, the synthesized curve image must be inverted. The curve after inversion is white, and the background is black. The black background in the middle of the two curves can be replaced by white by image filling, so that the difference of the two amplitude-frequency curve images can be represented. The larger the difference, the more white ranges. The smaller the difference, the smaller the white range. When the difference is very small, i.e. the two curves substantially coincide, there is substantially no filled portion.

And (4) calculating the areas of the white areas of the two images filled in the step (2) during the method for quantizing the difference of the amplitude-frequency curves. If the difference is smaller, the white portion in the image is smaller, and the corresponding area value is smaller. If the area is larger, the white portion in the image is larger, and the area value is larger. The area values under the two frequency bands are added, and the data of the area values are placed in the corresponding fault tags to reflect the difference.

And (5) repeating the operations in the steps (2) to (4), and after the step (4) is completed, obtaining the difference result of each group of faults of the images of the amplitude-frequency curve of the transformer to be tested in the image library under the two frequency bands. Expressed as an area value and placed behind the tag.

And (6) utilizing the area result of the step (5) to sequence the faults in the image library. And obtaining a group of amplitude-frequency curve images most consistent with the amplitude-frequency curve to be detected and corresponding label information. The fault type of the transformer can be diagnosed according to the label. The first few of the sorted results still have reference values. The fault grade interval of the fault corresponding to the transformer can be judged according to the first two unequal fault grades of the sequencing result so as to judge the running state and the fault degree of the transformer more accurately and comprehensively. The maintainer may also recall the combined image to observe the difference from the most consistent fault.

To verify the recognition accuracy of the present invention. The accuracy test was performed thereon, and the results are shown in table 1.

The faults acquired by simulation have randomness. Random fault type, fault location, fault class. The failure positions of the verification samples comprise a head part, a middle part, a tail part, a partial head part and a partial tail part. And verifying that the failure grade of the sample is different from 10-90%. The failure types of the validation sample contained all C, K, L three failure types. And the sample is verified to be not coincident with the sample in the image library.

Table 1 verification results table

As can be seen from table 1, C, K, L samples of different fault positions and different fault levels of the three fault types are included in the sample group. The results show that the diagnosis using the algorithm works well in this sample set. The diagnosis accuracy rate for the fault type is up to 100%, and the diagnosis accuracy rate for the fault grade interval is up to 91%.

The result of the verification of the fault type is shown in fig. 3, and the sample group includes a ground capacitance C fault group 4, a longitudinal capacitance K fault group 3, and an inductance L fault group 4. The diagnosis accuracy of each group of type faults is 100%, which shows that the method has good diagnosis effect on the fault types.

The verification of the failure level section is shown in fig. 4. The samples in the table 1 are arranged from small to large according to the ascending order, and the fault level of the samples and the upper limit and the lower limit of the fault level interval detected by the algorithm are presented in the form of a line graph. The result shows that the detection interval of only one group of samples does not contain a sample interval, and the diagnosis accuracy is up to 91%. The method has a good diagnosis effect on the fault grade interval.

Claims

1. The transformer winding fault diagnosis method based on morphological image processing is characterized by comprising the following steps of:

and 4, step 4: importing images under two frequency bands of the transformer winding fault to be detected, and carrying out preprocessing and morphological methods on the images in the image library established in the step 3 to finally obtain an area value;

2. The transformer winding fault diagnosis method based on morphological image processing as claimed in claim 1, characterized in that: in the step 1, before a certain type of transformer winding fault is diagnosed, a model is established and simulated for the transformer winding, a lumped parameter model is established to reflect the electrical characteristics of the transformer winding, and the transformer winding models under different faults are simulated by changing the element parameter values at different positions:

wherein, U₁A sweep frequency voltage signal input for the primary side; u shape₂Responding the output voltage signal for the secondary side; dB is amplitude-frequency data which changes along with the frequency of the sweep.

3. The transformer winding fault diagnosis method based on morphological image processing as claimed in claim 2, characterized in that: in step 1, the 46 sets of data include: normal data under normal conditions; fault data of 20%, 40%, 60%, 80% and 100% increase of head inductance, longitudinal capacitance and capacitance to ground; the middle part inductance, the longitudinal capacitance and the ground capacitance are increased by 20%, 40%, 60%, 80% and 100% of fault data; tail inductance, longitudinal capacitance, and capacitance to ground increase 20%, 40%, 60%, 80%, and 100% of fault data.

4. The transformer winding fault diagnosis method based on morphological image processing as claimed in claim 1, characterized in that: in the step 2, because the difference of the full-frequency section change curve images is small, the parts with large difference of the amplified amplitude-frequency curves are respectively a frequency section a: the frequency is 0-55 kHz, and the amplitude is-110-10 dB; frequency segment b: 80-200 kHz, and the amplitude is-600 to-210 dB.

5. The transformer winding fault diagnosis method based on morphological image processing as claimed in claim 1, characterized in that: in the step 3, 92 images can be obtained by intercepting the images of two frequency bands from the acquired 46 groups of data; and classifying the two images under each group of faults into one type, and labeling the two images with labels, wherein the labels comprise fault positions, fault types and fault grades of the faults.

6. The transformer winding fault diagnosis method based on morphological image processing as claimed in claim 1, characterized in that: in step 4, the imported image is the same as the two frequency bands in the image library, i.e. the same frequency range and the same amplitude range.

7. The transformer winding fault diagnosis method based on morphological image processing as claimed in claim 1, characterized in that: in the step 4, the image is preprocessed as follows: linearly adding the two frequency bands of the image to be detected and all the corresponding frequency bands in the image library to realize image synthesis; then, performing image negation, and adjusting gray value information of the synthesized image; and finally, filling the image by using a morphological method.

8. The transformer winding fault diagnosis method based on morphological image processing as claimed in claim 1, characterized in that: in the step 4, the area value of each synthesized image is obtained, and the two area values under each group of faults are added to obtain the total area value under each type of faults; since the image library contains 46 images, 46 area values are finally obtained, and the area values are placed behind the labels in the image library to be used as a basis for judgment of final diagnosis.

9. The transformer winding fault diagnosis method based on morphological image processing as claimed in claim 1, characterized in that: in the step 5, the images to be detected are sorted according to the area values, and according to the label result, which fault type the image to be detected belongs to is determined, specifically: according to the sequencing result, the fault type at the head of the ranking is a diagnosis result, and two unequal fault grades ranked at the top end are the upper limit and the lower limit of the diagnosis fault grade.

10. The transformer winding fault diagnosis method based on morphological image processing as claimed in claim 1, characterized in that: after the images are acquired, processing the input images and the images in the image library by using an image processing technology, wherein the program flow comprises the following steps:

(1) reading in an image group to be detected and an image library;

(3) inverting the new image and filling the image;

(6) and sorting by taking the area as a criterion.