GB2622910A - A method for identifying partial discharge types based on generated adaptive labels - Google Patents
A method for identifying partial discharge types based on generated adaptive labels Download PDFInfo
- Publication number
- GB2622910A GB2622910A GB2308907.1A GB202308907A GB2622910A GB 2622910 A GB2622910 A GB 2622910A GB 202308907 A GB202308907 A GB 202308907A GB 2622910 A GB2622910 A GB 2622910A
- Authority
- GB
- United Kingdom
- Prior art keywords
- partial discharge
- signals
- data
- network
- labels
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 64
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 18
- 238000012549 training Methods 0.000 claims abstract description 26
- 230000008569 process Effects 0.000 claims abstract description 21
- 238000009499 grossing Methods 0.000 claims abstract description 10
- 238000012544 monitoring process Methods 0.000 claims abstract description 7
- 230000006870 function Effects 0.000 claims description 37
- 230000000694 effects Effects 0.000 claims description 21
- 239000011159 matrix material Substances 0.000 claims description 16
- 230000003595 spectral effect Effects 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 8
- 238000010606 normalization Methods 0.000 claims description 8
- 230000001427 coherent effect Effects 0.000 claims description 6
- 230000009286 beneficial effect Effects 0.000 claims description 4
- 230000008859 change Effects 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 3
- 238000012937 correction Methods 0.000 description 8
- 238000013135 deep learning Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000001228 spectrum Methods 0.000 description 3
- 230000009466 transformation Effects 0.000 description 3
- 230000005540 biological transmission Effects 0.000 description 2
- 238000012806 monitoring device Methods 0.000 description 2
- 230000018199 S phase Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000004587 chromatography analysis Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000004321 preservation Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/50—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
- 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
- G01R31/12—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
-
- 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
- G01R31/12—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
- G01R31/1227—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials
- G01R31/1263—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials of solid or fluid materials, e.g. insulation films, bulk material; of semiconductors or LV electronic components or parts; of cable, line or wire insulation
- G01R31/1272—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials of solid or fluid materials, e.g. insulation films, bulk material; of semiconductors or LV electronic components or parts; of cable, line or wire insulation of cable, line or wire insulation, e.g. using partial discharge measurements
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Power Engineering (AREA)
- Testing Relating To Insulation (AREA)
Abstract
A method for identifying partial discharge types based on generated adaptive labels, comprising: step 1: Receive data of partial discharge fault signals of a transformer through radio monitoring equipment; step 2: Input the fault signals into a classification network to classify and process the partial discharge signals; step 3: Use the input partial discharge data to generate labels corresponding to different types of partial discharges, e.g. generating adaptive labels, and perform phase insensitive training; step 4: Position the classified partial discharge signals by measuring position by the covariance smoothing technique. The partial discharge signals can be classified as different discharge types through the self-learning of the classification network. A unidirectional LTSM training network may be used. A relu or linear rectifier function may be used to activate the data.
Description
Description
A method for identifying partial discharge types based on generated adaptive labels
Technical Field
This invention relates to the field of electronic information technology, particularly to a method for identifying partial discharge types based on generated adaptive labels.
Background
At present, there are various methods and technologies for transformer fault research, such as oil chromatography analysis method, ultrasonic method, ultra high frequency signal monitoring method, deep learning method, etc., which have made certain contributions in the direction of transformer fault research. However, most of the technologies are based on comparing the parametric signals under normal conditions with the fault signals when a fault occurs, so as to identify the working conditions of the transformer and whether a fault occurs. However, there is still relatively little research on the classification of faults, and based on different discharge types it is impossible to classify the collected partial discharge signals according to different discharge types through the self-learning of the classification network, and thereby determine the different fault types.
Invention Summary
The purpose of the present invention is to solve the above-mentioned problems by providing a method for identifying partial discharge types based on generated adaptive labels.
In order to achieve the above object, the technical scheme the present invention takes is:
Description
A method for identifying partial discharge types based on generated adaptive labels, comprising: Step 1: Receive data of partial discharge fault signals of the transformer through a radio monitoring equipment; Step 2: Input the fault signals of the transformer into a classification network to classify and process the partial discharge signals. The partial discharge signals are classified and outputted by an Ego-vise network. The specific steps are as follows: Step 2.1 Collect the partial discharge signals of the transformer through a radio collection equipment as the input data for the classification network; Step 2.2: The partial discharge data is input into the training network. The original data is firstly encoded and processed by the encoder to generate a regular data form that complies with subsequent operations, and then the feature information is extracted from the data through layer normalization operations and convolutional layer; Step 2.3: Segment the data output from the convolutional layer into data blocks with smaller amount of data, which is beneficial for subsequent network training; Step 2.4: Input the segmented and processed data blocks into the unidirectional LSTM training network to extract higher-dimensional features. After being processed by the fully connected layer and layer normalization, the data output by the LSTM is directly added to the data output by the encoder; Step 2.5: The data output after the above-mentioned addition is activated by the Relu function. The convolution layer is used again for feature extraction, and the output data blocks are reassembled into complete data by overlapping and adding operations;
Description
Step 2.6: The reassembled complete data is activated in two branches with the Sigmoid and Tanh functions respectively after passing through the same convolutional layer. The activated data is multiplied and pass through a convolutional layer and the Relu activation function, and then add to the output of the encoder; Step 2.7: Train the network through a certain amount of partial discharge data. When the training effect meets the requirements, input the original partial discharge data to output different partial discharge types, and thereby determine the fault type of the transformer; Step 3: Use the input partial discharge data to generate labels corresponding to different partial discharge types, that is, generating adaptive labels, and perform phase insensitive training. Once the classification effect of the network meets the requirements, the classification of the input partial discharge signals can be realized. Classify the input partial discharge data according to different labels, that is, classifying the fault types; Step 4: Position the classified partial discharge signals of different types respectively by using the method for measuring positioning by the covariance smoothing technique. After the above process, the classification and positioning of fault signals can be realized.
Furthermore, characterized in that the specific generation process of the adaptive labels described in step 3 is as follows: Step 3.1.1: Pass the partial discharge signals of the transformer through the classification network to obtain the corresponding classification results; Step 3.1.2: Based on different discharge types, generate corresponding categories of partial discharge signals from the output of the classification network as labels; Step 3.1.3: After obtaining the labels corresponding to different
Description
discharge types, the classification results can be optimized to improve the classification effect of the network.
Furthermore, characterized in that the specific implementation process of the phase insensitive training described in step 3 is as follows: Step 3.2.1: Adopting the single-variable method, using the same partial discharge signals, change their initial phases to generate partial discharge signals with random initial phases but with the same other parameters, use the generated partial discharge signals as inputs; Step 3.2.2: Input the signals obtained in Step 3.2.1 into the classification network, and train the classification effect of the network under different initial phases; Step 3.2.3: Use fixed labels as standards, and use the difference between the output result and the labels as the loss function for training; Step 3.2.4: When the classification effect meets the expected requirements under different initial phases, the network exhibits phase-insensitive characteristics.
Furthermore, characterized in that the method for measuring positioning by the covariance smoothing technique is as follows: Step 4.1: Use M antenna array elements to receive spatial signals; Step 4.2: Calculate the mutual covariance matrix of the signal data received by different receiving array elements, and then subtract each respective covariance matrix from it. This process can eliminate the irrelevant components in the signals, enhance the useful parts of the signals, and improve the positioning effect; Step 4.3: Perform rank supplementary operations when coherent signals appear to make the signals covariance matrix full rank, which is equal to the number of the sources of the signals, and remove correlation operations to reduce the impact of this coherence;
Description
Step 4.4: Use the eigenvector of the noise space to find the direction of the signals. Use this feature to construct the spatial spectral function, and find the spatial position corresponding to the spectral peak of the spectral function, which is the source direction of the signal, so as to realize the enhanced positioning function of the signals.
Compared with the prior art, the present invention has the following beneficial effects: It is possible to classify the discharge types of partial discharge signals, which is a function that current partial discharge signals monitoring devices do not have. Generally speaking, the data of partial discharge signals of a transformer collected through radio collection devices does not consist of a single discharge type. Instead, the partial discharge signals are formed by the overlapping of two or more types partial discharge signals. How to distinguish the collected electromagnetic signals according to different discharge types to obtain different discharge types of partial discharge signals, so as to determine the fault type? This is a very important technical problem, and the Ego-vise network can solve this problem very well. After the partial discharge signals are classified and output by the Ego-vise network, labels of different discharge types are generated according to the discharge types output by the network classification for network's self-supervised learning, backtracking the determination process of the network, checking the self-determination performance according to the labels generated in the previous step, optimizing the node parameters according to the loss function, thereby improving the classification performance of the network, and finally obtaining partial discharge signal data of different discharge types. Meanwhile combining with discharge Information such as the number of times of discharge, Ego-vise network is able to comprehensively determine the fault conditions of each different discharge
Description
type, so as to realize the monitoring and early warning functions for different types of fault occurrences.
The label samples are derived from the partial discharge signals through the output of the network. This improvement makes the label samples come from the actual partial discharge data, avoiding the big differences between the labels and the actual discharge types which cause an effect similar to overfitting, and makes the performance of the classification of the network is very high. While ensuring to improve the classification and identification ability of the network, this improvement generates different labels according to the collected data, which can better guarantee the quality of the labels and further ensure the reliability of the network classification results.
When classifying and identifying partial discharge signals of different phases, regarding the phase insensitivity of partial discharge signals, when the input of the network is the data of the same discharge type, a large amount of data with consistent partial discharge signals but with different phases is used for training to passivate the network's phase perception of partial discharge signals and form a phase insensitive characteristic in the process of determining the discharge type. When determining the category, the network only focus on the influence of other parameters. When determining the same partial discharge signals with different initial phases, the result is barely affected by the initial phase, which shows that the classification network is mainly aimed at the specific characteristics of different types of partial discharge signals, and excluding the influence of phase. This is more conducive to the accurately determine the type of partial discharge signals.
Since different fault signals have different signal characteristics, the subtle signal characteristics can be extracted by the method of deep learning to form a stable and reliable identification system to identify the types of the received transformer signals. The main theory is that the signals of different
Description
faults are different in the transmission paths, the signal strength and frequency, and the locations that the faults occur, which causes the fault signals generated in different locations to have different characteristics in the time-frequency domain of the signals. The deep learning network is used to capture this difference which can be used as a criterion for identifying the locations of the faults, and can be used to further position the partial discharge signals of the transformer, and determine the location of the fault, thereby accurately prompt the fault points, saving the time for manpower to investigate the causes of the faults, and reducing the probability of accidents.
Description of drawings
Fig. 1 is a schematic diagram of the Ego-vise network of the embodiment of the present invention.
FIG. 2 is a schematic diagram of the discharge type 1 of the embodiment of the present invention.
FIG. 2 is a schematic diagram of the discharge type 2 of the embodiment of the present invention.
FIG. 2 is a schematic diagram of the discharge type 3 of the embodiment of the present invention.
FIG. 2 is a schematic diagram of the discharge type 4 of the embodiment of the present invention.
Embodiments In order to make the technical means, creative features, goals and effects achieved by the present invention easy to understand, the present invention will be further described below in conjunction with specific embodiments.
As shown in Figures 1-5, a partial discharge type identification method based on adaptive label generation, comprising: Step 1: Receive data of partial discharge fault signals of the transformer through a radio monitoring equipment;
Description
Step 2: Input the fault signals of the transformer into a classification network to classify and process the partial discharge signals. The partial discharge signals are classified and outputted by an Ego-vise network. The specific steps are as follows: Step 2.1 Collect the partial discharge signals of the transformer through a radio collection equipment as the input data for the classification network; Step 2.2: The partial discharge data is input into the training network. The original data is firstly encoded and processed by the encoder to generate a regular data form that complies with subsequent operations, and then the feature information is extracted from the data through layer normalization operations and convolutional layer; Step 2.3: Segment the data output from the convolutional layer into data blocks with smaller amount of data, which is beneficial for subsequent network training; Step 2.4: Input the segmented and processed data blocks into the unidirectional LSTM training network to extract higher-dimensional features. After being processed by the fully connected layer and layer normalization, the data output by the LSTM is directly added to the data output by the encoder; Step 2.5: The data output after the above-mentioned addition is activated by the Relu function. The convolution layer is used again for feature extraction, and the output data blocks are reassembled into complete data by overlapping and adding operations; Step 2.6: The reassembled complete data is activated in two branches with the Sigmoid and Tanh functions respectively after passing through the same convolutional layer. The activated data is multiplied and pass through a convolutional layer and the Relu activation function, and then add to the output of the encoder;
Description
Step 2.7: Train the network through a certain amount of partial discharge data. When the training effect meets the requirements, input the original partial discharge data to output different partial discharge types, and thereby determine the fault type of the transformer; Step 3: Use the input partial discharge data to generate labels corresponding to different partial discharge types, that is, generating adaptive labels, and perform phase insensitive training. Once the classification effect of the network meets the requirements, the classification of the input partial discharge signals can be realized. Classify the input partial discharge data according to different labels, that is, classifying the fault types; wherein the specific generation process of the adaptive labels described in step 3 is as follows: Step 3.1.1: Pass the partial discharge signals of the transformer through the classification network to obtain the corresponding classification results; Step 3.1.2: Based on different discharge types, generate corresponding categories of partial discharge signals from the output of the classification network as labels; Step 3.1.3: After obtaining the labels corresponding to different discharge types, the classification results can be optimized to improve the classification effect of the network.
The specific implementation process of the phase insensitive training described in step 3 is as follows: Step 3.2.1: Adopting the single-variable method, using the same partial discharge signals, change their initial phases to generate partial discharge signals with random initial phases but with the same other parameters, use the generated partial discharge signals as inputs;
Description
Step 3.2.2: Input the signals obtained in Step 3.2.1 into the classification network, and train the classification effect of the network under different initial phases; Step 3.2.3: Use fixed labels as standards, and use the difference between the output result and the labels as the loss function for training; Step 3.2.4: When the classification effect meets the expected requirements under different initial phases, the network exhibits phase-insensitive characteristics.
It is possible to classify the discharge types of partial discharge signals, which is a function that current partial discharge signal monitoring devices do not have. Generally speaking, the data of partial discharge signals of a transformer collected through radio collection devices does not consist of a single discharge type. Instead, the partial discharge signals are formed by the overlapping of two or more types partial discharge signals. How to distinguish the collected electromagnetic signals according to different discharge types to obtain different discharge types of partial discharge signals, so as to determine the fault type? This is a very important technical problem, and the Ego-vise network can solve this problem very well. After the partial discharge signals are classified and output by the Ego-vise network, labels of different discharge types are generated according to the discharge types output by the network classification for network's self-supervised learning, backtracking the identification process of the network, checking the self-determination performance according to the labels generated in the previous step, optimizing the node parameters according to the loss function, thereby improving the classification performance of the network, and finally obtaining partial discharge signal data of different discharge types. Meanwhile combining with discharge Information such as the number of times of discharge, Ego-vise network is able to comprehensively determine
Description
the fault conditions of each different discharge type, so as to realize the monitoring and early warning functions for different types of fault occurrences. The Ego-vise classification network, as shown in Figure 1, has the following specific implementation process: 1. Firstly, collect the fault signals from the transformer to form I/Q data as the original data; 2. In the waveform section of the network, the fault signals from the transformer are pulse signals. These pulses appear at specific frequencies on the waveform, while the remaining frequency components are zero. Different fault signals exhibit certain patterns in terms of phase distribution and amplitude of the pulse signals; 3. The original I/Q data is firstly subjected to data preprocessing, where less distinguishable low-quality data is removed. Then, the data is input into the encoder module to undergo waveform encoding, which facilitates network training. The encoded data is normalized using a layer normalization function, constraining the data values to a range between 0 and 1. This is done to prevent the node parameters of the network from growing too large during training, which could impede convergence.
4.The data, after being encoded and layer-normalized, is passed through a convolutional layer to extract feature information. The resulting features are then sliced using a slicing operation, creating smaller segments of data from different batches. These segments are sequentially input into a unidirectional LSTM for processing. As there is correlation between different segments of data, this step involves using an LSTM network to ensure that this correlation is not lost Subsequently, the data passes through fully connected layers and undergoes layer normalization, which is done to prevent the node parameters from growing too large, then the processed data is combined with the waveform-encoded data outputted by the encoder module.
Description
5.The resulting data is then subjected to the activation of a linear rectifier function, allowing the data to propagate further. After passing through another convolutional layer to extract higher-dimensional features, the data blocks are concatenated in a concatenation mode to ensure the optimal utilization of the data. In this process, a dual-branch approach is adopted. Each branch passes through the same convolutional layer but uses different activation functions for activation. In other words, the same data is activated using two different activation functions, resulting in different expressive capacity.
6. After multiplying the two branches, the resulting data is passed through a convolutional layer to extract the relevant feature information. Following activation by a linear rectifier function, it is then added to the waveform-encoded data obtained earlier. The main purpose is to prevent the phenomenon of vanishing gradients caused by having too many layers in the network. Vanishing gradients occur when the gradients become progressively smaller during parameter backpropagation, eventually approaching zero and hindering convergence. This structure helps mitigate such problems. Finally, the data is restored to waveform data through the decoding process of the corresponding decoder. The decoder has two main functions: decoding the data to obtain the corresponding pulse signal waveforms and classifying the decoded data by comparing it with the corresponding labels, resulting in the output of discharge types.
The mentioned labels are generated by feeding the preprocessed waveform signals that after removing low-quality data into the classification, wherein the network utilizes clustering to automatically group the input data into different classes based on their distinct features, and the resulting classes are then adjusted to form labels that correspond to specific discharge types, which are used for comparison and matching purposes.
Since different fault signals have different signal characteristics, the
Description
subtle signal characteristics can be extracted by the method of deep learning to form a stable and reliable identification system to identify the types of the received transformer signals. The main theory is that the signals of different faults are different in the transmission paths, the signal strength and frequency, and the locations that the faults occur, which causes the fault signals generated in different locations to have different characteristics in the time-frequency domain of the signals. The deep learning network is used to capture this difference which can be used as a criterion for identifying the locations of the faults, and can be used to further position the partial discharge signals of the transformer, and determine the location of the fault, thereby accurately prompt the fault points, saving the time for manpower to investigate the causes of the faults, and reducing the probability of accidents. The specific positioning method used is the covariance smoothing technique for positioning measurements.
This method is primarily an improvement upon the MUSIC (Multiple Signal Classification) algorithm, introducing new techniques and methods to overcome the limitations of the traditional MUSIC algorithm and achieve better positioning results. The specific method involves the following steps: during the signal processing of the received data, the cross-covariance matrix between two matrices is calculated. After applying the appropriate transformations, the resulting matrices are averaged with their respective covariance matrices. The next step involves modifying the weighting coefficients of the noise subspace based on the actual noise characteristics in the scenario, generating a new noise subspace. Subsequently, a new spatial pseudospectrum function is obtained using the steps of the spatial spectrum function. Additionally, spatial smoothing techniques are applied to address the limitations of estimation caused by coherent signals.
In situations with low signal-to-noise ratio and a small number of snapshots, using the transformation based on cross-covariance has
Description
significant advantages. The cross-covariance matrix involves the multiplication between two different signals. These two signals, after subtracting the mean, have common components and non-common components. When the common components are multiplied, they always have the same sign, resulting in their enhancement and preservation. On the other hand, the non-common components tend to cancel each other out. In other words, cross-covariance can extract the common components of the two signals while suppressing the different components, and then take the average with the self-covariance matrix of the signals further enhances the statistical significance of the covariance matrix. This can improve the performance of Direction of Arrival (DOA) estimation.
The improved method for calculating noise eigenvalues is as follows: 21 = /1", + a ' , m = A +1, A + 2,-.N Equation 1 wherein 2," represents the initial noise eigenvalues, A' represents the new noise eigenvalues, and a represents the correction factor. In the noise subspace, different eigenvalues correspond to different effects on the spatial spectrum function. The correction factor controls the degree of divergence in the scattered signals and does not affect the vector of signal directions. As a result, it ensures stability in source estimation. In addition, the performance of the algorithm is also closely related to the value of the correction factor. When the correction factor is excessively large, the difference between the noise eigenvalues obtained from the eigenvalue decomposition of the autocorrelation matrix and the signal eigenvalues becomes too small, resulting in a lack of significant resolution.
However, this algorithm is validated under the assumption of mutually independent signal sources. When the antenna array receives coherent signals, new challenges arise. Therefore, it is necessary to combine spatial smoothing techniques with this algorithm to integrate the advantages of both methods for a more reliable and stable DOA estimation. From the
Description
above, it can be inferred that the first crucial step in the improved algorithm is to select an appropriate correction factor. Based on the theoretical foundation of information theory criteria, in a scenario where the number of signal sources can be estimated, the ratio between the maximum and minimum noise eigenvalues should be less than 2. In other words: 2A+1 + aA The specific process of the improved method is as follows: (1) Compute the cross-covariance of the received signal matrix, apply a transformation, and then average the transformed matrices with their respective covariance matrices; (2) Combine spatial smoothing techniques to mitigate the coherent processing of signals received by a uniform linear array; (3) Further decompose the autocorrelation matrix to obtain the noise eigenvalues, and then perform a sorting operation on the eigenvalues; (4) Substitute the eigenvalues into the following equation to determine the minimum integer G that satisfies the inequality: + o. to,u, <2 AN +O1GAN Equation 4 (5) Set the correction factor a = 0.1G which determines the corresponding value of a Based on the obtained eigenvalues A and Equation 4, by substituting 2 into Formula 4 can obtain the minimum integer G that satisfies the inequality, a linear relationship can be established between G and the correction factor, where a linear proportion is determined through <2 Equation 2 By substituting Formula 1 into above equation, it can be inferred: AAA ± a2.4±, Equation 3 Is
Description
experimental testing, and in this case, the linear proportion is determined as 0.1. Therefore, once the value of the minimum integer G is obtained, the correction factor a can be determined accordingly; (6) Further, the corresponding spatial spectrum function can be obtained, ultimately determining the estimated direction of arrival for the signals.
The improved covariance smoothing technique algorithm addresses the limitations of the traditional MUSIC algorithm, which is only effective in high signal-to-noise ratio and high snapshot scenarios. It also performs well in low signal-to-noise ratio and low snapshot situations. Furthermore, it exhibits better performance in fault signal positioning for transformers.
Step 1: Use M antenna array elements to receive spatial signals; Step 2: Calculate the mutual covariance matrix of the signal data received by different receiving array elements, and then subtract each respective covariance matrix from it. This process can eliminate the irrelevant components in the signals, enhance the useful parts of the signals, and improve the positioning effect; Step 3: Perform rank supplementary operations when coherent signals appear to make the signal covariance matrix full rank, which is equal to the number of signal sources, and remove correlation operations to reduce the impact of this coherence; Step 4: Use the eigenvector of the noise space to find the direction of the signals. Use this feature to construct the spatial spectral function, and find the spatial position corresponding to the spectral peak of the spectral function, which is the source direction of the signal, so as to realize the enhanced positioning function of the signals.
For those skilled in the art, it is evident that the present invention is not limited to the details of the exemplary embodiments described above. The
Description
invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the exemplary embodiments should be considered illustrative and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the above description. Thus, all changes falling within the meaning and scope of the claims, as well as equivalents of the claims, are intended to be embraced within the scope of the invention. Any reference signs in the claims should not be construed as limiting the claims in any way.
Furthermore, it should be understood that although the present specification describes the embodiments in a particular manner, each embodiment may not necessarily be limited to a single independent technical solution. The description in the specification is provided for clarity purposes only. Those skilled in the art should consider the specification as a whole, and the technical solutions in various embodiments can be appropriately combined to form other embodiments that would be understood by those skilled in the art.
Claims (4)
- Claims 1. A method for identifying partial discharge types based on generated adaptive labels, comprising: Step 1: Receive data of partial discharge fault signals of a transformer through a radio monitoring equipment; Step 2: Input the fault signals of the transformer into a classification network to classify and process the partial discharge signals. The partial discharge signals are classified and outputted by an Ego-vise network. The specific steps are as follows: Step 2.1 Collect the partial discharge signals of the transformer through a radio collection equipment as the input data for the classification network; Step 2.2: The partial discharge data is input into the training network. The original data is firstly encoded and processed by the encoder to generate a regular data form that complies with subsequent operations, and then the feature information is extracted from the data through layer normalization operations and convolutional layer; Step 2.3: Segment the data output from the convolutional layer into data blocks with smaller amount of data, which is beneficial for subsequent network training; Step 2.4: Input the segmented and processed data blocks into the unidirectional LSTM training network to extract higher-dimensional features. After being processed by the fully connected layer and layer normalization, the data output by the LSTM is directly added to the data output by the encoder; Step 2.5: The data output after the above-mentioned addition is activated by the Relu function. The convolution layer is used again for feature extraction, and the output data blocks are reassembled into complete data by overlapping and adding operations; Claims Step 2.6: The reassembled complete data is activated in two branches with the Sigmoid and Tanh functions respectively after passing through the same convolutional layer. The activated data is multiplied and pass through a convolutional layer and the Relu activation function, and then add to the output of the encoder; Step 2.7: Train the network through a certain amount of partial discharge data. When the training effect meets the requirements, input the original partial discharge data to output different partial discharge types, and thereby determine the fault type of the transformer; Step 3: Use the input partial discharge data to generate labels corresponding to different partial discharge types, that is, generating adaptive labels, and perform phase insensitive training. Once the classification effect of the network meets the requirements, the classification of the input partial discharge signals can be realized. Classify the input partial discharge data according to different labels, that is, classifying the fault types; Step 4: Position the classified partial discharge signals of different types respectively by using the method for measuring positioning by the covariance smoothing technique.
- 2. A method for identifying partial discharge types based on generated adaptive labels as claimed in claim 1, characterized in that: the specific generation process of the adaptive label described in step 3 is as follows: Step 3.1.1: Pass the partial discharge signals of the transformer through the classification network to obtain the corresponding classification results; Step 3.1.2: Based on different discharge types, generate corresponding categories of partial discharge signals from the output of the classification network as labels; Claims Step 3.1.3: After obtaining the labels corresponding to different discharge types, the classification results can be optimized to improve the classification effect of the network.
- 3. A method for identifying partial discharge types based on generated adaptive labels, as claimed in claim 1, characterized in that: the specific implementation process of the phase insensitive training described in step 3 is as follows: Step 3.2.1: Adopting the single-variable method, using the same partial discharge signals, change their initial phases to generate partial discharge signals with random initial phases but with the same other parameters, use the generated partial discharge signals as inputs; Step 3.2.2: Input the signals obtained in Step 3.2.1 into the classification network, and train the classification effect of the network under different initial phases; Step 3.2.3: Use fixed labels as standards, and use the difference between the output result and the labels as the loss function for training; Step 3.2.4: When the classification effect meets the expected requirements under different initial phases, the network exhibits phase-insensitive characteristics.
- 4. A method for identifying partial discharge types based on generated adaptive labels, as claimed in claim 1, characterized in that: the method for measuring positioning by the covariance smoothing technique is as follows: Step 4.1: Use M antenna array elements to receive spatial signals; Step 4.2: Calculate the mutual covariance matrix of the signal data received by different receiving array elements, and then subtract each respective covariance matrix from it. This process can eliminate the irrelevant components in the signals, enhance the useful parts of the signals, and improve the positioning effect; Claims Step 4.3: Perform rank supplementary operations when coherent signals appear to make the signals covariance matrix full rank, which is equal to the number of the sources of the signals, and remove correlation operations to reduce the impact of this coherence; Step 4.4: Use the eigenvector of the noise space to find the direction of the signals. Use this feature to construct the spatial spectral function, and find the spatial position corresponding to the spectral peak of the spectral function, which is the source direction of the signal, so as to realize the enhanced positioning function of the signals.
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211205399.8A CN115267462B (en) | 2022-09-30 | 2022-09-30 | Partial discharge type identification method based on self-adaptive label generation |
Publications (3)
| Publication Number | Publication Date |
|---|---|
| GB202308907D0 GB202308907D0 (en) | 2023-07-26 |
| GB2622910A true GB2622910A (en) | 2024-04-03 |
| GB2622910A8 GB2622910A8 (en) | 2024-04-17 |
Family
ID=83758127
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| GB2308907.1A Pending GB2622910A (en) | 2022-09-30 | 2023-06-14 | A method for identifying partial discharge types based on generated adaptive labels |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN115267462B (en) |
| GB (1) | GB2622910A (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20210046356A (en) * | 2019-10-18 | 2021-04-28 | 한전케이디엔주식회사 | Diagnosis apparatus of partial discharge |
| CN114035007A (en) * | 2021-11-22 | 2022-02-11 | 国家电网有限公司 | Non-contact transformer bushing partial discharge detection system |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107907799B (en) * | 2017-11-10 | 2020-10-09 | 国网浙江省电力公司电力科学研究院 | Method and system for identifying partial discharge defect type based on convolutional neural network |
| KR101958903B1 (en) * | 2018-05-31 | 2019-03-15 | 수원대학교산학협력단 | PARTIAL DISCHARGE CLASSIFICATION SYSTEM USING RBFNNs(RADIAL BASIS FUNCTION Neural NETWORKS) MODEL |
| CN109116203A (en) * | 2018-10-31 | 2019-01-01 | 红相股份有限公司 | Power equipment partial discharges fault diagnostic method based on convolutional neural networks |
| CN110309010B (en) * | 2019-06-05 | 2022-08-19 | 国网江苏省电力有限公司电力科学研究院 | Partial discharge network training method and device for phase resolution of power equipment |
| CN110412431A (en) * | 2019-08-05 | 2019-11-05 | 国网湖南省电力有限公司 | A kind of diagnostic method and diagnostic system of the shelf depreciation defect type of power equipment |
| CN113065522A (en) * | 2021-04-26 | 2021-07-02 | 哈尔滨理工大学 | Transformer partial discharge type identification method based on deep neural network |
| CN114417926A (en) * | 2022-01-19 | 2022-04-29 | 山东大学 | Power equipment partial discharge pattern recognition method and system based on deep convolution generation countermeasure network |
| CN114839492A (en) * | 2022-05-24 | 2022-08-02 | 国网江苏省电力有限公司电力科学研究院 | Method and device for identifying GIS partial discharge type based on MOBILE NETV3 |
-
2022
- 2022-09-30 CN CN202211205399.8A patent/CN115267462B/en active Active
-
2023
- 2023-06-14 GB GB2308907.1A patent/GB2622910A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20210046356A (en) * | 2019-10-18 | 2021-04-28 | 한전케이디엔주식회사 | Diagnosis apparatus of partial discharge |
| CN114035007A (en) * | 2021-11-22 | 2022-02-11 | 国家电网有限公司 | Non-contact transformer bushing partial discharge detection system |
Also Published As
| Publication number | Publication date |
|---|---|
| GB202308907D0 (en) | 2023-07-26 |
| GB2622910A8 (en) | 2024-04-17 |
| CN115267462B (en) | 2022-12-23 |
| CN115267462A (en) | 2022-11-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Jin et al. | A novel adaptive EEMD method for switchgear partial discharge signal denoising | |
| CN108469560B (en) | Electromagnetic interference objective complexity evaluation method based on rapid S-transform time-frequency space model | |
| CN112834878B (en) | Transformer partial discharge defect type identification method and system | |
| Liu et al. | Dual-channel convolutional network-based fault cause identification for active distribution system using realistic waveform measurements | |
| Yang et al. | Partial discharge signal extraction method based on EDSSV and low rank RBF neural network | |
| Tang et al. | Jamming recognition based on AC-VAEGAN | |
| Chen et al. | Automatic modulation classification of radar signals utilizing X-net | |
| CN107356843A (en) | The partial discharge of transformer method for diagnosing faults of small echo is synchronously extruded based on gradient threshold | |
| CN104614647A (en) | Complex wavelet transform partial discharge location test method and device | |
| CN112183225A (en) | Underwater target signal feature extraction method based on probability latent semantic analysis | |
| CN116189711B (en) | Transformer fault identification method and device based on acoustic wave signal monitoring | |
| CN108594156B (en) | Improved current transformer saturation characteristic identification method | |
| Deng et al. | Selection of optimal wavelet basis for signal denoising | |
| GB2622910A (en) | A method for identifying partial discharge types based on generated adaptive labels | |
| Chen et al. | Unlocking signal processing with image detection: A frequency hopping detection scheme for complex EMI environments using STFT and CenterNet | |
| CN108761202A (en) | The harmonic detecting method that pole symmetric mode decomposition and Hilbert transform are combined | |
| CN116667920A (en) | Transient signal detection method combining principal component analysis and Hilbert-Huang transform | |
| Limin et al. | Low probability of intercept radar signal recognition based on the improved AlexNet model | |
| CN113158714B (en) | Fault feature self-adaptive extraction method based on wavelet entropy and EEMD | |
| Madhu et al. | A novel algorithm for denoising of simulated partial discharge signals using adaptive wavelet thresholding methods | |
| CN115345216A (en) | FMCW radar interference elimination method fusing prior information | |
| CN114113837A (en) | Acoustic feature-based transformer live-line detection method and system | |
| Wang et al. | Research on speech enhancement based on deep neural network | |
| Govalkar et al. | Siamese Network based Pulse and Signal Attribute Identification | |
| Quadri | A review of noise cancellation techniques for cognitive radio |