Low-frequency Discharge Type Recognition Model Based on Period Normalization, and Its Training Method and System

Abstract

The present invention belongs to the technical field of partial discharge classification and recognition, and discloses a low-frequency discharge type recognition model based on period normalization, its training method and system, so as to solve the problem of the blank in low-frequency discharge type recognition. The training method includes: obtaining a power frequency discharge data set and a low-frequency discharge data set, as well as the offset data of the low-frequency discharge compared with the power frequency discharge in terms of phase, amplitude and discharge time; performing period normalization correction on the data in the power frequency discharge data set based on the offset data to obtain a low-frequency discharge type recognition sample set; constructing a discharge type recognition neural network and inputting each sample data in the low-frequency discharge type recognition sample set to train the network and output the recognition result corresponding to each sample data; obtaining a preset weight vector corresponding to each recognition result, and performing recognition result fusion based on each recognition result and its corresponding weight vector. The present invention fills the blank in low-frequency discharge type recognition and improves the efficiency and accuracy of low-frequency discharge recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of partial discharge classification and recognition, and particularly relates to a low-frequency discharge type recognition model based on period normalization, and a training method and system thereof. Background Art

[0002] The phenomenon of partial discharge is commonly found in areas with weak insulation and non-uniform electric field distribution. Its characteristics are distinct, manifested as nanosecond-level short intermittent pulses, and these pulses instantaneously release electromagnetic waves in a specific frequency band (300 MHz to 3 GHz). Detecting partial discharge is of multiple important significances for maintaining the safety of power equipment. It can effectively prevent equipment failures, extend the service life of equipment, ensure the stable operation of the power system, and promote the continuous progress and innovation of related technologies. Therefore, in the operation and maintenance management of power equipment, high attention should be paid to the detection of partial discharge.

[0003] Currently, in the standard power frequency (50 Hz) environment, various live detection technologies such as the ultra-high frequency method, high frequency method, ultrasonic method, and infrared method have been widely used in the detection of partial discharge. However, for the research on the characteristics of partial discharge under low frequency (such as 20 Hz), there is still a lack of sufficient engineering practice experience and unified reference standards at home and abroad. Flexible low-frequency AC power transmission is a new and efficient power transmission technology based on power electronics technology, which can improve the transmission capacity and flexible regulation ability of the power grid by reducing the transmission frequency, reducing the line impedance, and reducing the charging reactive power of cable lines. It has broad application prospects in scenarios such as power grid interconnection, medium- and long-distance offshore wind power transmission, and island power supply. Due to the fact that the characteristics of partial discharge under low-frequency environment may be very different from those under power frequency conditions, it is difficult to directly migrate and apply the traditional discharge type recognition methods based on power frequency data. In addition, the applicability of the discharge type recognition sample library accumulated under power frequency and its specific application method under low-frequency environment still need to be deeply explored and verified. Summary of the Invention

[0004] Based on the above-mentioned drawbacks and deficiencies existing in the prior art, one of the purposes of the present invention is to at least solve one or more of the above problems existing in the prior art. In other words, one of the purposes of the present invention is to provide a low-frequency discharge type recognition model based on period normalization, and a training method and system thereof that meet one or more of the foregoing requirements, so as to fill the gap in the recognition of partial discharge types under low-frequency environment, and strive to improve the efficiency and accuracy of the recognition of partial discharge types under low-frequency environment, thereby bringing substantial optimization and improvement to the safety monitoring and maintenance work of power equipment.

[0005] To achieve the above-mentioned invention purpose, the present invention adopts the following technical solutions:

[0006] In the first aspect, the present invention provides a training method for a low-frequency discharge type recognition model based on period normalization, including the steps:

[0007] S1. Obtain the power frequency discharge data set and the low frequency discharge data set;

[0008] S2. Based on the power frequency discharge data set and the low frequency discharge data set, obtain the offset data of the low frequency discharge compared with the power frequency discharge in terms of phase, amplitude and discharge time;

[0009] S3. Based on the offset data, perform cycle normalization correction on the data in the power frequency discharge data set to obtain a low frequency discharge type recognition sample set, and the low frequency discharge type recognition sample set includes PRPD map sample data, map sample data and characteristic parameter sample data;

[0010] S4. Construct a discharge type recognition neural network and input each sample data in the low frequency discharge type recognition sample set into it to train the network and output the recognition result corresponding to each sample data;

[0011] S5. Obtain the preset weight vectors corresponding to each of the recognition results, and perform recognition result fusion based on each of the recognition results and their corresponding weight vectors to complete the training of the low frequency discharge type recognition model.

[0012] As a preferred solution, the cycle normalization correction of the data in the power frequency discharge data set based on the offset data in step S3 includes performing normalization correction on the phase data in the power frequency discharge data, specifically:

[0013] Obtain the phases corresponding to each discharge type in the power frequency discharge data ;

[0014] Obtain the phases corresponding to each discharge type in the low frequency discharge data ;

[0015] Based on the and the corresponding to the same discharge type, perform difference calculation to obtain the phase deviation corresponding to each discharge type , ;

[0016] Based on the and the phase deviation corresponding to the same discharge type perform calculation to obtain the correction value for normalizing and correcting the phases corresponding to each discharge type in the power frequency discharge data , .

[0017] As a preferred solution, the periodic normalization correction of the data in the power frequency discharge dataset based on the offset data in step S3 includes normalizing and correcting the amplitude data of each discharge data sample, specifically:

[0018] Obtain the discharge amplitude of each discharge data sample V ;

[0019] Obtain all the discharge amplitudes V The maximum amplitude value V max And the minimum amplitude value V min ;

[0020] Based on the discharge amplitude V、 The maximum amplitude value V max And the minimum amplitude value V min Calculate to obtain the correction value for normalizing and correcting the amplitude in each discharge data sample V 修正 , .

[0021] As a preferred solution, the periodic normalization correction of the data in the power frequency discharge dataset based on the offset data in step S3 includes normalizing and correcting the discharge time interval data in the power frequency discharge data, specifically:

[0022] Obtain the preset correction coefficient k;

[0023] Obtain the discharge time interval corresponding to each discharge type in the power frequency discharge data ;

[0024] Based on the correction coefficient k and each discharge time interval , obtain the correction value for normalizing and correcting the discharge time interval corresponding to each discharge type in the power frequency discharge data , .

[0025] As a preferred solution, the characteristic parameter sample data includes the mean value of the discharge time interval , the standard deviation of the discharge time interval , the entropy value of the discharge time interval , the Weibull parameters of the discharge times - amplitude (N - V), the mean value of the amplitude , the standard deviation of the amplitude , the entropy value of the amplitude .

[0026] As a preferred solution, the step of constructing a discharge type recognition neural network and inputting each sample data in the low-frequency discharge type recognition sample set thereto to train the network and output a recognition result corresponding to each sample data includes the steps of:

[0027] Construct a PRPD pattern network, a pattern network, and a BP neural network;

[0028] Divide the low-frequency discharge type recognition sample set into a training data set and a validation data set, and the ratio of the training data set to the validation data set corresponding to the same sample data is 4:1;

[0029] Input the PRPD pattern sample data, pattern sample data, and feature parameter sample data in the low-frequency discharge type recognition sample set into the corresponding networks respectively to train the networks and output corresponding training set confusion matrices and validation set confusion matrices.

[0030] As a preferred solution, the step of constructing the PRPD pattern network or the pattern network includes the steps of:

[0031] Set multiple layers of neurons to form an input layer, an output layer, and an intermediate layer;

[0032] Set module A, module B, and module C in the intermediate layer, and set 1×1 convolutions for adjusting the number of channels of the input features in module A, module B, and module C.

[0033] As a preferred solution, the step of obtaining a preset weight vector corresponding to each of the recognition results and performing recognition result fusion based on each of the recognition results and its corresponding weight vector includes the steps of:

[0034] Obtain a preset weight vector corresponding to the PRPD pattern training set confusion matrix Q i PRPD ;

[0035] Obtain a preset weight vector corresponding to the pattern training set confusion matrix ;

[0036] Obtain a preset weight vector corresponding to the feature parameter training set confusion matrix Q k BP ;

[0037] Based on the Q i PRPD , and Qk BP Calculate the fused recognition vector Q and select Q the index value with the highest probability in as the discharge recognition result;

[0038] The calculation formula of the recognition vector Q is:

[0039] ,

[0040] wherein Q i PRPD represents the weight vector of a certain input sample recognized as the i-th type of discharge by the PRPD pattern network, represents that a certain input sample passes through the weight vector of the pattern network recognized as the j-th type of discharge, Q k BP represents the weight vector of a certain input sample recognized as the k-th type of discharge by the BP neural network.

[0041] In a second aspect, the present invention provides a training system for a low-frequency discharge type recognition model based on cycle normalization, based on the method for training a low-frequency discharge type recognition model according to the first aspect:

[0042] It includes an acquisition module, a cycle normalization correction module, a construction module, a training module, and a result fusion module;

[0043] The acquisition module is used to acquire the power frequency discharge data set and the low-frequency discharge data set, as well as the offset data of the low-frequency discharge compared with the power frequency discharge in terms of phase, amplitude, and discharge time;

[0044] The cycle normalization correction module performs cycle normalization correction on each discharge data based on the offset data to obtain a low-frequency discharge type recognition sample set, and the low-frequency discharge type recognition sample set includes PRPD pattern sample data, pattern sample data, and characteristic parameter sample data;

[0045] The construction module is used to construct a discharge type recognition neural network;

[0046] The training module is used to input each sample data in the low-frequency discharge type recognition sample set to train the network and output the recognition result corresponding to each sample data;

[0047] The acquisition module is further used to acquire the preset weight vectors corresponding to each of the recognition results;

[0048] The result fusion module performs recognition result fusion based on each of the recognition results and their corresponding weight vectors to complete the training of the low-frequency discharge type recognition model.

[0049] In a third aspect, the present invention provides a low-frequency discharge type recognition model based on cycle normalization, which is trained by using the training method described in the first aspect.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. By performing cycle normalization on the discharge type recognition sample library accumulated at the power frequency (50 Hz), the application of the power frequency database at the low frequency (20 Hz) is made more accurate and reliable, reducing the need to re-establish a low-frequency sample library and write a new recognition program, and improving work efficiency and cost-effectiveness.

[0052] 2. The introduction of the phase normalization processing method effectively solves the phase shift problem in the discharge data collected at different frequencies, ensures that the recognition model can accurately process low-frequency data, eliminates the influence of frequency changes on the accuracy of phase difference recognition, improves the adaptability and accuracy of the recognition model, and can be further extended to the recognition of partial discharge types at other voltage frequencies.

[0053] 3. A recognition method of multi-network recognition result fusion is proposed. Combining the complementary advantages of the PRPD pattern, pattern and various characteristic parameters, through weight allocation and result fusion, the possible misjudgment or limitations of a single network are effectively reduced, the overall recognition error rate is significantly reduced, and the comprehensive accuracy and robustness of the recognition are improved.

[0054] 4. The constructed PRPD pattern network and pattern network adopt an architecture with multiple modules and multiple structural levels, including a branch structure and a residual structure. This design not only reduces the total number of parameters in the network, but also increases the depth of the network, can extract more types of abstract features and high-order features while retaining the original input features, enhances the learning ability and recognition ability of the network, and improves the efficiency and accuracy of discharge pattern recognition.

[0055] Further or more detailed beneficial effects will be described in combination with specific embodiments in the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0057] Figure 1 It is a schematic flow chart of the low-frequency discharge type recognition model training method described in Embodiment 1 of the present invention.

[0058] Figure 2 It is a schematic diagram of the UHF detection of partial discharge in GIS of the prior art described in Embodiment 1 of the present invention.

[0059] Figure 3 It is a schematic diagram of the air gap discharge collected by the UHF partial discharge detector described in Embodiment 1 of the present invention, where a is the PRPD pattern, b is the pattern, and c is the pattern.

[0060] Figure 4 It is the prior art described in Embodiment 1 of the present invention using Figure 3 and a convolutional neural network (CNN) for the recognition of partial discharge types under power frequency.

[0061] Figure 5 It is a schematic diagram of the classical BP neural network described in Embodiment 1 of the present invention.

[0062] Figure 6 It is a schematic diagram of the convolutional neural network constructed for discharge type recognition in Embodiment 1 of the present invention, where a is the overall structure diagram, b is the structure diagram of module A, c is the structure diagram of module B, and d is the structure diagram of module C.

[0063] Figure 7 It is in Embodiment 1 of the present invention based on Figure 6 The confusion matrix of the validation set obtained by training the shown convolutional neural network and the power frequency discharge data set, where a is the confusion matrix of the PRPD pattern validation set and b is the confusion matrix of the pattern validation set.

[0064] Figure 8 It is the result of directly recognizing the discharge data at low frequency after training the convolutional neural network shown based on the power frequency PRPD pattern and the pattern data in Embodiment 1 of the present invention, where a is the recognition confusion matrix of the PRPD pattern and b is the Figure 6 confusion matrix of the pattern recognition. confusion matrix of the pattern recognition.

[0065] Figure 9 It is in Embodiment 1 of the present invention based on Figure 5 The confusion matrix of the BP neural network shown, which is trained by the power frequency discharge characteristic parameter data set, for recognizing the power frequency data validation set and the low-frequency discharge data, where a is the confusion matrix of the 50Hz characteristic parameter validation set and b is the confusion matrix of the 20Hz characteristic parameter recognition.

[0066] Figure 10 It is a schematic flowchart of the method for training the low-frequency discharge type recognition model described in the first embodiment of the present invention.

[0067] Figure 11 It is a schematic diagram of the cycle normalization effect described in the first embodiment of the present invention, where a, b, and c are the PRPD diagrams of spike discharges under power frequency, the PRPD diagrams of spike discharges at low frequencies, and the PRPD diagrams of corrected spike discharges, respectively.

[0068] Figure 12 It is a schematic diagram of the result obtained by training based on the PRPD diagram sample data in the low-frequency discharge type recognition sample set and the PRPD diagram network described in the first embodiment of the present invention, where a is the result of the training set and b is the result of the validation set.

[0069] Figure 13 It is a schematic diagram of the result of training the PRPD diagram network to recognize the low-frequency discharge data set based on the low-frequency discharge type recognition sample set described in the first embodiment of the present invention.

[0070] Figure 14 It is based on the low-frequency discharge type recognition sample set in the first embodiment of the present invention diagram sample data and the diagram network for training, where a is the result of the training set and b is the result of the validation set.

[0071] Figure 15 It is based on the training of the low-frequency discharge type recognition sample set in the first embodiment of the present invention diagram network to recognize the low-frequency discharge data set.

[0072] Figure 16 It is a schematic diagram of the confusion matrix of the validation set and its recognition result for the low-frequency discharge data set after training based on the feature parameter sample data in the low-frequency discharge type recognition sample set and the BP neural network described in the first embodiment of the present invention, where a is the result of the validation set and b is the result of the low-frequency discharge data set.

[0073] Figure 17 It is a schematic diagram of the confusion matrix of the PRPD diagram training network described in the first embodiment of the present invention.

[0074] Figure 18 It is a schematic diagram of the fusion result described in the first embodiment of the present invention. Detailed implementation manners

[0075] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention.

[0076] In the following description, multiple embodiments of the present invention are provided, and different embodiments can be replaced or combined. Therefore, the present invention can also be considered to include all possible combinations of the same and / or different embodiments described. Thus, if one embodiment includes features A, B, and C, and another embodiment includes features B and D, then the present invention should also be considered to include embodiments that contain one or more of all other possible combinations of A, B, C, and D, even though such embodiments may not be explicitly described in the following content in words.

[0077] The following description provides examples and does not limit the scope, applicability, or examples set forth in the claims. Changes can be made to the functions and arrangements of the described elements without departing from the scope of the content of the present invention. Various processes or components can be appropriately omitted, substituted, or added to each example. For example, the described method can be executed in a different order than the described order, and various steps can be added, omitted, or combined. In addition, the features described in some examples can be combined into other examples.

[0078] To facilitate a better understanding of the embodiments of the present invention, before explaining the specific implementation manners of the present invention in detail, its application scenarios will be described first.

[0079] The low-frequency discharge type recognition model based on period normalization, its training method, and system described in the embodiments of this specification are applied to the fault diagnosis and maintenance process of power equipment. In these scenarios, the application of the low-frequency discharge type recognition model based on period normalization, its training method, and system aims to make the recognition of the partial discharge type of power equipment more accurate and efficient, so that the fault diagnosis is more timely and accurate, in order to provide a high-quality user experience and reliable performance.

[0080] Embodiment 1:

[0081] As Figure 1 shown, this embodiment provides a training method for a low-frequency discharge type recognition model based on period normalization, including the steps of:

[0082] S1. Obtain a power frequency discharge data set and a low-frequency discharge data set;

[0083] S2. Based on the power frequency discharge data set and the low-frequency discharge data set, obtain the offset data of the low-frequency discharge compared to the power frequency discharge in terms of phase, amplitude, and discharge time;

[0084] S3. Based on the offset data, perform period normalization correction on the data in the power frequency discharge data set to obtain a low-frequency discharge type recognition sample set, and the low-frequency discharge type recognition sample set includes PRPD map sample data, map sample data, and characteristic parameter sample data;

[0085] S4. Construct a neural network for identifying discharge types and input the sample data in the low-frequency discharge type identification sample set into it to train the network and output the identification results corresponding to each sample data;

[0086] S5. Obtain the preset weight vectors corresponding to each of the identification results, and perform identification result fusion based on each of the identification results and their corresponding weight vectors to complete the training of the low-frequency discharge type identification model.

[0087] It can be understood that partial discharge usually occurs between regions with a relatively low insulation level and an uneven electric field. This kind of discharge is intermittent, with an extremely short duration of nanosecond level, and the pulsed signal generated during the discharge process will release electromagnetic waves of specific frequencies, usually between 300 MHz and 3 GHz. As Figure 2 shown, existing technologies such as the ultra-high frequency partial discharge detection method (UHF) use ultra-high frequency receiving devices (sensors) to detect and analyze these electromagnetic waves. By means of an appropriate A / D conversion device, the pulsed signal of partial discharge can be effectively reproduced, thereby improving the detection accuracy and reliability. Usually, the discharge data processed by an ultra-high frequency partial discharge detector retains the amplitude, phase, and time information of partial discharge, and can create various discharge patterns for subsequent discharge type identification, such as the PRPD pattern (as shown in a of Figure 3 ), pattern (as shown in b of Figure 3 ), pattern (as shown in c of Figure 3 ), characteristic parameters, etc. After machine learning learns the characteristics of these pictures and numerical values, it can realize the identification of discharge types. The existing technology uses the above-mentioned discharge patterns and convolutional neural network (CNN) to identify the partial discharge type under power frequency, and the process is as shown in Figure 4 . Specifically, the commonly used characteristic parameters based on the PRPD pattern such as the pulse number distribution , average discharge amount distribution, and their distributions are different on the positive and negative half axes of the alternating voltage, and , , , can also be obtained. In the existing methods, some also extract characteristic parameters from the patterns and introduce the skewness S k , steepness K u , the asymmetry between the positive and negative half cycles A sy , and the similarity of the pattern contours between the positive and negative half cycles C c. When using characteristic parameters to identify the discharge type, these characteristic parameters are classified using an artificial neural network (BP neural network). The structure diagram of the neural network is as shown in Figure 5 . A typical BP neural network generally consists of three layers, namely the input layer, output layer, and intermediate hidden layer shown in Figure 5 . Usually, it is necessary to first tell the computer which information is a feature. After calculating the characteristic parameters, the number of neurons in the input layer is the same as theirs. The hidden layer is set according to the categories to be classified and the types of characteristic parameters. The number of neurons in the output layer is set to the number of categories to be classified. After extracting the characteristic parameters and building the network, the training set and validation set can be set, and the backpropagation algorithm can be used to train the BP neural network model. According to the evaluation results, the hyperparameters of the neural network, such as the learning rate, regularization parameter, etc., are adjusted to improve the generalization ability of the model.

[0088] In this embodiment, based on the classic neural network models of InceptionV3 and Resnet, a convolutional neural network for discharge type recognition is built as shown in Figure 6 . Based on this convolutional neural network, a special high-frequency partial discharge detector is used to collect four types of GIS partial discharges under power frequency. The corresponding labels from 0 to 3 are floating discharge, air gap discharge, tip discharge, and surface discharge. The processed PRPD maps and maps are respectively sent into the above convolutional neural network for training. The confusion matrix of the obtained validation set is as shown in Figure 7 . Figure 7 In the confusion matrix, the horizontal axis represents the discharge categories predicted by the convolutional neural network, and the vertical axis represents the actual discharge categories of the discharges. Therefore, only the numbers on the diagonal are the number of correctly identified samples. From the above two figures, it can be obtained that after being identified by the above convolutional neural network, the correct rates of the four discharge types identified by the PRPD map are 99.18%, 97.78%, 97%, and 97.62% respectively. The correct rates of the Figure 7 map are 99.78%, 99.38%, 92.69%, and 94.76% respectively. It can be seen from Figure 8 that the correct rate of this convolutional neural network is relatively high under power frequency. Next, the sample maps of the four discharge types at a low frequency of 20 Hz are directly input into the above-trained network, and the recognition results are as shown in Figure 8 . Among them, a is the PRRP map, and b is the result of the Figure 8 map. It can be seen from The atlas shapes are similar, but the network directly trained with power frequency data performs poorly in directly identifying low-frequency discharge atlases. From the recognition results of the PRPD atlas, for the 0th category (suspended discharge) and the 1st category (air gap discharge) at low frequencies, misidentifications are caused by phase shifts. The same problem also exists for the 2nd category (tip discharge) and the 3rd category (surface discharge). The atlas recognition effect is also poor, and many 0th category discharges are misidentified as 1st category discharges. It can be seen from this that the network trained with the accumulated power frequency discharge data is not effective in directly diagnosing the partial discharge types of low-frequency power transmission and transformation equipment. If the diagnostic program is trained with the data accumulated at low frequencies, there is a problem of insufficient accumulated samples in the newly put into operation low-frequency power transmission project.

[0089] In this embodiment, a method of fusing the recognition results of multiple networks is creatively proposed to improve the diagnosis of low-frequency discharge types. Based on the discharge data collected by the UHF partial discharge detector, 26 characteristic parameters are extracted. The BP neural network used is the Multi-LayerPerceptron Classifier. The number of neurons in the input layer is 26, there is one hidden layer with 13 neurons, and the output is the number of discharge types, which is 4. The BP neural network built above is used to train the discharge data at power frequency as Figure 9 shown, where a is the confusion matrix of the 50Hz characteristic parameter verification set, and b is the confusion matrix of the 20Hz characteristic parameter recognition. By calculation, it is obtained that the above BP neural network is used to classify the extracted discharge characteristic parameters. The recognition accuracies of the verification set are 95.91%, 93.89%, 96.57%, and 98.10% respectively, and the recognition accuracies of the discharge data at 20Hz are 98.65%, 93.15, 75.49%, and 42.57% respectively. Since the number of extracted features is 26, and the statistical features of the characteristic parameters are considered, the influence degree of frequency on the recognition accuracy of the above neural network is different for different discharge types.

[0090] In summary, the various GIS discharge data accumulated at the original power frequency and the discharge type recognition network based on deep learning are directly used to identify the discharge types at the low frequency of 20Hz, and the effect is not good. In the existing low-frequency power transmission project, the identification of GIS partial discharge types faces the problems of redesigning the neural network and accumulating the discharge sample library, with a large amount of work and a long time. Therefore, this embodiment proposes a training method for a low-frequency discharge type recognition model based on cycle normalization, and its process is as Figure 10 shown.

[0091] Specifically, this embodiment provides a preferred implementation manner of step S3. The cycle normalization correction of the data in the power frequency discharge data set based on the offset data includes performing normalization correction on the phase data in the power frequency discharge data, specifically:

[0092] Obtain the phases corresponding to each discharge type in the power frequency discharge data ;

[0093] Obtain the phases corresponding to each discharge type in the low frequency discharge data ;

[0094] Based on the and the corresponding to the same discharge type, perform a difference calculation to obtain the phase deviation corresponding to each discharge type , ;

[0095] Based on the and the phase deviation corresponding to the same discharge type perform a calculation to obtain a correction value for normalizing and correcting the phases corresponding to each discharge type in the power frequency discharge data , .

[0096] More specifically, in this embodiment, based on experiments, the discharge characteristics of typical partial discharges of GIS at 50HZ and 20Hz are obtained. From the existing experimental results, it can be seen that for the typical partial discharges of GIS equipment under power frequency, the partial discharges of GIS at low frequency have the characteristics of phase right shift, pulse repetition rate decrease, and increased discharge time interval. To eliminate the influence of the pulse repetition rate on the network recognition effect, when training the pattern recognition network and the characteristic parameter recognition network, according to the measured results, the number of pulses of a single sample in the power frequency sample library is normalized to the number of pulses of a single sample at low frequency, that is, the same number of sample pulses is adopted. Secondly, the phases of the power frequency sample library are normalized. Specifically, for each different type of discharge, according to the measured phase deviation , the phases of the power frequency sample library are normalized and corrected, and the phases of the original power frequency sample library data are corrected according to to obtain new phase information for training the network .

[0097] Specifically, this embodiment provides a preferred implementation manner of step S3. The cycle normalization and correction of the data in the power frequency discharge data set based on the offset data includes normalizing and correcting the amplitude data of each discharge data sample. Specifically:

[0098] Obtain the discharge amplitude of each discharge data sample V ;

[0099] Obtain the maximum amplitude V in all the discharge amplitudes V max and the minimum amplitude Vmin ;

[0100] Based on the discharge amplitude V、 Amplitude maximum value V max and amplitude minimum value V min perform calculations to obtain a correction value for normalizing and correcting the amplitude of each discharge data sample V 修正 , .

[0101] It can be understood that for amplitude normalization, it is to eliminate the differences in discharge amplitudes under different voltages and different frequencies.

[0102] Specifically, this embodiment provides a preferred implementation manner of step S3. The cycle normalization and correction of the data in the power frequency discharge data set based on the offset data includes normalizing and correcting the discharge time interval data in the power frequency discharge data. Specifically:

[0103] Obtain a preset correction coefficient k;

[0104] Obtain the discharge time intervals corresponding to each discharge type in the power frequency discharge data ;

[0105] Based on the correction coefficient k and each of the discharge time intervals , obtain a correction value for normalizing and correcting the discharge time intervals corresponding to each discharge type in the power frequency discharge data , .

[0106] More specifically, this embodiment provides a preferred implementation manner of the correction coefficient k, k = 50 / 20 = 2.5.

[0107] Specifically, this embodiment provides a preferred implementation manner of step S4. The construction of a discharge type recognition neural network and inputting each sample data in the low-frequency discharge type recognition sample set to output a recognition result corresponding to each sample data includes the steps:

[0108] Construct a PRPD pattern network, pattern network and BP neural network;

[0109] Divide the low-frequency discharge type recognition sample set into a training data set and a validation data set, and the ratio of the training data set to the validation data set corresponding to the same sample data is 4:1;

[0110] The PRPD pattern sample data in the low-frequency discharge type recognition sample set, The atlas sample data and the characteristic parameter sample data are respectively input into the corresponding networks to output the corresponding training set confusion matrix and validation set confusion matrix.

[0111] It should be specifically noted that in this embodiment, the PRPD atlas network, the atlas network are both convolutional neural networks.

[0112] To show the effect after normalization, this embodiment provides Figure 11 , Figure 11 where a, b, and c are respectively the PRPD atlas of spike discharge under power frequency, the PRPD atlas of spike discharge under low frequency, and the corrected PRPD atlas of spike discharge. It can be seen from Figure 11 that after normalization, the phase interval where the PRPD atlas cluster is located is corrected, and at the same time, the pulse repetition rate of each sample atlas is normalized to be consistent with that at 20 Hz of low frequency. The max pooling and convolution stride operations in the CNN are designed to sample from specific regions of the input and extract key features. When the phase moves, the pooling region may not be aligned with the features of the shape after phase shift, resulting in the loss of important information and possible misidentification. Therefore, the above phase normalization operation is adopted.

[0113] More specifically, in this embodiment, the normalized power frequency PRPD atlas (i.e., the PRPD atlas sample data in the low-frequency discharge type recognition sample set) is input into the CNN convolutional neural network constructed in this embodiment for training. The sample data set is divided into a training data set and a validation data set, and their proportions are 80% and 20% respectively. The results of the training set and validation set obtained by the normalized PRPD atlas network are as Figure 12 shown. It can be seen from Figure 12 that the recognition accuracy rates for the normalized PRPD atlas itself are relatively high. The accuracy rates of the training set are 98.91%, 97.36%, 96.04%, and 98.10% respectively, and the accuracy rates of the validation set are 97.55%, 97.22%, 94.85%, and 96.19% respectively. It can be seen that the training network architecture constructed in this embodiment is reasonably designed. The 20 Hz PRPD atlas (i.e., the low-frequency discharge data set) is input into the PRPD atlas network, and the results are as Figure 13 shown. It can be seen that the recognition accuracy rates of the network trained with the normalized PRPD atlas for the PRPD atlas at 20 Hz are relatively high compared to the network trained with unnormalized data, which are 98.82%, 100%, 71.15%, and 87.66 respectively.

[0114] Similar to the above normalized PRPD atlas network, the results of the training set and validation set obtained by the network trained with the normalized atlas are as Figure 14 shown. It can be seen from Figure 14 that for the normalized The recognition accuracy of the spectrogram itself is also relatively high. The accuracy rates of the training set are 98.84%, 98.27%, 90.69%, and 100% respectively, and the accuracy rates of the validation set are 99.18%, 97.50%, 86.70%, and 99.05% respectively. For the spectrogram with 20Hz input to the network trained with the above normalized data, the results are as Figure 15 shown. It can be seen that the network trained with the normalized spectrogram has a higher recognition accuracy for the spectrogram at 20Hz compared to the network trained with non-normalized data, which are 100%, 82.65%, 77.87%, and 88.16% respectively. However, for some discharge types such as spike discharges, the recognition effect is slightly inferior. Generally speaking, the recognition accuracy of the network trained with non-normalized data has been improved significantly.

[0115] Specifically, this embodiment provides a preferred implementation manner. Constructing the PRPD spectrogram network or the spectrogram network includes the steps:

[0116] Set multiple layers of neurons to form an input layer, an output layer, and an intermediate layer;

[0117] Set module A, module B, and module C in the intermediate layer, and set 1×1 convolutions for adjusting the number of channels of the input features in both module A, module B, and module C.

[0118] More specifically, the 1×1 convolution in module ABC can adjust the number of channels of the input features. By changing the number of output channels, the feature expression ability and complexity of each layer in the network can be controlled, making the network more flexible, adapting to different data features and task requirements, and realizing the dimensionality reduction operation of the feature map, which helps to reduce the computational amount and the number of parameters of the network, while maintaining the rich expression ability of the network for data. In addition, the 1×1 convolution can also introduce non-linearity, increase the expression ability of the network, and help to learn more complex features and can adjust the number of channels of the feature map. At the same time, there is an operation of decomposing the n×n convolution into 1×n convolution and n×1 convolution, decomposing the total number of parameters into 2n, significantly reducing the number of parameters and the computational complexity of the network, especially in deep networks, which helps to accelerate the training and inference processes. Each 1×1 convolution kernel alone has the effect of a non-linear activation function, such as ReLU, which can increase the non-linear expression ability of the network and enable the network to better capture complex features and patterns in the data. In addition, module B introduces a residual module, which can retain the original features while preventing problems such as network degradation and gradient disappearance due to the network being too deep during training.

[0119] Specifically, this embodiment provides a preferred implementation manner, where the mean value of the discharge time interval of the characteristic parameter sample data , the standard deviation of the discharge time interval , the entropy value of the discharge time interval , the Weibull parameter of the discharge times - amplitude (N - V), and the mean value of the amplitude , the standard deviation of the amplitude , the entropy value of the amplitude .

[0120] Based on the UHF - collected GIS partial discharge data, this embodiment extracts 26 characteristic parameters as shown in Table 1. Based on the previously proposed BP neural network, the phase, amplitude, and time interval among the 26 characteristic parameters are normalized, and then the normalized data is used to train the above - mentioned neural network. The confusion matrices of the normalized data validation set and the normalized data training network for low - frequency recognition results are as Figure 16 shown. It can be seen from Figure 16 that the recognition accuracy of the BP neural network trained with the normalized data for the normalized data is also relatively high. The accuracy rates of the validation set are 95.91%, 93.89%, 96.57%, and 98.10% respectively. When the characteristic parameters of 20 Hz are input into the network trained with the above - mentioned normalized data, the recognition accuracy is improved compared with the network trained with the non - normalized data, which are 98.65%, 92.69%, 71.15%, and 57.18% respectively. However, for some discharge types such as spike discharge and surface discharge, the recognition effect is not ideal. Therefore, it can be seen that the diagnostic effect of only using a single network or model for discharge type may not be ideal, and accordingly, a strategy for fusing the recognition results of multiple models is proposed.

[0121] Table 1 Characteristic parameters

[0122]

[0123] Specifically, this embodiment provides a preferred implementation manner of step S5. The obtaining of the preset weight vectors corresponding to each of the recognition results and the fusion of the recognition results based on each of the recognition results and its corresponding weight vector include the steps:

[0124] Obtain the preset weight vector corresponding to the confusion matrix of the PRPD pattern training set Q i PRPD ;

[0125] Obtain the preset weight vector corresponding to the pattern training set confusion matrix ;

[0126] Obtain the preset weight vector corresponding to the confusion matrix of the characteristic parameter training set Qk BP ;

[0127] Based on the Q i PRPD , and Q k BP compute the fused recognition vector Q , and select Q the index value with the highest probability in as the discharge recognition result;

[0128] The calculation formula of the recognition vector Q is:

[0129] ,

[0130] In the formula, Q i PRPD represents the weight vector of a certain input sample recognized as the i-th type of discharge by the PRPD pattern network, represents that a certain input sample passes through the weight vector of the pattern network recognized as the j-th type of discharge, Q k BP represents the weight vector of a certain input sample recognized as the k-th type of discharge by the BP neural network.

[0131] To further improve the accuracy of the above low-frequency discharge type recognition method, the recognition results of the above three discharge type recognition networks are fused to achieve accurate recognition of the partial discharge type of low-frequency GIS equipment. The fusion strategy is to form a corresponding weight matrix from the confusion matrix of each model training set. After each sample is recognized by the three networks, a specific recognition result will be given. Based on this recognition result, the corresponding weight vector is selected, and the weight vectors of the three results are added and fused to obtain a fusion diagnosis result. Suppose the confusion matrix of the training set obtained by using the above network for training the PRPD pattern is as Figure 17 shown. For this embodiment, it is the confusion matrix obtained after training the network with normalized data. Among them, the abscissa is still the predicted label, and the ordinate is the actual label. Select four elements in each column, and divide each element by the sum of the four elements as its recognition weight vector. The following formula shows the weight vector corresponding to the recognition result of 0 for this network Q 0 PRPD :

[0132]

[0133] It can be understood that the meaning of the first element in the phasor is that the accuracy of identifying this sample as the discharge of type 0 is A / (A + E + I + M), and so on for the meaning of the elements in each position. The actual meaning of this weight matrix is the probability of correct identification. Using the above fusion strategy, the confusion matrix of the fusion recognition of the 20Hz discharge sample by the trained network with the above normalized data is as follows Figure 18 shown. It can be seen that after the fusion of multiple recognition results, the problems that the accuracy of identifying spike discharges by the original spectrogram network is not high and the accuracy of identifying creeping discharges by the BP neural network is not high are improved, and the accuracy rates are both increased to about 85%.

[0134] Embodiment 2:

[0135] This embodiment provides a training system for a low-frequency discharge type recognition model based on cycle normalization, based on the method for training a low-frequency discharge type recognition model based on cycle normalization described in Embodiment 1:

[0136] It includes an acquisition module, a cycle normalization correction module, a construction module, a training module, and a result fusion module;

[0137] The acquisition module is used to acquire the power frequency discharge data set and the low-frequency discharge data set, as well as the offset data of the low-frequency discharge compared with the power frequency discharge in terms of phase, amplitude, and discharge time;

[0138] The cycle normalization correction module performs cycle normalization correction on each discharge data based on the offset data to obtain a low-frequency discharge type recognition sample set, and the low-frequency discharge type recognition sample set includes PRPD spectrogram sample data, spectrogram sample data, and characteristic parameter sample data;

[0139] The construction module is used to construct a discharge type recognition neural network;

[0140] The training module is used to input each sample data in the low-frequency discharge type recognition sample set to train the network and output the recognition results corresponding to each sample data;

[0141] The acquisition module is further used to acquire the preset weight phasor corresponding to each of the recognition results;

[0142] The result fusion module performs fusion of the recognition results based on each of the recognition results and its corresponding weight phasor to complete the training of the low-frequency discharge type recognition model.

[0143] Embodiment 3:

[0144] This embodiment provides a low-frequency discharge type recognition model based on cycle normalization, which is trained by using the training method described in Embodiment 1.​

[0145] It should be noted that, for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0146] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0147] The foregoing are only exemplary embodiments of the present disclosure, and the scope of the present disclosure cannot be limited thereby. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure still fall within the scope covered by the present disclosure. After considering the specification and practicing the present disclosure herein, those skilled in the art will readily conceive of other embodiments of the present disclosure. The present invention is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include known common knowledge or conventional technical means in the technical field not described in the present disclosure. The specification and embodiments are only regarded as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A training method for a low-frequency discharge type recognition model based on cycle normalization, characterized in that, Including the steps: S1. Obtain the power frequency discharge dataset and the low-frequency discharge dataset; S2. Based on the power frequency discharge dataset and the low-frequency discharge dataset, obtain the offset data of the low-frequency discharge compared to the power frequency discharge in terms of phase, amplitude, and discharge time; S3. Perform periodic normalization correction on the data in the power frequency discharge dataset based on the offset data to obtain a low-frequency discharge type identification sample set, where the low-frequency discharge type identification sample set includes PRPD pattern sample data, pattern sample data, and characteristic parameter sample data, and the characteristic parameter sample data includes the mean value of the discharge time interval , the standard deviation of the discharge time interval , the entropy value of the discharge time interval , the Weibull parameter of the discharge number - amplitude (N - V), the mean value of the amplitude , the standard deviation of the amplitude , the entropy value of the amplitude ; S4. Construct a PRPD pattern network and input the PRPD pattern sample data in the low-frequency discharge type recognition sample set into it, construct a pattern network and input the pattern sample data in the low-frequency discharge type recognition sample set into it, construct a BP neural network and input the feature parameter sample data in the low-frequency discharge type recognition sample set into it to train the network and output the recognition results corresponding to each sample data; S5. Obtain the preset weight phasors corresponding to each of the recognition results, and perform recognition result fusion based on each of the recognition results and their corresponding weight phasors to complete the training of the low-frequency discharge type recognition model.

2. The training method of a low-frequency discharge type recognition model based on cycle normalization according to claim 1, wherein The cycle normalization correction of the data in the power frequency discharge dataset based on the offset data in step S3 includes performing normalization correction on the phase data in the power frequency discharge data. Specifically: Obtain the phases corresponding to each discharge type in the power frequency discharge data ; Obtain the phases corresponding to each discharge type in the low-frequency discharge data ; Based on the and the corresponding to the same discharge type, perform a difference calculation to obtain the phase deviation corresponding to each discharge type , ; Based on the and the phase deviation perform calculations to obtain a correction value for normalizing and correcting the phases corresponding to each discharge type in the power frequency discharge data , .

3. A training method for a low-frequency discharge type recognition model based on cycle normalization according to claim 1, characterized in that, The cycle normalization correction of the data in the power frequency discharge dataset based on the offset data in step S3 includes performing normalization correction on the amplitude data of each discharge data sample. Specifically: Obtain the discharge amplitude of each discharge data sample V ; Obtain all the discharge amplitudes V the maximum value of the amplitudes V max and the minimum value of the amplitudes V min ; Based on the discharge amplitude V、 Amplitude maximum value V max and amplitude minimum value V min Perform calculations to obtain correction values for normalizing and correcting the amplitudes corresponding to each discharge type in each discharge data sample V 修正 , .

4. A training method for a low-frequency discharge type recognition model based on cycle normalization according to claim 1, characterized in that The cycle normalization correction of the data in the power frequency discharge dataset based on the offset data in step S3 includes performing normalization correction on the discharge time interval data in the power frequency discharge data. Specifically: Obtain the preset correction coefficient k; Obtain the discharge time intervals corresponding to each discharge type in the power frequency discharge data ; Based on the correction coefficient k and each of the discharge time intervals , a correction value for normalizing and correcting the discharge time intervals corresponding to each discharge type in the power frequency discharge data is obtained , .

5. A training method for a low-frequency discharge type recognition model based on cycle normalization according to claim 1, characterized in that, The step S4 of constructing the discharge type recognition neural network and inputting each sample data in the low-frequency discharge type recognition sample set to train the network and output the recognition results corresponding to each sample data includes the steps: Construct a PRPD map network, a map network, and a BP neural network; Divide the low-frequency discharge type recognition sample set into a training data set and a validation data set, and the ratio of the training data set to the validation data set corresponding to the same sample data is 4:1; Input the PRPD pattern sample data, pattern sample data, and characteristic parameter sample data in the low-frequency discharge type recognition sample set into the corresponding networks respectively to train the networks and output the corresponding confusion matrices of the training set and the validation set.

6. A training method for a low-frequency discharge type recognition model based on cycle normalization according to claim 5, characterized in that Constructing the PRPD map network or The map network includes the steps of: Set multiple layers of neurons to form an input layer, an output layer, and an intermediate layer; Set module A, module B, and module C in the intermediate layer, and set 1×1 convolutions for adjusting the number of channels of the input features in module A, module B, and module C.

7. A method for training a low-frequency discharge type recognition model based on cycle normalization according to claim 6, characterized in that, The step S5 of obtaining the preset weight phasors corresponding to each of the recognition results and performing recognition result fusion based on each of the recognition results and their corresponding weight phasors includes the steps: Obtain the weight vector corresponding to the confusion matrix of the preset PRPD map training set Q i PRPD ; Obtain the preset weight vector corresponding to the confusion matrix of the spectral training set ; Obtain the weight vector corresponding to the confusion matrix of the preset feature parameter training set Q k BP ; Based on the above-mentioned Q i PRPD , and Q k BP calculate the fused recognition vector Q , and select Q the index value with the highest probability in it as the discharge recognition result; The calculation formula of the recognition vector Q is: ， In the formula, Q i PRPD represents the weight vector of a certain input sample recognized as the i-th type of discharge by the PRPD pattern network. represents that the certain input sample passes through the pattern network and is recognized as the weight vector of the j-th type of discharge. Q k BP represents the weight vector of the certain input sample recognized as the k-th type of discharge by the BP neural network.

8. A training system for a low-frequency discharge type recognition model based on cycle normalization, based on the training method for a low-frequency discharge type recognition model based on cycle normalization according to any one of claims 1-7, characterized in that: It includes an acquisition module, a cycle normalization correction module, a construction module, a training module, and a result fusion module; The acquisition module is used to obtain the power frequency discharge dataset and the low-frequency discharge dataset, as well as the offset data of the low-frequency discharge compared to the power frequency discharge in terms of phase, amplitude, and discharge time; The cycle normalization correction module performs cycle normalization correction on each discharge data based on the offset data to obtain a low-frequency discharge type identification sample set, where the low-frequency discharge type identification sample set includes PRPD pattern sample data, pattern sample data, and characteristic parameter sample data; The construction module is used to construct a discharge type recognition neural network; The training module is used to input each sample data in the low-frequency discharge type recognition sample set to train the network and output the recognition results corresponding to each sample data; The acquisition module is further used to obtain the preset weight phasors corresponding to each of the recognition results; The result fusion module performs recognition result fusion based on each of the recognition results and their corresponding weight phasors to complete the training of the low-frequency discharge type recognition model.

9. A low-frequency discharge type recognition model based on cycle normalization, characterized in that: Trained by using the training method according to any one of claims 1-7.

Images