A GIS device insulation defect evaluation method, system, device and medium

Abstract

The application discloses a GIS device insulation defect evaluation method and system, a plurality of partial discharge feature vectors are obtained by extracting partial discharge information of the GIS device, a historical fault case text feature vector is obtained by extracting a historical fault case text of the GIS device, the historical fault case text feature vector and the plurality of partial discharge feature vectors are summed to obtain a plurality of vector sequences, the plurality of vector sequences are input into an LSTM model to obtain a state feature vector, the dimension of the state feature vector is reduced to the number of states by using a full connection layer, and then the state feature vector is converted into a probability distribution, the probability distribution includes a plurality of states, each state corresponds to a respective probability, and a state with the highest probability in the probability distribution is output as a defect evaluation result of the device, the defect of the GIS device is evaluated by using a neural network by considering the currently collected partial discharge detection data and the historical case data, and better evaluation effect can be obtained.

Description

Technical Field

[0001] This invention relates to the field of GIS equipment insulation defect severity assessment technology, and in particular to a method and system for evaluating GIS equipment insulation defects. Background Technology

[0002] Gas-insulated switchgear (GIS) is widely used in power grids due to its small footprint, high reliability, and excellent insulation performance. However, during manufacturing, transportation, and assembly, GIS equipment can easily introduce potential hazards such as particles and debris, leading to insulation defects like partial discharge (PD) and potentially causing major equipment failures.

[0003] Traditional GIS equipment insulation condition assessment primarily relies on relevant standards and expert experience, employing static evaluation through periodic inspections and power outage experiments. To reduce the economic losses and human resource waste caused by periodic power outages, some scholars have proposed using partial discharge statistical features combined with expert experience for GIS equipment condition assessment. With the rapid development of information technology and big data technology, intelligent algorithms such as machine learning and deep learning are widely used in equipment insulation condition assessment. Among them, the high-voltage switchgear reliability assessment model based on radial basis functions has good learning and classification capabilities, fast training speed, and low error rate; the fractal box dimension method can identify the contact state of GIS disconnect switches based on features; and the improved sparrow swarm search algorithm optimized support vector machine can effectively identify the type of partial discharge inside GIS equipment. However, besides the partial discharge characteristics inside the GIS reflecting the insulation defect status of the equipment, the equipment's historical operating conditions and the occurrence of insulation faults in similar equipment can also affect the current insulation condition assessment results. For example, if the partial discharge characteristics are similar to those of similar equipment that has experienced insulation faults, then even if the current internal partial discharge is not severe, it should still be given sufficient attention. Current intelligent diagnostic algorithms can only assess the insulation defect status of equipment by analyzing the currently collected partial discharge detection data, resulting in low levels of intelligence and accuracy in the assessment. Summary of the Invention

[0004] This invention provides a method and system for evaluating insulation defects in GIS equipment. By constructing an evaluation method and using partial discharge data and historical case records to evaluate insulation defects in GIS equipment, the invention improves the intelligence level and accuracy of insulation status evaluation for GIS equipment.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a method for evaluating insulation defects in GIS equipment, comprising:

[0006] A partial discharge model identification neural network was used to extract multiple partial discharge feature vectors from the partial discharge information of GIS equipment over several days. Historical fault case text feature vectors were also extracted from the historical fault case text of the GIS equipment.

[0007] The text feature vectors of historical fault cases are summed with multiple partial discharge feature vectors to obtain multiple vector sequences. These multiple vector sequences are then input into a preset LSTM model to calculate the state feature vector.

[0008] After reducing the dimension of the state feature vector to the number of states using a fully connected layer, the Softmax function is used to transform it into a probability distribution. The state with the highest probability from the probability distribution is output as the defect evaluation result of the GIS equipment. The probability distribution includes multiple states, each with its own probability.

[0009] In this embodiment, multiple partial discharge feature vectors are extracted from the partial discharge information of GIS equipment over several days using a partial discharge model recognition neural network. Historical fault case texts of the GIS equipment are extracted to obtain historical fault case text feature vectors. These historical fault case text feature vectors are summed with the multiple partial discharge feature vectors to obtain multiple vector sequences. These multiple vector sequences are input into a preset LSTM model to calculate state feature vectors. A fully connected layer is used to reduce the dimension of the state feature vectors to the number of states, and then a Softmax function is used to transform them into a probability distribution. The probability distribution includes multiple states, each with its own probability. The state with the highest probability from the probability distribution is output as the defect evaluation result of the GIS equipment. This method, by considering the currently collected partial discharge detection data and historical case data, utilizes a neural network to evaluate the defects of the GIS equipment, achieving better evaluation results.

[0010] As a preferred option, it also includes:

[0011] After validating the defect evaluation results using K-fold cross-validation, a confusion matrix is ​​then used to evaluate the defect evaluation results.

[0012] As a preferred approach, a partial discharge model identification neural network is used to extract multiple partial discharge feature vectors from the multi-day partial discharge information of the GIS equipment, specifically:

[0013] After zero-padding the partial discharge information, the zero-padding data is input into a convolutional layer for convolution, and the output is a partial discharge feature matrix. The convolution process is as follows:

[0014]

[0015] Where σ(·) represents the nonlinear activation function, This represents the value in the i-th row and j-th column of the output partial discharge feature matrix after convolution;

[0016] Pooling is performed on the partial discharge feature matrix to obtain the model feature vector. This feature vector is then input into a fully connected layer for classification. The dimensionality of the feature vector is reduced to a preset number of partial discharge modes, outputting the dimensionality-reduced feature vector. The fully connected layer is represented as follows:

[0017] f p =σ(W p ·f d +b p )

[0018] in, For the weights and biases of the fully connected layer, The feature vector of the classification pattern after dimensionality reduction;

[0019] After using the Softmax function to transform the dimensionality-reduced feature vectors into a pre-probability distribution, multiple partial discharge feature vectors are output.

[0020] As a preferred approach, the historical fault case text of GIS equipment is extracted to obtain the historical fault case text feature vector, specifically:

[0021] The BERT embedding model is used to model each character in the historical failure case text, and the historical failure case text is transformed into an embedding vector matrix.

[0022] After performing sequence analysis on the word embedding vectors in the embedding vector matrix of the text using an LSTM neural network, historical case features of the text are extracted.

[0023] This embodiment employs a BERT embedding model to model each character in the historical fault case text, transforming the historical fault case text into an embedding vector matrix. An LSTM neural network is then used to perform sequence analysis on the character embedding vectors in the embedding vector matrix to extract the historical case features. By using this method to transform the historical fault case text of the device into a character embedding matrix, and then utilizing a Long Short-Term Memory (LSTM) neural network to mine information from this matrix, the representation feature vectors of the historical cases can be obtained. This method can further extract sequence information, improving the accuracy of insulation defect severity assessment.

[0024] As a preferred approach, after performing sequence analysis on the word embedding vectors in the text's embedding vector matrix using an LSTM neural network, historical case features of the text are extracted, specifically:

[0025] The word embedding vector is processed using an LSTM neural network to output historical case features. The processing procedure is as follows:

[0026]

[0027]

[0028]

[0029]

[0030]

[0031] Among them, W f W i With W o b represents the weight. f b i With b o The vectors [·,·] represent the bias components, and [·,·] represent the vector concatenation operation. tanh (·) is the tanh activation function, W c b c For weights and biases, the sign is... For the term-by-term multiplication of vectors, f t Represents the forget gate, c t-1 Indicates the cell state at the previous moment, i t Indicates the input gate, c t Indicates cell state, f hc =h l Indicates the characteristics of historical cases.

[0032] As a preferred solution, to address the same technical problem, the present invention also provides a GIS equipment insulation defect evaluation system, including a data processing module, a state feature vector calculation module, and a defect evaluation result calculation module.

[0033] The data processing module is used to extract multiple partial discharge feature vectors from the partial discharge information of the GIS equipment over several days using a partial discharge model recognition neural network, and to extract historical fault case text feature vectors from the historical fault case text of the GIS equipment.

[0034] The state feature vector calculation module is used to sum the text feature vectors of historical fault cases with multiple partial discharge feature vectors to obtain multiple vector sequences. The multiple vector sequences are then input into a preset LSTM model to calculate the state feature vector.

[0035] The defect evaluation result calculation module uses a fully connected layer to reduce the dimension of the state feature vector to the number of states, and then uses the Softmax function to transform it into a probability distribution. The state with the highest probability from the probability distribution is output as the defect evaluation result of the GIS equipment. The probability distribution includes multiple states, each with its own probability.

[0036] As a preferred option, a verification module is also included.

[0037] The verification module is used to verify the defect evaluation results using K-fold cross-validation, and then to evaluate the defect evaluation results using a confusion matrix.

[0038] As a preferred solution, the data processing module includes convolutional units, pooling units, and dimensionality reduction units.

[0039] The convolutional unit is used to pad the partial discharge information with zeros. The zero-padded data is then input into the convolutional layer for convolution, outputting the partial discharge feature matrix. The convolution process is as follows:

[0040]

[0041] Where σ(·) represents the nonlinear activation function, This represents the value in the i-th row and j-th column of the output partial discharge feature matrix after convolution;

[0042] The pooling unit performs pooling operations on the partial discharge feature matrix to obtain the model feature vector. This model feature vector is then input into the fully connected layer for classification. The dimensionality of the feature vector is reduced to a preset number of partial discharge modes, outputting the dimensionality-reduced feature vector. The fully connected layer is represented as follows:

[0043] f p =σ(W p ·f d +b p )

[0044] in, For the weights and biases of the fully connected layer, The feature vector of the classification pattern after dimensionality reduction;

[0045] The dimensionality reduction unit is used to transform the dimensionality-reduced feature vectors into a pre-probability distribution using the Softmax function, and then outputs multiple partial discharge feature vectors.

[0046] As a preferred embodiment, the present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the GIS equipment insulation defect evaluation method as shown in this embodiment.

[0047] As a preferred embodiment, the present invention also provides a computer program that, when executed by a processor, implements the steps of the GIS equipment insulation defect evaluation method as described in this embodiment.

[0048] The present invention has the following beneficial effects:

[0049] This method employs a partial discharge model recognition neural network to extract multiple partial discharge feature vectors from multi-day partial discharge information of GIS equipment. It also extracts historical fault case text feature vectors from the historical fault case text, summing these vectors with the partial discharge feature vectors to obtain multiple vector sequences. These vector sequences are then input into a pre-defined LSTM model to calculate state feature vectors. A fully connected layer reduces the dimension of the state feature vectors to the number of states, and a Softmax function is used to transform them into a probability distribution. This probability distribution includes multiple states, each with its own probability. The state with the highest probability from the probability distribution is output as the defect evaluation result for the GIS equipment. This method, by considering both currently collected partial discharge detection data and historical case data, utilizes a neural network to evaluate the defects of GIS equipment, achieving better evaluation results. Attached Figure Description

[0050] Figure 1 : A schematic flowchart of an embodiment of the GIS equipment insulation defect evaluation method provided by the present invention;

[0051] Figure 2 : A schematic diagram of the method flow of an embodiment of the GIS equipment insulation defect evaluation method provided by the present invention;

[0052] Figure 3 : A confusion matrix diagram of an embodiment of the GIS equipment insulation defect evaluation method provided by the present invention;

[0053] Figure 4 : A schematic diagram of the system structure of another embodiment of the GIS equipment insulation defect evaluation method provided by the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1

[0056] Please refer to Figure 1The present invention provides a method for evaluating insulation defects in GIS equipment, which includes steps 101 to 103, the specific details of which are as follows:

[0057] Step 101: Use a partial discharge model recognition neural network to extract multiple partial discharge feature vectors from the partial discharge information of the GIS equipment over several days, and extract the historical fault case text feature vectors from the historical fault case text of the GIS equipment.

[0058] Optionally, a partial discharge model recognition neural network is used to extract multiple partial discharge feature vectors from the multi-day partial discharge information of the GIS equipment, specifically:

[0059] After zero-padding the partial discharge information, the zero-padding data is input into a convolutional layer for convolution, and the output is a partial discharge feature matrix. The convolution process is as follows:

[0060]

[0061] Where σ(·) represents the nonlinear activation function, This represents the value in the i-th row and j-th column of the output partial discharge feature matrix after convolution;

[0062] Pooling is performed on the partial discharge feature matrix to obtain the model feature vector. This feature vector is then input into a fully connected layer for classification. The dimensionality of the feature vector is reduced to a preset number of partial discharge modes, outputting the dimensionality-reduced feature vector. The fully connected layer is represented as follows:

[0063] f p =σ(W p ·f d +b p )

[0064] in, For the weights and biases of the fully connected layer, The feature vector of the classification pattern after dimensionality reduction;

[0065] After using the Softmax function to transform the dimensionality-reduced feature vectors into a pre-probability distribution, multiple partial discharge feature vectors are output.

[0066] Optionally, the historical fault case text of GIS equipment can be extracted to obtain the historical fault case text feature vector, specifically:

[0067] The BERT embedding model is used to model each character in the historical failure case text, and the historical failure case text is transformed into an embedding vector matrix.

[0068] After performing sequence analysis on the word embedding vectors in the embedding vector matrix of the text using an LSTM neural network, historical case features of the text are extracted.

[0069] Optionally, after performing sequence analysis on the word embedding vectors in the text's embedding vector matrix using an LSTM neural network, historical case features of the text can be extracted, specifically:

[0070] The historical case features are output by processing the word embedding vectors in the LSTM neural network. The processing procedure is as follows:

[0071]

[0072]

[0073]

[0074]

[0075]

[0076] Among them, W f W i With W o b represents the weight. f b i With b o The vectors [·,·] represent the bias components, and [·,·] represent the vector concatenation operation. tanh (·) is the tanh activation function, W c b c For weights and biases, the sign is... For the term-by-term multiplication of vectors, f t Represents the forget gate, c t-1 Indicates the cell state at the previous moment, i t Indicates the input gate, c t Indicates cell state, f hc =h l Indicates the characteristics of historical cases, i t Indicates the input gate, o t This indicates the output gate.

[0077] In this embodiment, as Figure 2As shown, firstly, the partial discharge mode features of the equipment are extracted. Since different partial discharge modes have varying impacts on the insulation state of the equipment, the partial discharge feature information significantly affects the accuracy of the state assessment. Furthermore, the PRPS data of power equipment is essentially a two-dimensional matrix, exhibiting some similarity to image formats. Therefore, this invention uses a Convolutional Neural Network (CNN) to identify the partial discharge modes within the GIS equipment and extract the partial discharge feature vectors from the equipment over the past five days. The partial discharge mode recognition CNN used in this invention consists of two convolutional layers, two pooling layers, one fully connected layer, and one normalized exponential function layer (Softmax). The specific steps are as follows:

[0078] 1) Pad the PRPS data with zeros to ensure that the output data of subsequent convolutional layers has a consistent format;

[0079] 2) The zero-padded PRPS data is input into a convolutional layer for convolution. The convolutional layer is the core of the CNN neural network; it has multiple convolutional kernels and can extract local features of the partial discharge signal. One convolutional kernel of the convolutional layer is Ω. s×s The input normalized partial discharge PRPS data is The convolution operation of a convolutional layer can be represented by the following formula:

[0080]

[0081] Where σ(·) represents the nonlinear activation function, This represents the value in the i-th row and j-th column of the output partial discharge feature matrix after convolution;

[0082] 3) Pooling is performed on the partial-amplitude feature matrix to reduce the data size while preserving the original data features and reducing model overfitting. After two convolutions and pooling operations, the final pattern feature vector of the partial-amplitude data is obtained.

[0083] 4) Input the pattern feature vector into a fully connected layer for classification, reducing the dimension of the feature vector to the number of partial discharge patterns, d². The fully connected layer can be represented by the following formula:

[0084] f p =σ(W p ·f d +b p )

[0085] in, For the weights and biases of the fully connected layer, The feature vector of the classification pattern after dimensionality reduction;

[0086] 5) Use the Softmax function to optimize the output f of the fully connected layer. p This is transformed into a probability distribution, thereby enabling pattern recognition of partial discharge data.

[0087] Secondly, the textual features of historical failure cases are extracted. The historical failure case texts are converted into word embedding matrices, and a Long Short-Term Memory (LSTM) neural network is used to mine information from these matrices to obtain the representation feature vectors of the historical cases. The specific steps are as follows:

[0088] The BERT embedding model is used to model each character in the text, representing each character with a word embedding vector. This transforms the historical fault case text into an embedding vector matrix. The length and direction of the word embedding vector in the embedding space can reflect the semantics of a character to some extent. Characters with similar semantics have two embedding vectors with similar lengths and directions. On the other hand, characters with opposite semantics have larger distances between their embedding vectors. As an example of this embodiment, the historical fault case text of GIS is T. hc = <w1,w2,…,w l >, where w i Let be the i-th character in the text, and l be the length of the text. Then the BERT model can be represented as:

[0089] Φ=f bert (T hc )

[0090] Among them, f bert (·) represents the BERT embedding model. The column vectors in the embedding representation matrix representing historical GIS cases. This represents the character embedding vector corresponding to the i-th character in the text, with dimension d. e .

[0091] Using an LSTM neural network to extract historical case features, the order of characters in the historical failure case text also affects the semantic information of the entire text. Different arrangements of the same characters often express different meanings. Therefore, it is necessary to perform sequence analysis on the character embedding vectors in the text's embedding representation matrix and extract the historical case feature vectors of the text. Let the input character embedding vector at time t be... The output is h t The cell state is c t Then the forget gate f of LSTM at time t t Input gate i t Output gate o t It can be expressed by the following formula:

[0092]

[0093]

[0094]

[0095]

[0096]

[0097] Among them, W f W i With W o b represents the weight. f b i With b o The vectors [·,·] represent the bias components, and [·,·] represent the vector concatenation operation. tanh (·) is the tanh activation function, W c b c For weights and biases, the sign is... For the term-by-term multiplication of vectors, f t Represents the forget gate, c t-1 Indicates the cell state at the previous moment, i t Indicates the input gate, c t Indicates cell state, f hc =h l Indicates the characteristics of historical cases, i t Indicates the input gate, o t This indicates the output gate.

[0098] Therefore, the forget gate filters the cell state from the previous moment, selecting which information needs to be forgotten, while the input gate filters and updates the information to the current memory state, i.e., the cell state, based on the current input information. Ultimately, the feature vector of the entire historical fault case text is the output f of the last corresponding word embedding vector sequence. hc =h l This vector contains information about the entire text and will be used in conjunction with the local feature vector to evaluate the state of the GIS.

[0099] Step 102: Summing the text feature vectors of historical fault cases with the multiple partial discharge feature vectors to obtain multiple vector sequences, and inputting the multiple vector sequences into a preset LSTM model to calculate the state feature vector.

[0100] In this embodiment, the case representation feature vector is summed with the obtained five partial discharge feature vectors. The development trend of partial discharge in GIS equipment also affects the insulation defect status of the equipment to some extent. When the partial discharge becomes more severe, the insulation status of the equipment becomes more critical. Therefore, the status of the equipment needs to be evaluated based on the partial discharge signal sequence within 5 days and integrated into the equipment's historical case information.

[0101] The summed vector sequence is then input into the LSTM. After the following steps, the final state feature vector of the device is obtained. The implementation steps are as follows:

[0102]

[0103]

[0104]

[0105]

[0106]

[0107] Among them, W f W i With W o b represents the weight. f b i With b o The vectors [·,·] represent the bias components, and [·,·] represent the vector concatenation operation. tanh (·) is the tanh activation function, W c b c For weights and biases, the sign is... For the term-by-term multiplication of vectors, f t Represents the forget gate, c t-1 Indicates the cell state at the previous moment, i t Indicates the input gate, c t Indicates cell state, f hc =h l Indicates the characteristics of historical cases, i t Indicates the input gate, o t This indicates the output gate.

[0108] As an example of this embodiment, the state of the device is evaluated based on the partial discharge signal sequence within 5 days and incorporated into the device's historical case information. Therefore, the case feature vector is summed with the 5 partial discharge feature vectors to finally obtain 5 state feature vectors.

[0109] Step 103: After reducing the dimension of the state feature vector to the number of states using a fully connected layer, use the Softmax function to transform it into a probability distribution. Output the state with the highest probability from the probability distribution as the defect evaluation result of the GIS equipment. The probability distribution includes multiple states, each with its own probability.

[0110] In this embodiment, a fully connected layer is used to reduce the dimension of the feature vector to the number of states, and the Softmax function is used to transform it into a probability distribution. The probability distribution contains four elements, each representing the probability that the GIS device belongs to one of the four states. The sum of the four elements is 1. The final severity of the GIS is determined by which state in the probability distribution has the highest probability. The highest probability state is the severity of the insulation defect of the device.

[0111] Optionally, the method may also include using K-fold cross-validation to verify the defect evaluation results, and then using a confusion matrix to evaluate the defect evaluation results.

[0112] In this embodiment, the proposed method for assessing the severity of insulation defects in GIS equipment is validated using K-fold cross-validation. The parameter scanning method is used to set s=3, d1=256, d2=5, and d... e =768. On the other hand, the LSTM layer used for historical case feature extraction has 1 layer and a hidden size of 256, while the LSTM layer used for state feature extraction has 1 layer and a hidden size of 128. The activation function σ(·) is a decaying trimmed linear unit with a decay rate of -0.2. The model's loss function is the cross-entropy function. To further verify the method's preference for different insulation state data, a confusion matrix is ​​used for evaluation. The confusion matrix of the validation results on 1200 validation set data is shown below. Figure 3 As shown in the experimental results of the confusion matrix, this method can accurately and effectively assess the severity of different equipment insulation defects.

[0113] This method accurately identifies samples with varying degrees of defect severity on the validation set. Specifically, the model's precision for the four insulation states is 91.50%, 86.45%, 86.64%, and 88.01%, respectively, while the recall rates are 93.33%, 89.33%, 84.33%, and 85.67%. This method effectively extracts continuous features from the samples and treats the severity levels of insulation defects in GIS equipment as continuous rather than discrete, aligning with common assessment practices. Compared to traditional models, the F-measure is improved by 19.39%, and with the support of historical fault case texts, it can treat samples of various insulation defect states equally.

[0114] This invention offers the following advantages: It extracts multiple partial discharge feature vectors from multi-day partial discharge information of GIS equipment using a partial discharge model recognition neural network; extracts historical fault case text feature vectors from the historical fault case text of the GIS equipment; sums the historical fault case text feature vectors with the multiple partial discharge feature vectors to obtain multiple vector sequences; inputs these multiple vector sequences into a preset LSTM model to calculate state feature vectors; reduces the dimension of the state feature vectors to the number of states using a fully connected layer; and then uses a Softmax function to transform them into a probability distribution. This probability distribution includes multiple states, each with its own probability. The state with the highest probability from the probability distribution is output as the defect evaluation result of the GIS equipment. This method, by considering both currently collected partial discharge detection data and historical case data, utilizes a neural network to evaluate the defects of GIS equipment, achieving better evaluation results.

[0115] Example 2

[0116] Accordingly, see Figure 4 , Figure 4 This is a schematic diagram of the structure of a GIS equipment insulation defect evaluation system provided by the present invention. As shown in the figure, the GIS equipment insulation defect evaluation system includes a data processing module 401, a state feature vector calculation module 402, and a defect evaluation result calculation module 403.

[0117] Among them, the data processing module 401 is used to extract multiple partial discharge feature vectors from the partial discharge information of the GIS equipment over multiple days using a partial discharge model recognition neural network, and to extract the historical fault case text feature vectors from the historical fault case text of the GIS equipment.

[0118] The state feature vector calculation module 402 is used to sum the text feature vector of historical fault cases with multiple partial discharge feature vectors to obtain multiple vector sequences, and input the multiple vector sequences into a preset LSTM model to calculate the state feature vector.

[0119] The defect evaluation result calculation module 403 is used to reduce the dimension of the state feature vector to the number of states using a fully connected layer, and then use the Softmax function to transform it into a probability distribution. The state with the highest probability in the probability distribution is output as the defect evaluation result of the GIS equipment. The probability distribution includes multiple states, each with its own probability.

[0120] Optionally, a 404 verification module is also included.

[0121] The verification module 404 is used to verify the defect evaluation results using K-fold cross-validation, and then to evaluate the defect evaluation results using a confusion matrix.

[0122] Optionally, it also includes a data processing module 401, comprising a convolution unit 4011, a pooling unit 4012, and a dimensionality reduction unit 4013.

[0123] Convolutional unit 4011 is used to pad the partial discharge information with zeros, then input the zero-padded data into the convolutional layer for convolution operation, and output the partial discharge feature matrix. The convolution process is as follows:

[0124]

[0125] Where σ(·) represents the nonlinear activation function, This represents the value in the i-th row and j-th column of the output partial discharge feature matrix after convolution;

[0126] Pooling unit 4012 is used to perform pooling operations on the partial discharge feature matrix to obtain model feature vectors. These feature vectors are then input into a fully connected layer for classification. The dimensionality of the feature vectors is reduced to a preset number of partial discharge modes, outputting the dimensionality-reduced feature vectors. The fully connected layer is expressed as follows:

[0127] f p =σ(W p ·f d +b p )

[0128] in, For the weights and biases of the fully connected layer, The feature vector of the classification pattern after dimensionality reduction;

[0129] The dimensionality reduction unit 4013 is used to transform the dimensionality-reduced feature vectors into a pre-probability distribution using the Softmax function, and then output multiple partial discharge feature vectors.

[0130] Optionally, the present invention also provides an electronic device, a memory, and a processor for storing a computer program and implementing the GIS equipment insulation defect evaluation method as described in Embodiment 1 when executing the computer program.

[0131] Optionally, the present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the GIS equipment insulation defect evaluation method as described in Embodiment 1.

[0132] For a more detailed explanation of the working principle and procedures of this embodiment, please refer to the relevant description in Embodiment 1.

[0133] Compared to existing technologies, this method employs a partial discharge model recognition neural network to extract multiple partial discharge feature vectors from multi-day partial discharge information of GIS equipment. It also extracts historical fault case text feature vectors from the historical fault case text, sums these feature vectors with the multiple partial discharge feature vectors to obtain multiple vector sequences. These vector sequences are then input into a pre-defined LSTM model to calculate state feature vectors. A fully connected layer reduces the dimension of the state feature vectors to the number of states, and a Softmax function is used to transform them into a probability distribution. This probability distribution includes multiple states, each with its own probability. The state with the highest probability from the probability distribution is output as the defect evaluation result for the GIS equipment. This method, by considering both currently collected partial discharge detection data and historical case data, utilizes a neural network to evaluate the defects of GIS equipment, achieving better evaluation results.

[0134] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for evaluating insulation defects in GIS equipment, characterized in that, include: A partial discharge model identification neural network was used to extract multiple partial discharge feature vectors from the partial discharge information of GIS equipment over several days. Historical fault case text feature vectors were also extracted from the historical fault case text of the GIS equipment. The text feature vectors of the historical fault cases are summed with the multiple partial discharge feature vectors to obtain multiple vector sequences. The multiple vector sequences are then input into a preset LSTM model to calculate the state feature vector. After reducing the dimension of the state feature vector to the number of states using a fully connected layer, the Softmax function is used to transform it into a probability distribution. The state with the highest probability from the probability distribution is output as the defect evaluation result of the GIS device. The probability distribution includes multiple states, each with its own probability.

2. The method for evaluating insulation defects in GIS equipment as described in claim 1, characterized in that, Also includes: After validating the defect evaluation results using K-fold cross-validation, a confusion matrix is ​​then used to evaluate the defect evaluation results.

3. The method for evaluating insulation defects in GIS equipment as described in claim 1, characterized in that, The partial discharge model identification neural network is used to extract multiple partial discharge feature vectors from the multi-day partial discharge information of GIS equipment, specifically: After zero-padding the partial discharge information, the zero-padding data is input into a convolutional layer for convolution, and the output is a partial discharge feature matrix. The convolution process is as follows: Where σ(·) represents the nonlinear activation function, P represents the value in the i-th row and j-th column of the output partial discharge feature matrix after convolution. d (i+m, j+n) represents the normalized partial discharge PRPS data, where i+m and j+n represent the values ​​at the i+m and j+n positions in the zero-padded data, respectively, and Ω(m, n) represents the convolution kernel, where m and n represent the index variables. Pooling is performed on the partial discharge feature matrix to obtain the model feature vector. The model feature vector is then input into a fully connected layer for classification. The dimensionality of the feature vector is reduced to a preset number of partial discharge modes, and the dimensionality-reduced feature vector is output. The fully connected layer is expressed as follows: f p =σ(W p ·f d +b p ) in, For the weights and biases of the fully connected layer, f is the feature vector of the classification pattern after dimensionality reduction. d Represents the model feature vector; After converting the dimensionality-reduced feature vectors into a pre-probability distribution using the Softmax function, multiple partial discharge feature vectors are output.

4. The method for evaluating insulation defects in GIS equipment as described in claim 1, characterized in that, The extraction of historical fault case texts from GIS equipment to obtain historical fault case text feature vectors is specifically as follows: The BERT embedding model is used to model each character in the historical failure case text, and the historical failure case text is transformed into an embedding vector matrix. After performing sequence analysis on the word embedding vectors in the embedding vector matrix of the text using an LSTM neural network, historical case features of the text are extracted.

5. The method for evaluating insulation defects in GIS equipment as described in claim 4, characterized in that, After performing sequence analysis on the word embedding vectors in the text's embedding vector matrix using an LSTM neural network, the historical case features of the text are extracted, specifically as follows: The word embedding vector is processed using an LSTM neural network to output historical case features. The processing procedure is as follows: Among them, W f W i With W o b represents the weight. f b i With b o The vectors [·,·] represent the bias components, and [·,·] represent the vector concatenation operation. tanh (·) is the tanh activation function, W c b c For weights and biases, the sign is... For the term-by-term multiplication of vectors, f t Represents the forget gate, c t-1 Indicates the cell state at the previous moment, i t Indicates the input gate, c t Indicates cell state, f hc =h l Indicates characteristics of historical cases, o t Indicates the output gate, h t The current hidden state. This is the word embedding vector.

6. A GIS equipment insulation defect evaluation system, characterized in that, It includes a data processing module, a state feature vector calculation module, and a defect evaluation result calculation module. The data processing module is used to extract multiple partial discharge feature vectors from the partial discharge information of the GIS equipment over several days using a partial discharge model recognition neural network, and to extract historical fault case text feature vectors from the historical fault case text of the GIS equipment. The state feature vector calculation module is used to sum the text feature vector of the historical fault case with the multiple partial discharge feature vectors to obtain multiple vector sequences, and input the multiple vector sequences into a preset LSTM model to calculate the state feature vector; The defect evaluation result calculation module is used to reduce the dimension of the state feature vector to the number of states using a fully connected layer, and then use the Softmax function to transform it into a probability distribution. The state with the highest probability from the probability distribution is output as the defect evaluation result of the GIS device. The probability distribution includes multiple states, each with its own probability.

7. The GIS equipment insulation defect evaluation system as described in claim 6, characterized in that, It also includes a verification module. The verification module is used to verify the defect evaluation results using K-fold cross-validation, and then to evaluate the defect evaluation results using a confusion matrix.

8. The GIS equipment insulation defect evaluation system as described in claim 6, characterized in that, The data processing module includes convolutional units, pooling units, and dimensionality reduction units. The convolutional unit is used to pad the partial discharge information with zeros, input the zero-padded data into the convolutional layer for convolution, and output the partial discharge feature matrix. The convolution process is as follows: Where σ(·) represents the nonlinear activation function, P represents the value in the i-th row and j-th column of the output partial discharge feature matrix after convolution. d (i+m, j+n) represents the zero-padded data, i+m and j+n represent the values ​​at the i+m and j+n positions in the zero-padded data, respectively, Ω(·,·) represents the convolution kernel, and m and n represent the index variables. The pooling unit is used to perform a pooling operation on the partial discharge feature matrix to obtain a model feature vector. This model feature vector is then input into a fully connected layer for classification. The dimensionality of the feature vector is reduced to a preset number of partial discharge modes, and the reduced-dimensional feature vector is output. The fully connected layer is expressed as follows: f p =σ(W p ·f d +b p ) in, For the weights and biases of the fully connected layer, b is the feature vector of the classification pattern after dimensionality reduction. p For the bias term, f d Represents the model feature vector; The dimensionality reduction unit is used to transform the dimensionality-reduced feature vector into a pre-probability distribution using the Softmax function, and then output multiple partial discharge feature vectors.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to implement the GIS equipment insulation defect evaluation method as described in any one of claims 1 to 5 when executing the computer program.

10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the GIS equipment insulation defect evaluation method as described in any one of claims 1 to 5.

Images