Oil-immersed transformer anomaly detection method based on capsule network
Through the capsule network-based method, the problems of oil-immersed transformer discharge and overheating fault detection are solved, and accurate judgment and efficient detection of faults are achieved, reducing the repetition of manpower and material testing.
Patent Information
- Application Number
- CN202411791084.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-06
- Publication Date
- 2025-04-15
AI Technical Summary
The prior art is difficult to effectively detect and predict discharge and overheating failures of oil-immersed transformers, resulting in equipment failures and grid abnormalities.
The abnormality detection method based on the capsule network is adopted, through vector input and vector output, combined with dynamic routing mechanism, the characteristic losses caused by traditional pooling operations are reduced, and whether the oil-immersed transformer has discharge and overheating faults occur in real time.
It realizes accurate judgment of discharge and overheating faults of oil-immersed transformers, reduces the repetition of manpower and material testing, and improves the accuracy and efficiency of detection.
Smart Images
Figure CN119830167A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of an abnormality detection method for an oil-immersed transformer, and in particular to an abnormality detection method for an oil-immersed transformer based on a capsule network. Background Art
[0002] Oil-immersed transformer is one of the important equipment in power distribution station. It uses oil as the main insulation means of transformer and relies on oil as cooling medium, such as oil-immersed self-cooling, oil-immersed air cooling, etc. Its safe and stable operation is an important guarantee to ensure the safety and reliability of power grid and electrical equipment. However, with the operation of oil-immersed transformer, irregular operation and untimely maintenance will lead to a series of possible problems, such as overheating defects, discharge defects, insulation moisture defects and winding deformation defects. These abnormal defects may cause the oil-immersed transformer equipment to fail to work, or even affect the surrounding equipment and cause large-scale abnormalities in the power supply and distribution of the entire area, bringing inestimable losses.
[0003] Among them, oil leakage detection for oil-immersed transformers belongs to the field of transformer oil leakage detection. There are two reasons for oil leakage. On the one hand, it is a problem left over from the transformer design and manufacturing process. On the other hand, it is leakage caused by improper installation and maintenance of the transformer. The detection method for this problem is to set up a multi-point test disk, pass air into the heat sink, form abnormal oil leakage replica points, and find the corresponding abnormal oil leakage points based on this.
[0004] The detection of the inflation status of oil-immersed transformers is mainly based on the pressure, temperature, humidity and other data collected by the signal acquisition unit. After the data is converted, it is combined with environmental information for real-time monitoring. If an abnormality occurs, a corresponding alarm is issued and transmitted to the remote terminal.
[0005] Partial discharge detection for oil-immersed transformers refers to the detection of partial discharge caused by local and repeated breakdown at the defects of the transformer insulation due to bubbles, impurities, etc. when an alternating voltage is applied. Existing methods include combining ultrasonic and current detection to collect and detect abnormalities in the operation of the transformer.
[0006] The abnormal oil level detection of the oil-immersed transformer refers to the problem that the oil level reading of the transformer changes with the oil temperature, and this change makes it difficult to judge the abnormal oil level. The existing method is to obtain the actual oil level value and the actual oil temperature value of the transformer, and combine the set oil level value fluctuation range and the oil temperature value error value to obtain the upper and lower limits of the oil level. If the actual oil level value is not within the interval formed by the upper limit, the oil level is judged to be abnormal.
[0007] Abnormal fan air volume detection for oil-immersed transformers refers to the detection of insufficient fan air volume due to old fans, fan dust accumulation, fan damage, etc., which in turn affects the oil temperature and transformer abnormalities. There are currently few methods in this area. The existing methods include analyzing the characteristic information contained in the dynamic changes of the top oil temperature of the transformer to determine whether the fan operation status is abnormal.
[0008] Abnormal detection of the oil bladder of an oil-immersed transformer refers to abnormal detection of the oil bladder installed in the transformer oil storage cabinet to prevent the oil bladder from failing to work, resulting in the inability to alleviate the change in oil volume, and then causing the internal pressure to rise, leading to irreversible consequences. Existing methods include improving the design of the oil bladder device and inspecting it when the oil bladder is fully inflated and expanded.
[0009] In order to ensure the safe operation, long life and high efficiency of oil-immersed transformers, a series of abnormal detection must be performed. Summary of the invention
[0010] In order to solve the above technical problems, the present invention provides a capsule network that innovatively proposes a "vector input, vector output" method compared to the classic machine learning method, and adopts a dynamic routing mechanism to reduce the feature loss caused by traditional pooling operations, and accurately judge whether discharge and overheating faults occur in real time, thereby reducing repeated manpower and material detection. A capsule network-based oil-immersed transformer abnormality detection method.
[0011] The capsule network-based oil-immersed transformer abnormality detection method of the present invention comprises the following steps:
[0012] S1. Obtain data sample set. The data set uses 686 groups of measured data from the power supply company and DGA data downloaded from some literature, including 52 groups of normal data and 634 groups of fault data. The data set attributes include 6 attributes: H2, CH4, C2H6, C2H4, C2H2, and fault type. The specific content is shown in Table 1. Due to the complex causes of oil-immersed transformer failures, there are many types of faults. However, whether it is caused by external or internal reasons, the fault will eventually manifest in the form of heat and electricity. Therefore, the fault type is divided into 4 states according to the possible state of the transformer in actual situation: normal, overheating fault, discharge fault and discharge and overheating fault. The specific types are shown in Table 2:
[0013] Table 1 Dataset attributes
[0014]
[0015] Table 2 Fault type name
[0016]
[0017] S2, 3-sigma detects and removes abnormal data. Due to sensor failure, unreliable transmission, system maintenance or insufficient data storage, DGA data often contains abnormal and inaccurate values. Low-quality data easily complicates feature analysis and reduces model training effect. By referring to the "3-sigma" method to detect abnormal data in the data set, the implementation of "3-sigma" is based on the empirical rule that 99.7% of the values in the distribution fall within three standard deviations of the mean; the formula of the "3-sigma" method is shown in formula (1):
[0018]
[0019] Where x is a data point, representing a specific data value (such as H2 concentration), x0 is a set of all data points, avg(x0) is the mean of the data point set, and std(x0) represents the standard deviation of the data point set;
[0020] S3. Data processing. As the data vary greatly, as shown in Table 3, some data differ by 3 orders of magnitude. Therefore, for the convenience of training, the data needs to be processed. For the H2 attribute, the ratio of H2 to the total gas is taken. For CH4 and the remaining hydrocarbon gases, the ratio of the gas to the total hydrocarbon is taken, that is:
[0021] H2=H2 / (H2+CH4+C2H6+C2H4+C2H2)
[0022] CH4=CH4 / (CH4+C2H6+C2H4+C2H2)
[0023] C2H6=C2H6 / (CH4+C2H6+C2H4+C2H2)
[0024] C2H4=C2H4 / (CH4+C2H6+C2H4+C2H2)
[0025] C2H2=C2H2 / (CH4+C2H6+C2H4+C2H2)
[0026] Table 3 Partial data of the dataset
[0027]
[0028] Some of the data after data preprocessing are shown in Table 4:
[0029] Table 4 Part of the data after data preprocessing
[0030]
[0031] Since the data values are all between [0, 1), the data normalization is also completed. Then the data samples are divided into a training validation set and a test set in a ratio of 4:1, that is, the training validation set contains 548 data and the test set contains 138 data;
[0032] S4. Input the training data set into the capsule network for training. In the capsule network, unlike the convolutional neural network, the capsule network uses vectors instead of neurons as input and output, and uses the length of the activity vector to represent the probability of entity existence, and uses its direction to represent the instantiation parameter, such as Figure 2 Shown is a simple model of capsule network;
[0033] Specifically, the input vector V1 is multiplied by a weight matrix W1 to obtain a "prediction vector" uj|i, as shown in formula (2):
[0034]
[0035] Among them, ui represents the i-th vector of the upper layer (input layer), Wij represents the weight matrix corresponding to the i-th vector of the upper layer, cij is the coupling coefficient determined by the iterative dynamic routing process, and bij is generally initialized to 0, as shown in formula (3):
[0036]
[0037] The total input to capsule Sj is the weighted sum of all “prediction vectors” uj|i from the next layer, as shown in formula (4):
[0038]
[0039] Different from the activation function of the neural network, the Squashing function is used in the capsule network to compress the capsule Sj, as shown in formula (5):
[0040]
[0041] Finally, update the value of the temporary variable bij, as shown in formula (6):
[0042]
[0043] In the capsule network, the training loss is composed of the margin loss and the reconstruction loss. The margin loss is shown in formula (7):
[0044] L k =T k max(0,m + -‖V k ‖)2 +λ(1-T k )max(0,‖V k ‖-m - ) 2 (7)
[0045] Among them, Tk = 1 if and only if the k category exists, otherwise Tk = 0, m+ = 0.9, m- = 0.1, λ is the proportional coefficient and takes 0.5. Based on the above capsule network model, an oil-immersed transformer anomaly detection model based on capsule network is established, and the training and validation sets are used for model training;
[0046] S5. Use the test data set to test the accuracy of the model. Input the divided data set into the trained model to test the precision, recall and F1 value of the model on the training and validation sets, as well as the accuracy on the test set. The calculation formula is as follows:
[0047]
[0048] Among them, TP represents the number of true positive results, FP represents the number of false positive results, TN represents the number of true negative results, and FN represents the number of false negative results. Precision (P) represents the proportion of true positive examples in the predicted results to all positive examples in the predicted results. The higher the precision rate, the higher the proportion of true positive examples in the model prediction results, and the stronger the model's ability to recognize positive samples. Recall (R) represents the proportion of true positive examples in the predicted results to all actual positive examples. The higher the recall rate, the higher the proportion of positive samples that the model can successfully predict, and the more comprehensive the model's recognition ability. The F1 value is the harmonic mean of the precision rate and the recall rate. The higher the F1 value, the better the model performs in terms of accuracy and reliability.
[0049] Preferably, the training set, the test set and the validation set are divided into different proportions for testing, so that in the training and validation sets, the accuracy of the training and validation sets, the recall rate of the training and validation sets and the F1 value of the training and validation sets are finally obtained. Figure 3 , Figure 4 and Figure 5 shown.
[0050] Preferably, the capsule network is used to perform the test on the test set, and the test results are shown in Table 5:
[0051] Table 5 Test results of the test set
[0052]
[0053] According to the test results obtained in Table 5, the capsule network model achieves a comprehensive accuracy of 90% in the oil-immersed transformer test data set. Among them, due to the small proportion of normal labels and discharge and overheating fault labels in the overall data set, the training and recognition effects are slightly poor. However, for discharge faults and overheating faults, the recognition accuracy can reach more than 92%, effectively identifying the anomalies of the oil-immersed transformer based on the gas data.
[0054] Preferably, a novel capsule network is used to train the existing oil-immersed transformer abnormal data set to obtain a model that can accurately detect three types of abnormalities; compared with the classic machine learning method, the capsule network innovatively proposes a "vector input, vector output" method, and adopts a dynamic routing mechanism to reduce the feature loss caused by traditional pooling operations, so that the lower-level features are only passed to the matching high-level, more features of the original data are retained, and the accuracy of feature recognition is improved. Compared with the method that requires the addition of additional equipment or additional operations such as using ultrasonic and current detection and passing air into the heat sink to find abnormal points, the method based on the capsule network can effectively save manpower and material costs. There is no need to use manual on-site operation judgment. It only needs to use the gas data in the oil-immersed transformer collected by the sensor and input it into the capsule network model. It can make accurate judgments on whether discharge and overheating faults occur in real time, reducing repeated manpower and material detection.
[0055] Preferably, the training set, test set and validation set are divided into 6:2:2 for training and testing, that is, the training set contains 411 data, the validation set contains 137 data, and the test set contains 138 data.
[0056] Compared with the prior art, the beneficial effects of the present invention are as follows: by collecting abnormal data of oil-immersed transformers, using capsule networks for training, and combining with real-time monitoring data of oil-immersed transformers, real-time judgment and analysis of transformer faults caused by single-point abnormalities and multiple-point abnormalities are performed. Compared with the classic machine learning method, the capsule network innovatively proposes a "vector input, vector output" method, and adopts a dynamic routing mechanism to reduce the feature loss caused by traditional pooling operations, so that the lower-level features are only passed to the matching high-level, more features of the original data are retained, and the accuracy of feature recognition is improved. Compared with the method that requires the addition of additional equipment or additional operations such as using ultrasonic and current detection and passing air into the heat sink to find abnormal points, the method based on the capsule network can effectively save manpower and material costs. There is no need to use manual on-site operation judgment. It only needs to use the gas data in the oil-immersed transformer collected by the sensor and input the capsule network model to make accurate judgments on whether discharge and overheating faults occur in real time, thereby reducing repeated manpower and material detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1It is a schematic diagram of the structure of the technical process of abnormality detection of oil-immersed transformer based on capsule network of the present invention;
[0058] Figure 2 It is a schematic diagram of the capsule network model structure;
[0059] Figure 3 It is a schematic diagram of the accuracy structure of the training and validation sets;
[0060] Figure 4 It is a schematic diagram of the recall structure of the training and validation sets;
[0061] Figure 5 It is a schematic diagram of the F1 value structure of the training and validation sets. DETAILED DESCRIPTION
[0062] In order to facilitate understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.
[0063] Example 1
[0064] like Figures 1 to 5 As shown, the capsule network-based oil-immersed transformer abnormality detection method of the present invention comprises the following steps:
[0065] S1. Obtain a data sample set. The data set uses 686 groups of measured data from the power supply company and DGA data downloaded from some literature, including 52 groups of normal data and 634 groups of fault data. The data set attributes include 6 attributes: H2, CH4, C2H6, C2H4, C2H2, and fault type. The specific content is shown in Table 1. Due to the complex causes of oil-immersed transformer failures, there are many types of fault types. However, whether it is caused by external or internal reasons, the fault will eventually manifest in the form of heat and electricity. Therefore, the fault type is divided into 4 states according to the possible state of the transformer in actual conditions: normal, overheating fault, discharge fault, and discharge and overheating fault. The specific types are shown in Table 2:
[0066] Table 1 Dataset attributes
[0067]
[0068] Table 2 Fault type name
[0069]
[0070] S2, 3-sigma detects and removes abnormal data. Due to sensor failure, unreliable transmission, system maintenance or insufficient data storage, DGA data often contains abnormal and inaccurate values. Low-quality data easily complicates feature analysis and reduces model training effect. By referring to the "3-sigma" method to detect abnormal data in the data set, the implementation of "3-sigma" is based on the empirical rule that 99.7% of the values in the distribution fall within three standard deviations of the mean; the formula of the "3-sigma" method is shown in formula (1):
[0071]
[0072] Where x is a data point, representing a specific data value (such as H2 concentration), x0 is a set of all data points, avg(x0) is the mean of the data point set, and std(x0) represents the standard deviation of the data point set;
[0073] S3. Data processing. As the data vary greatly, as shown in Table 3, some data differ by 3 orders of magnitude. Therefore, for the convenience of training, the data needs to be processed. For the H2 attribute, the ratio of H2 to the total gas is taken. For CH4 and the remaining hydrocarbon gases, the ratio of the gas to the total hydrocarbon is taken, that is:
[0074] H2=H2 / (H2+CH4+C2H6+C2H4+C2H2)
[0075] CH4=CH4 / (CH4+C2H6+C2H4+C2H2)
[0076] C2H6=C2H6 / (CH4+C2H6+C2H4+C2H2)
[0077] C2H4=C2H4 / (CH4+C2H6+C2H4+C2H2)
[0078] C2H2=C2H2 / (CH4+C2H6+C2H4+C2H2)
[0079] Table 3 Partial data of the dataset
[0080]
[0081] Some of the data after data preprocessing are shown in Table 4:
[0082] Table 4 Part of the data after data preprocessing
[0083]
[0084] Since the data values are all between [0, 1), the data normalization is also completed. Then the data samples are divided into a training validation set and a test set in a ratio of 4:1, that is, the training validation set contains 548 data and the test set contains 138 data;
[0085] S4. Input the training data set into the capsule network for training. In the capsule network, unlike the convolutional neural network, the capsule network uses vectors instead of neurons as input and output, and uses the length of the activity vector to represent the probability of entity existence, and uses its direction to represent the instantiation parameter, such as Figure 2 Shown is a simple model of capsule network;
[0086] Specifically, the input vector V1 is multiplied by a weight matrix W1 to obtain a "prediction vector" uj|i, as shown in formula (2):
[0087]
[0088] Among them, ui represents the i-th vector of the upper layer (input layer), Wij represents the weight matrix corresponding to the i-th vector of the upper layer, cij is the coupling coefficient determined by the iterative dynamic routing process, and bij is generally initialized to 0, as shown in formula (3):
[0089]
[0090] The total input to capsule Sj is the weighted sum of all “prediction vectors” uj|i from the next layer, as shown in formula (4):
[0091]
[0092] Different from the activation function of the neural network, the Squashing function is used in the capsule network to compress the capsule Sj, as shown in formula (5):
[0093]
[0094] Finally, update the value of the temporary variable bij, as shown in formula (6):
[0095]
[0096] In the capsule network, the training loss is composed of the margin loss and the reconstruction loss. The margin loss is shown in formula (7):
[0097] L k =T k max(0,m + -‖V k ‖)2 +λ(1-T k )max(0,‖V k ‖-m - ) 2 (7)
[0098] Among them, Tk = 1 if and only if the k category exists, otherwise Tk = 0, m+ = 0.9, m- = 0.1, λ is the proportional coefficient and takes 0.5. Based on the above capsule network model, an oil-immersed transformer anomaly detection model based on capsule network is established, and the training and validation sets are used for model training;
[0099] S5. Use the test data set to test the accuracy of the model. Input the divided data set into the trained model to test the precision, recall and F1 value of the model on the training and validation sets, as well as the accuracy on the test set. The calculation formula is as follows:
[0100]
[0101]
[0102] Among them, TP represents the number of true positive results, FP represents the number of false positive results, TN represents the number of true negative results, and FN represents the number of false negative results. Precision (P) represents the ratio of true positive examples in the predicted results to positive examples in all predicted results. The higher the precision, the higher the ratio of true positive examples in the predicted results of the model, and the stronger the recognition ability of the model for positive samples. Recall (R) represents the ratio of true positive examples in the predicted results to all actual positive examples. The higher the recall, the higher the ratio of positive examples that the model can successfully predict, and the more comprehensive the recognition ability of the model. The F1 value is the harmonic mean of the precision and recall. The higher the F1 value, the better the model performs in terms of accuracy and reliability.
[0103] By dividing the training set, test set and validation set into different proportions for testing, the accuracy of the training and validation sets, the recall rate of the training and validation sets and the F1 value of the training and validation sets are finally obtained in the training and validation sets. Figure 3 , Figure 4 and Figure 5 As shown;
[0104] In this embodiment, by collecting abnormal data of oil-immersed transformers, using capsule networks for training, and combining real-time monitoring data of oil-immersed transformers, real-time judgment and analysis of transformer faults caused by single-point abnormalities and multiple-point abnormalities are performed. Compared with the classic machine learning method, the capsule network innovatively proposes a "vector input, vector output" method, and adopts a dynamic routing mechanism to reduce the feature loss caused by traditional pooling operations, so that the lower-level features are only passed to the matching high-level, more features of the original data are retained, and the accuracy of feature recognition is improved. Compared with the method that requires the addition of additional equipment or additional operations such as using ultrasonic and current detection and passing air into the heat sink to find abnormal points, the method based on the capsule network can effectively save manpower and material costs. There is no need to use manual on-site operation judgment. It only needs to use the gas data in the oil-immersed transformer collected by the sensor and input the capsule network model to make accurate judgments on whether discharge and overheating faults occur in real time, thereby reducing repeated manpower and material detection.
[0105] Example 2
[0106] On the basis of Example 1, Figures 1 to 5 As shown, the capsule network is used to test on the test set, and the test results are shown in Table 5:
[0107] Table 5 Test results of the test set
[0108]
[0109] According to the test results in Table 5, the capsule network model achieves a comprehensive accuracy of 90% in the oil-immersed transformer test data set. Among them, due to the small proportion of normal labels and discharge and overheating fault labels in the overall data set, the training and recognition effects are slightly poor, but for discharge faults and overheating faults, the recognition accuracy can reach more than 92%, effectively identifying the abnormalities of oil-immersed transformers based on gas data;
[0110] A novel capsule network is used to train the existing oil-immersed transformer anomaly dataset to obtain a model that can accurately detect three types of anomalies.
[0111] The training set, test set and validation set are divided into 6:2:2 for training and testing, that is, the training set contains 411 data, the validation set contains 137 data, and the test set contains 138 data;
[0112] In this embodiment, compared with the classic machine learning method, the capsule network innovatively proposes a "vector input, vector output" method, and adopts a dynamic routing mechanism to reduce the feature loss caused by traditional pooling operations, so that the lower-level features are only passed to the matching high-level, retaining more features of the original data and improving the accuracy of feature recognition. Compared with the method of using ultrasonic and current detection, passing air into the heat sink to find abnormal points, etc., which requires the addition of additional equipment or additional operations, the method based on the capsule network can effectively save manpower and material costs. There is no need to use manual on-site operation judgment. It only needs to use the gas data in the oil-immersed transformer collected by the sensor and input it into the capsule network model. It can make accurate judgments on whether discharge and overheating faults occur in real time, reducing repeated manpower and material detection.
[0113] The main functions achieved by the present invention are: compared with the classic machine learning method, the capsule network innovatively proposes the "vector input, vector output" method, and adopts a dynamic routing mechanism to reduce the feature loss caused by the traditional pooling operation, so that the lower-level features are only transmitted to the matching high-level, retaining more features of the original data, and improving the accuracy of feature recognition. Methods that require the addition of additional equipment or additional operations, such as ultrasonic and current detection, and the introduction of air into the heat sink to find abnormal points, can effectively save manpower and material costs, without the need for manual on-site operation and judgment.
[0114] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. The oil-immersed transformer abnormality detection method based on capsule network is characterized by: The following steps are involved: S1. Obtain a data sample set. The data set uses 686 sets of measured data from the power supply company and DGA data downloaded from some literature, including 52 sets of normal data and 634 sets of fault data. The data set attributes include 6 attributes: H2, CH4, C2H6, C2H4, C2H2, and fault type. S2 and 3-sigma detect and remove abnormal data. Due to sensor failure, unreliable transmission, system maintenance or insufficient data storage, DGA data often contains abnormal and inaccurate values. Low-quality data can easily complicate feature analysis and reduce model training results. S4. Input the training data set into the capsule network for training. In the capsule network, unlike the convolutional neural network, the capsule network uses vectors instead of neurons as input and output, and uses the length of the activity vector to represent the probability of entity existence, and uses its direction to represent the instantiation parameter; S5. Use the test data set to test the accuracy of the model. Input the divided data set into the trained model to test the precision, recall and F1 value of the model on the training and validation sets, as well as the accuracy on the test set. The calculation formula is as follows: Among them, TP represents the number of true positive results, FP represents the number of false positive results, TN represents the number of true negative results, FN represents the number of false negative results, and Precision (P) represents the ratio of true positive examples in the predicted results to positive examples in all predicted results.
2. The method for detecting abnormality of an oil-immersed transformer based on capsule network according to claim 1, characterized in that: By dividing the training set, test set, and validation set into different proportions for testing, the accuracy of the training and validation sets, the recall rate of the training and validation sets, and the F1 value of the training and validation sets are finally obtained in the training and validation sets.
3. The method for detecting abnormality of an oil-immersed transformer based on capsule network according to claim 2, characterized in that: The capsule network is used to test the test set, and the test results are shown in Table 5: Table 5 Test results of the test set According to the test results obtained in Table 5, the capsule network model achieves a comprehensive accuracy of 90% in the oil-immersed transformer test data set. Among them, due to the small proportion of normal labels and discharge and overheating fault labels in the overall data set, the training and recognition effects are slightly poor. However, for discharge faults and overheating faults, the recognition accuracy rate reaches more than 92%, effectively identifying the anomalies of the oil-immersed transformer based on the gas data.
4. The method for detecting abnormality of an oil-immersed transformer based on capsule network according to claim 1, characterized in that: A novel capsule network is used to train the existing oil-immersed transformer anomaly dataset, and a model is obtained that can accurately detect three types of anomalies.
5. The method for detecting abnormality of an oil-immersed transformer based on capsule network according to claim 2, characterized in that: The training set, test set and validation set are divided into 6:2:2 for training and testing, that is, the training set contains 411 data, the validation set contains 137 data, and the test set contains 138 data.