Transformer fault diagnosis method and device based on spatiotemporal sequence classification algorithm

Abstract

In some embodiments of the present application, a transformer fault diagnosis method and device based on a spatiotemporal sequence classification algorithm are disclosed. The method uses a magnetic flux sensor to obtain a first leakage flux signal related to the internal time and space of the transformer. Using a preset spatiotemporal sequence classification algorithm model, the method determines whether the transformer has a small-turn short-circuit fault. When a small-turn short-circuit fault occurs, an alarm message is sent to a staff member, enabling staff to promptly identify potential fault hazards and providing a power grid operating company with an early warning of accidents caused by transformer short circuits, thereby preventing accidents from occurring. The method includes: receiving a first leakage flux signal from the magnetic flux sensor; inputting the first leakage flux signal into a preset spatiotemporal sequence classification algorithm model to output a transformer label; and sending an alarm message to a staff member if the transformer label is a fault label.

Description

Technical Field

[0001] The present application relates to the technical field of transformers, and in particular to a transformer fault diagnosis method and device based on a spatiotemporal sequence classification algorithm. Background Art

[0002] Before a major transformer safety incident occurs, there will be some precursor signs. Proper fault predictive maintenance can avoid accidents and disasters, and prevent economic and even personal safety losses and losses to power grid companies and users.

[0003] When a small-turn short circuit occurs in a transformer, it does not immediately stop functioning. Instead, it continues to operate in a pathological state. This has already had a negative impact on the power system, such as voltage fluctuations and a decrease in power quality. Small-turn short circuits can cause abnormal local voltage fluctuations, impacting power supply quality and potentially terminating the functioning of connected electrical equipment.

[0004] While the transformer's operating parameters may not change significantly in this state, if left unchecked, the consequences could be even more serious. These include further damage to the transformer windings, increased current from the short circuit, and increased short-circuit current, leading to winding deformation and a sharp increase in copper losses, further exacerbating the short-circuit fault. The elevated temperature accelerates aging and damage to the winding insulation, ultimately leading to winding burnout. The overvoltage generated by the short circuit may cause breakdown of the transformer's main or longitudinal insulation, ultimately leading to more serious faults such as phase-to-phase shorts and winding-to-ground shorts. The impact of the short circuit could also damage other transformer components, such as leads and tap changers. Short-circuit faults can also trigger a chain reaction within the power system, causing system oscillations, voltage collapse, and even widespread power outages. Short circuits in transformers located near hub substations or critical transmission lines can have a particularly wide-ranging impact.

[0005] This shows that if the signs of transformer faults can be diagnosed at the early stage, that is, small-turn short circuits, significant property and personal losses can be avoided. Therefore, how to identify transformer faults has become a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention

[0006] The present application provides a transformer fault diagnosis method and device based on a time-space sequence classification algorithm. The method obtains a first leakage flux signal about the time and space inside the transformer through a magnetic flux sensor, and uses a preset time-space sequence classification algorithm model to determine whether the transformer has a small-turn short-circuit fault. When a small-turn short-circuit fault occurs, an alarm message is sent to the staff end, so that the staff can discover the hidden fault in time, provide the power grid operating company with accident warnings caused by transformer short circuits, and prevent accidents from occurring.

[0007] In a first aspect, an embodiment of the present application provides a transformer fault diagnosis method based on a spatiotemporal sequence classification algorithm, wherein a plurality of magnetic flux sensors are arranged around the transformer; the method comprises:

[0008] receiving a first leakage magnetic flux signal sent by the magnetic flux sensor;

[0009] Inputting the first leakage magnetic flux signal into a preset spatiotemporal sequence classification algorithm model to output a transformer label; wherein the preset spatiotemporal sequence classification algorithm model is trained based on the second leakage magnetic flux signal sent by the magnetic flux sensor in a fault state of a small-turn short circuit occurring at different locations of the transformer and in a normal operating state;

[0010] If the transformer tag is a fault tag, an alarm message is sent to the staff end.

[0011] In some embodiments, the transformer is a three-phase transformer; the number of the magnetic flux sensors is eight, and the magnetic flux sensors acquire leakage flux signals in the x, y, and z directions; the preset spatiotemporal sequence classification algorithm model includes a shape transformation module, a first one-dimensional convolution module, a second one-dimensional convolution module, a message passing graph module, and a fully connected layer module; the training steps of the preset spatiotemporal sequence classification algorithm model include:

[0012] determining a duty cycle of the three-phase transformer based on an operating frequency of the three-phase transformer;

[0013] In a fault state in which a small-turn short circuit occurs at different positions of the three-phase transformer and in a normal operating state, respectively, based on a preset number of acquisitions by the magnetic flux sensor in one working cycle, a second leakage magnetic flux signal sent by the magnetic flux sensor of the three-phase transformer is obtained;

[0014] Determining a time slot length based on the preset number of acquisitions and the preset working cycle;

[0015] Determining a total amount of sensing data in the second leakage magnetic flux signals collected by all magnetic flux sensors in one working cycle based on the number of the magnetic flux sensors and the number of directions included in the second leakage magnetic flux signal;

[0016] The second leakage flux signal is formed into data of a first target matrix shape based on the total number of sensing data and the time slot length, wherein the first target matrix shape is batch_size, 24, 128;

[0017] Inputting the input data of the first target matrix shape to the shape transformation module to output data of a second target matrix shape, wherein the second target matrix shape is batch_size×24×8,1,16;

[0018] Passing the data in the second target matrix shape through the first one-dimensional convolution module to output first target data;

[0019] Inputting the first target data into the second one-dimensional convolution module to output second target data;

[0020] inputting the second target data into the message passing graph module to output third target data;

[0021] The third target data is input into the fully connected layer module.

[0022] In some embodiments, the first one-dimensional convolution module includes a first one-dimensional convolution layer, a first one-dimensional batch normalization layer, a first activation layer, a first maximum pooling layer, and a random dropout layer; the step of passing the data of the second target matrix shape through the first one-dimensional convolution module to output the first target data includes:

[0023] Input the data of the second target matrix shape to the first one-dimensional convolutional layer to output data of the third target matrix shape, where the third target matrix shape is batch_size×24×8,64,16;

[0024] Input the data of the third target matrix shape into the first one-dimensional batch normalization layer;

[0025] Input the data output from the first one-dimensional batch normalization layer into the first activation layer;

[0026] Input the data output from the first activation layer to the first maximum pooling layer to output data of a fourth target matrix shape, where the fourth target matrix shape is batch_size×24×8,64,9;

[0027] The data in the fourth target matrix shape is input to the random deactivation layer to output the first target data.

[0028] In some embodiments, the second one-dimensional convolution module includes a second one-dimensional convolution layer, a second one-dimensional batch normalization layer, a second activation layer, a second maximum pooling layer, a first shape transformation unit, a first linear layer, a third one-dimensional batch normalization layer, and a second shape transformation unit; and the step of inputting the first target data into the second one-dimensional convolution module to output the second target data includes:

[0029] Input the first target data into the second one-dimensional convolutional layer to output data of a fifth target matrix shape, where the fifth target matrix shape is batch_size×24×8,18,7;

[0030] Input the data of the fifth target matrix shape into the second one-dimensional batch normalization layer;

[0031] Input the data output by the second one-dimensional batch normalization layer into the second activation layer;

[0032] Input the data output by the second activation layer to the second maximum pooling layer to output data of a sixth target matrix shape, where the sixth target matrix shape is batch_size×24×8,18,4;

[0033] Inputting the data of the sixth target matrix shape into the first shape transformation unit to output data of a seventh target matrix shape, wherein the seventh target matrix shape is batch_size×24,576;

[0034] Input the data of the seventh target matrix shape to the first linear layer to output data of an eighth target matrix shape, where the eighth target matrix shape is batch_size×24,32;

[0035] Inputting the data of the eighth target matrix shape into the third one-dimensional batch normalization layer;

[0036] The data output by the third one-dimensional batch normalization layer is input to the second shape transformation unit to output second target data, where the matrix shape of the second target data is batch_size, 24, 32.

[0037] In some embodiments, the message passing graph module includes a relationship matrix unit, a second linear layer, a fourth one-dimensional batch normalization layer, and a third activation layer; the step of inputting the second target data into the message passing graph module to output the third target data includes:

[0038] inputting the second target data into the relationship matrix unit to generate a relationship matrix;

[0039] Performing a matrix multiplication operation on the relationship matrix and the second target data to obtain data in a tenth target matrix shape, where the tenth target matrix shape is batch_size, 24, 32;

[0040] Input the data of the tenth target matrix shape into the second linear layer to obtain data of the eleventh target matrix shape, where the eleventh target matrix shape is batch_size, 24, 16;

[0041] Input the data of the eleventh target matrix shape into the fourth one-dimensional batch normalization layer;

[0042] The data output by the fourth one-dimensional batch normalization layer is input into the third activation layer to obtain third target data.

[0043] In some embodiments, the relationship matrix unit includes a fourth activation layer and a softmax layer; the step of inputting the second target data into the relationship matrix unit to generate a relationship matrix includes:

[0044] Transpose the second target data, and perform matrix multiplication on the transposed second target data and the second target data to obtain data in a twelfth target matrix shape, where the twelfth target matrix shape is batch_size×24×24;

[0045] Performing a dot product operation on the data of the twelfth target matrix shape and the attenuation matrix, and inputting the resultant data into the fourth activation layer;

[0046] The data output by the fourth activation layer is input into the softmax layer to generate a relationship matrix.

[0047] In some embodiments, the fully connected layer module includes a third shape change unit, a third linear layer, a fourth linear layer, a fifth linear layer, a fifth activation layer, a sixth activation layer, and a seventh activation layer; and the step of inputting the third target data into the fully connected layer module includes:

[0048] Input the third target data to the third shape changing unit to output data of a thirteenth target matrix shape, wherein the thirteenth target matrix shape is batch_size, 384;

[0049] Inputting the data of the thirteenth target matrix shape into the fifth activation layer through the third linear layer to obtain data of a fourteenth target matrix shape, wherein the fourteenth target matrix shape is batch_size,128;

[0050] Inputting the data of the fourteenth target matrix shape into the sixth activation layer through the fourth linear layer to obtain data of a fifteenth target matrix shape, wherein the fifteenth target matrix shape is batch_size,64;

[0051] The data of the fifteenth target matrix shape is input into the seventh activation layer through the fifth linear layer to obtain third target data, where the shape of the third target data is batch_size,28.

[0052] In some embodiments, the fault labels include high-voltage U-phase high short circuit, high-voltage U-phase middle short circuit, high-voltage U-phase low short circuit, high-voltage V-phase high short circuit, high-voltage V-phase middle short circuit, high-voltage V-phase low short circuit, high-voltage W-phase high short circuit, high-voltage W-phase middle short circuit, high-voltage W-phase low short circuit, medium-voltage U-phase high short circuit, medium-voltage U-phase middle short circuit, medium-voltage U-phase low short circuit, medium-voltage V-phase high short circuit, medium-voltage V-phase middle short circuit, medium-voltage V-phase low short circuit, medium-voltage W-phase high short circuit, medium-voltage W-phase middle short circuit, medium-voltage W-phase low short circuit, low-voltage U-phase high short circuit, low-voltage U-phase middle short circuit, low-voltage U-phase low short circuit, low-voltage V-phase high short circuit, low-voltage V-phase middle short circuit, low-voltage V-phase low short circuit, low-voltage W-phase high short circuit, low-voltage W-phase middle short circuit or low-voltage W-phase low short circuit.

[0053] In some embodiments, the four magnetic flux sensors are evenly distributed on one side of the three-phase transformer and are centrally symmetrical with the remaining four magnetic flux sensors about a center point of a coil of the three-phase transformer.

[0054] In a second aspect, an embodiment of the present application provides a transformer fault diagnosis device based on a spatiotemporal sequence classification algorithm, wherein a plurality of magnetic flux sensors are arranged around the transformer; the device comprises:

[0055] a receiving unit, configured to receive a first leakage magnetic flux signal sent by the magnetic flux sensor;

[0056] An input unit is configured to input the first leakage flux signal into a preset spatiotemporal sequence classification algorithm model to output a transformer label; wherein the preset spatiotemporal sequence classification algorithm model is trained based on the second leakage flux signal sent by the flux sensor in a fault state of a small-turn short circuit occurring at different locations of the transformer and in a normal operating state;

[0057] The sending unit is used to send an alarm message to the staff end if the transformer tag is a fault tag.

[0058] In the above embodiment, a transformer fault diagnosis method and device based on a spatiotemporal sequence classification algorithm is provided. The method obtains a first leakage flux signal about the internal time and space of the transformer through a magnetic flux sensor, and uses a preset spatiotemporal sequence classification algorithm model to determine whether the transformer has a small-turn short-circuit fault. When a small-turn short-circuit fault occurs, an alarm message is sent to a staff end, so that the staff can promptly discover the fault hidden danger, provide the power grid operating company with an accident warning caused by the transformer short circuit, and prevent the accident. The method includes: receiving the first leakage flux signal sent by the magnetic flux sensor; inputting the first leakage flux signal into a preset spatiotemporal sequence classification algorithm model to output a transformer label; wherein the preset spatiotemporal sequence classification algorithm model is trained based on the second leakage flux signal sent by the magnetic flux sensor in the fault state of a small-turn short circuit at different locations of the transformer and the normal operating state; if the transformer label is a fault label, an alarm message is sent to the staff end. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A flowchart of a transformer fault diagnosis method based on a spatiotemporal sequence classification algorithm according to some embodiments is exemplarily shown;

[0060] Figure 2 A schematic diagram illustrating a distribution of magnetic flux sensors according to some embodiments is exemplified;

[0061] Figure 3 A flowchart of another transformer fault diagnosis method based on a spatiotemporal sequence classification algorithm according to some embodiments is exemplarily shown;

[0062] Figure 4 A schematic structural diagram of a transformer fault diagnosis device based on a spatiotemporal sequence classification algorithm according to some embodiments is exemplarily shown. DETAILED DESCRIPTION

[0063] In order to make the purpose and implementation of this application clearer, the exemplary implementation of this application will be clearly and completely described below in conjunction with the drawings in the exemplary embodiments of this application. Obviously, the described exemplary embodiments are only part of the embodiments of this application, not all of the embodiments.

[0064] It should be noted that the brief descriptions of terms in this application are only for the purpose of facilitating the understanding of the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise specified, these terms should be understood according to their ordinary and usual meanings.

[0065] In the specification and claims of this application and the accompanying drawings, the terms "first," "second," "third," etc. are used to distinguish similar or similar objects or entities, and are not necessarily intended to limit a particular order or sequence, unless otherwise noted. It should be understood that the terms used in this manner are interchangeable under appropriate circumstances.

[0066] The terms "comprise," "include," and "have," and any variations thereof, are intended to cover but not exclude inclusion; for example, a product or device comprising a list of components is not necessarily limited to all the components expressly listed but may include other components not expressly listed or inherent to such product or device.

[0067] Before a major transformer safety incident occurs, there will be some precursor signs. Proper fault predictive maintenance can avoid accidents and disasters, and prevent economic and even personal safety losses and losses to power grid companies and users.

[0068] When a small-turn short circuit occurs in a transformer, it does not immediately stop functioning. Instead, it continues to operate in a pathological state. This has already had a negative impact on the power system, such as voltage fluctuations and a decrease in power quality. Small-turn short circuits can cause abnormal local voltage fluctuations, impacting power supply quality and potentially terminating the functioning of connected electrical equipment.

[0069] While the transformer's operating parameters may not change significantly in this state, if left unchecked, the consequences could be even more serious. These include further damage to the transformer windings, increased current from the short circuit, and increased short-circuit current, leading to winding deformation and a sharp increase in copper losses, further exacerbating the short-circuit fault. The elevated temperature accelerates aging and damage to the winding insulation, ultimately leading to winding burnout. The overvoltage generated by the short circuit may cause breakdown of the transformer's main or longitudinal insulation, ultimately leading to more serious faults such as phase-to-phase shorts and winding-to-ground shorts. The impact of the short circuit could also damage other transformer components, such as leads and tap changers. Short-circuit faults can also trigger a chain reaction within the power system, causing system oscillations, voltage collapse, and even widespread power outages. Short circuits in transformers located near hub substations or critical transmission lines can have a particularly wide-ranging impact.

[0070] This shows that if the signs of transformer faults can be diagnosed at the early stage, that is, small-turn short circuits, significant property and personal losses can be avoided. Therefore, how to identify transformer faults has become a technical problem that needs to be solved urgently by those skilled in the art.

[0071] In order to solve the above technical problems, an embodiment of the present application provides a transformer fault diagnosis method based on a spatiotemporal sequence classification algorithm. The method obtains a first leakage flux signal about the time and space inside the transformer through a magnetic flux sensor, and uses a preset spatiotemporal sequence classification algorithm model to determine whether the transformer has a small-turn short-circuit fault. When a small-turn short-circuit fault occurs, an alarm message is sent to the staff end, so that the staff can promptly discover the fault hidden dangers, provide the power grid operating company with an accident warning caused by the transformer short circuit, and prevent the accident from occurring. The method includes: receiving the first leakage flux signal sent by the magnetic flux sensor; inputting the first leakage flux signal into a preset spatiotemporal sequence classification algorithm model to output a transformer label; wherein the preset spatiotemporal sequence classification algorithm model is trained based on the second leakage flux signal sent by the magnetic flux sensor under the fault state of a small-turn short circuit at different positions of the transformer and the normal operating state; if the transformer label is a fault label, an alarm message is sent to the staff end.

[0072] Figure 1 A flowchart of a transformer fault diagnosis method based on a spatiotemporal sequence classification algorithm according to some embodiments is exemplarily shown. The method includes S100-S300.

[0073] S100: Receive a first leakage magnetic flux signal sent by the magnetic flux sensor.

[0074] In the embodiment of the present application, the first leakage flux signal is a leakage flux signal collected by the flux sensor when the transformer is working.

[0075] In an embodiment of the present application, a plurality of magnetic flux sensors are arranged around the transformer. In one example, eight magnetic flux sensors may be provided. In some embodiments, the transformer is a three-phase transformer. Four of the eight magnetic flux sensors are evenly distributed on one side of the three-phase transformer and are centrally symmetrical with the remaining four magnetic flux sensors about the center point of the coil of the three-phase transformer. Figure 2 As shown, Figure 2 The locations of the eight magnetic flux sensors are shown in FIG.

[0076] In an embodiment of the present application, a transformer fault diagnosis method based on a spatiotemporal sequence classification algorithm can be built into a transformer terminal unit.

[0077] S200. Input the first leakage flux signal into a preset spatiotemporal sequence classification algorithm model to output a transformer label; wherein the preset spatiotemporal sequence classification algorithm model is trained based on the second leakage flux signal sent by the flux sensor under the fault state of a small-turn short circuit at different positions of the transformer and the normal operating state.

[0078] In some embodiments, the step of inputting the first leakage flux signal into a preset spatiotemporal sequence classification algorithm model to output a transformer label includes: forming target data in the shape of a first target matrix based on the total number of sensor data and the time slot length of the first leakage flux signal, and inputting the target data into the preset spatiotemporal sequence classification algorithm model to output the transformer label. The method for determining the total number of sensor data and the time slot length is described in detail below and will not be repeated here. In this embodiment, the purpose of forming the target data in the shape of a first target matrix from the first leakage flux signal is to convert the first leakage flux signal into the same format as the sample data used to train the preset spatiotemporal sequence classification algorithm model.

[0079] In some embodiments, the transformer is a three-phase transformer; the number of the magnetic flux sensors is eight, and the magnetic flux sensors acquire leakage flux signals in the x, y, and z directions; the preset spatiotemporal sequence classification algorithm model includes a shape transformation module, a first one-dimensional convolution module, a second one-dimensional convolution module, a message passing graph module, and a fully connected layer module; Figure 3 The flowchart of another transformer fault diagnosis method based on a spatiotemporal sequence classification algorithm according to some embodiments is exemplified. The training steps of the preset spatiotemporal sequence classification algorithm model include S400-S1300:

[0080] S400: Determine a duty cycle of the three-phase transformer based on an operating frequency of the three-phase transformer.

[0081] In one example, the operating frequency of the three-phase transformer is 50 Hz, and the operating period of the three-phase transformer is 1 / 50 second.

[0082] S500. In a fault state in which a small-turn short circuit occurs at different positions of the three-phase transformer and in a normal operating state, obtaining a second leakage magnetic flux signal sent by the magnetic flux sensor in a preset working cycle based on a preset number of acquisition times of the magnetic flux sensor in one working cycle.

[0083] In this embodiment, in order to obtain a preset spatiotemporal sequence classification algorithm model that can accurately predict the transformer label, the samples used to train the preset spatiotemporal sequence classification algorithm model include the second leakage flux signal under the fault state of a small-turn short circuit occurring at different positions of the three-phase transformer, and also include the second leakage flux signal under normal operating state.

[0084] In this embodiment, the second leakage magnetic flux signal may be actually collected or obtained by simulation software.

[0085] In this embodiment, the transformer label includes a fault label and a normal operation label. The fault label includes high voltage U phase short circuit at high, high voltage U phase short circuit at middle, high voltage U phase short circuit at low, high voltage V phase short circuit at high, high voltage V phase short circuit at middle, high voltage V phase short circuit at low, high voltage W phase short circuit at high, high voltage W phase short circuit at middle, high voltage W phase short circuit at low, medium voltage U phase short circuit at high, medium voltage U phase short circuit at middle, medium voltage U phase short circuit at low Short circuit, medium voltage V phase high short circuit, medium voltage V phase middle short circuit, medium voltage V phase low short circuit, medium voltage W phase high short circuit, medium voltage W phase middle short circuit, medium voltage W phase low short circuit, low voltage U phase high short circuit, low voltage U phase middle short circuit, low voltage U phase low short circuit, low voltage V phase high short circuit, low voltage V phase middle short circuit, low voltage V phase low short circuit, low voltage W phase high short circuit, low voltage W phase middle short circuit or low voltage W phase low short circuit.

[0086] In one example, the magnetic flux sensor is preset to collect data 16 times in one working cycle. The preset working cycle is eight working cycles. In this case, in both a fault state where a small-turn short circuit occurs at different locations of the three-phase transformer and a normal operating state, the second leakage magnetic flux signal sent by the magnetic flux sensor in the eight working cycles is obtained based on the preset 16 collection times in one working cycle.

[0087] S600: Determine a time slot length based on the preset number of acquisition times and the preset working cycle.

[0088] In this embodiment, the preset number of acquisition times may be 16 times, and the preset acquisition period is 8. In this case, the time slot length is 16×8, ie, 128.

[0089] S700: Determine the total amount of sensing data in the second leakage magnetic flux signals collected by all magnetic flux sensors in one working cycle based on the number of the magnetic flux sensors and the number of directions included in the second leakage magnetic flux signal.

[0090] In this embodiment, determining the total amount of sensing data in the second magnetic leakage flux signal collected by all magnetic flux sensors during a working cycle based on the number of magnetic flux sensors and the number of directions included in the second magnetic leakage flux signal may include multiplying the number of magnetic flux sensors by the number of directions included in the second magnetic leakage flux signal to obtain the total amount of sensing data in the second magnetic leakage flux signal collected by all magnetic flux sensors during the working cycle. In this embodiment, the second magnetic leakage flux signal includes sub-signals in the x, y, and z directions, so the number of directions included in the second magnetic leakage flux signal is three. If there are eight magnetic flux sensors, the total amount of sensing data is 24.

[0091] S800 , forming data of a first target matrix shape based on the total number of sensing data and the time slot length using the second leakage flux signal, wherein the first target matrix shape is batch_size, 24,128.

[0092] In an embodiment of the present application, a large number of second magnetic leakage flux signals can be obtained to form data in the shape of multiple first target matrices. The model is trained using data in the shape of multiple first target matrices. The present application can also obtain other second magnetic leakage flux signals as a verification data set, and use the verification data set to verify the accuracy of the model. In this embodiment, the ratio of the number of second magnetic leakage flux signals in the training data set to the number of second magnetic leakage flux signals in the verification data set can be 4:1. It can be understood that the greater the number of second magnetic leakage flux signals, the more comprehensive the coverage of defects and normal operation conditions, and the better the training effect.

[0093] S900. Input the input data of the first target matrix shape to the shape transformation module to output data of the second target matrix shape, wherein the second target matrix shape is batch_size×24×8,1,16.

[0094] In this embodiment, the shape transformation module primarily segments the data of the first target matrix shape based on the periodic characteristics of each magnetic flux sensor and sequentially reassembles it into a matrix of a new shape to facilitate subsequent feature extraction. In this embodiment, the first target matrix shape is batch_size,24,128, where 24 represents a total of eight magnetic flux sensors. The second leakage flux signal collected by each magnetic flux sensor includes data from three directions. Multiplying the two together yields the total number of sensor data, i.e., 24. Each flux sensor collects data 16 times during one operating cycle of the flux transformer. The algorithm inputs data from eight three-phase transformer operating cycles at a time. Therefore, for each direction of each flux sensor, the algorithm requires 128 time slots of data. Based on the operating characteristics of eight operating cycles and 16 time slots per operating cycle, the shape transformation module further divides the 128 time slots of data into an 8×16 shape, and then integrates the previous data dimensions. Therefore, after the input data of shape batch_size,24,128 passes through the shape transformation module, the matrix shape becomes batch_size×24×8,1,16. The extra dimension of shape 1 in the middle is used to store the features to be extracted later.

[0095] S1000. Pass the data in the shape of the second target matrix through the first one-dimensional convolution module to output first target data.

[0096] In some embodiments, the first one-dimensional convolution module includes a first one-dimensional convolution layer, a first one-dimensional batch normalization layer, a first activation layer, a first maximum pooling layer, and a random dropout layer; the step of passing the data of the second target matrix shape through the first one-dimensional convolution module to output the first target data includes:

[0097] The data of the second target matrix shape is input to the first one-dimensional convolutional layer to output data of a third target matrix shape, where the shape of the third target matrix is ​​batch_size×24×8,64,16.

[0098] In this embodiment, the first one-dimensional convolutional layer primarily extracts more data features in the time dimension. After the input data passes through this layer, the matrix shape transforms from batch_size × 24 × 8, 1, 16 to batch_size × 24 × 8, 64, 16. The 64 represents 64 data features extracted from every 16 time slots of data.

[0099] The data of the third target matrix shape is input into the first one-dimensional batch normalization layer.

[0100] In this embodiment, the first one-dimensional batch normalization layer is mainly used to standardize the input distribution, which can reduce overfitting of the neural network and improve the generalization ability of the algorithm.

[0101] The data output from the first one-dimensional batch normalization layer is input to the first activation layer.

[0102] In this embodiment, the first activation layer can solve the gradient vanishing problem during model training, while enhancing the model's nonlinear learning ability. The first activation layer does not change the data dimension and shape.

[0103] Input the data output from the first activation layer to the first maximum pooling layer to output data of a fourth target matrix shape, where the fourth target matrix shape is batch_size×24×8,64,9;

[0104] In this embodiment, the function of the first maximum pooling layer is to reduce the data dimension, reduce the amount of calculation, and eliminate some noise. After the first maximum pooling layer, the data shape becomes batch_size×24×8,64,9.

[0105] The data in the fourth target matrix shape is input to the random deactivation layer to output the first target data.

[0106] In this embodiment, the random dropout layer can improve the generalization ability of the neural network and reduce the risk of overfitting. The random dropout layer will not change the shape of the data dimension.

[0107] S1100: Input the first target data into the second one-dimensional convolution module to output second target data;

[0108] In this embodiment, the main function of the second one-dimensional convolution module is to compress and refine data features and prepare the data for entering the message passing graph module in the next step.

[0109] The second one-dimensional convolution module includes a second one-dimensional convolution layer, a second one-dimensional batch normalization layer, a second activation layer, a second maximum pooling layer, a first shape transformation unit, a first linear layer, a third one-dimensional batch normalization layer and a second shape transformation unit; the step of inputting the first target data into the second one-dimensional convolution module to output the second target data includes:

[0110] Input the first target data into the second one-dimensional convolutional layer to output data of a fifth target matrix shape, where the fifth target matrix shape is batch_size×24×8,18,7;

[0111] In this embodiment, the second one-dimensional convolution layer compresses and refines data features. After the data passes through the second one-dimensional convolution layer, the shape becomes batch_size×24×8,18,7.

[0112] Input the data of the fifth target matrix shape into the second one-dimensional batch normalization layer;

[0113] In this embodiment, the second one-dimensional batch normalization layer is used to standardize data distribution and reduce overfitting.

[0114] Input the data output by the second one-dimensional batch normalization layer into the second activation layer;

[0115] In this embodiment, the second activation layer is still used to reduce the risk of vanishing gradients and enhance the model's nonlinear learning capabilities. The data dimension and shape remain unchanged after the second one-dimensional batch normalization layer and the second activation layer.

[0116] Input the data output by the second activation layer to the second maximum pooling layer to output data of a sixth target matrix shape, where the sixth target matrix shape is batch_size×24×8,18,4;

[0117] In this embodiment, the second max pooling layer is used to reduce the data dimension and further extract the main information. After the data passes through the second max pooling layer, the dimension shape becomes batch_size × 24 × 8,18,4.

[0118] Inputting the data of the sixth target matrix shape into the first shape transformation unit to output data of a seventh target matrix shape, wherein the seventh target matrix shape is batch_size×24,576;

[0119] In this embodiment, the first shape transformation unit mainly flattens the other data of each magnetic flux sensor direction in batch_size. After the shape transformation, the dimensional shape of the data becomes batch_size×24,576.

[0120] Input the data of the seventh target matrix shape to the first linear layer to output data of an eighth target matrix shape, where the eighth target matrix shape is batch_size×24,32;

[0121] In this embodiment, the function of the first linear layer is to further reduce the data dimension and refine the data features. After passing through the first linear layer, the data dimension shape becomes batch_size×24,32.

[0122] Inputting the data of the eighth target matrix shape into the third one-dimensional batch normalization layer;

[0123] In this embodiment, the function of the third one-dimensional batch normalization layer is still to standardize the data distribution and reduce overfitting. This layer does not change the dimensional shape of the data.

[0124] Input the data output by the third one-dimensional batch normalization layer to the second shape transformation unit to output second target data, where the matrix shape of the second target data is batch_size, 24, 32;

[0125] In this embodiment, the last second shape transformation unit changes the data dimension shape to batch_size, 24, 32 in preparation for subsequent model calculation.

[0126] S1200: Input the second target data into the message passing graph module to output third target data.

[0127] In some embodiments, the message passing graph module includes a relationship matrix unit, a second linear layer, a fourth one-dimensional batch normalization layer, and a third activation layer; the step of inputting the second target data into the message passing graph module to output the third target data includes:

[0128] The second target data is input into the relationship matrix unit to generate a relationship matrix.

[0129] In some embodiments, the relationship matrix unit includes a fourth activation layer and a softmax layer; the step of inputting the second target data into the relationship matrix unit to generate a relationship matrix includes:

[0130] Transpose the second target data, and perform matrix multiplication on the transposed second target data and the second target data to obtain data in a twelfth target matrix shape, where the twelfth target matrix shape is batch_size×24×24;

[0131] Performing a dot product operation on the data of the twelfth target matrix shape and the attenuation matrix, and inputting the resultant data into the fourth activation layer;

[0132] The data output by the fourth activation layer is input into the softmax layer to generate a relationship matrix.

[0133] In this embodiment, the output of the second one-dimensional convolution module is first transposed and then matrix-multiplied with itself to obtain a matrix with a dimensional shape of batch_size × 24 × 24. This operation uses products to calculate the characteristic relationship matrix of each magnetic flux sensor with other magnetic flux sensors at each time slot, i.e., the data in the shape of the twelfth target matrix. This matrix is ​​then dot-producted with an attenuation matrix to weaken the relationships with magnetic flux sensors that are more distant from each other. This attenuation matrix is ​​a square matrix with the same number of rows and columns as the characteristic relationship matrix, with 1s on the diagonal and gradually decreasing to 0s on both sides. The dimensional shape is batch_size, 24, 24. The result is then passed through the fourth activation layer. This fourth activation layer uses leaky_relu, which can retain a very small value even when the value is negative to avoid neuron death. Finally, the softmax layer further strengthens the directional data of magnetic flux sensors with close temporal and spatial relationships and weakens the directional data of magnetic flux sensors with more distant temporal and spatial relationships.

[0134] Performing a matrix multiplication operation on the relationship matrix and the second target data to obtain data in a tenth target matrix shape, where the tenth target matrix shape is batch_size, 24, 32;

[0135] In this embodiment, after the relationship matrix is ​​calculated, a matrix multiplication operation is performed with the output result of the second one-dimensional convolution module. At this time, the data dimension shape is batch_size, 24, 32.

[0136] Input the data of the tenth target matrix shape into the second linear layer to obtain data of the eleventh target matrix shape, where the eleventh target matrix shape is batch_size, 24, 16;

[0137] In this embodiment, the second linear layer is used to reduce the dimension and compress the data features. At this time, the data dimension shape is batch_size, 24, 16.

[0138] Input the data of the eleventh target matrix shape into the fourth one-dimensional batch normalization layer;

[0139] The data output by the fourth one-dimensional batch normalization layer is input into the third activation layer to obtain third target data.

[0140] In this embodiment, data in the shape of the eleventh target matrix is ​​input into the third activation layer through the fourth one-dimensional batch normalization layer. The functions of these two layers are the same as the previous ones, which are to standardize data distribution and improve nonlinear learning ability, thereby improving the generalization ability of the model.

[0141] S1300: Input the third target data into the fully connected layer module.

[0142] The fully connected layer module includes a third shape change unit, a third linear layer, a fourth linear layer, a fifth linear layer, a fifth activation layer, a sixth activation layer, and a seventh activation layer; the step of inputting the third target data into the fully connected layer module includes:

[0143] Input the third target data to the third shape changing unit to output data of a thirteenth target matrix shape, wherein the thirteenth target matrix shape is batch_size, 384;

[0144] Inputting the data of the thirteenth target matrix shape into the fifth activation layer through the third linear layer to obtain data of a fourteenth target matrix shape, wherein the fourteenth target matrix shape is batch_size,128;

[0145] Inputting the data of the fourteenth target matrix shape into the sixth activation layer through the fourth linear layer to obtain data of a fifteenth target matrix shape, wherein the fifteenth target matrix shape is batch_size,64;

[0146] The data of the fifteenth target matrix shape is input into the seventh activation layer through the fifth linear layer to obtain third target data, where the shape of the third target data is batch_size,28.

[0147] In this example, the linear layer and activation layer compress and extract data features and improve the model's generalization ability, respectively. The data passes through three linear and activation layers, and its dimensions change to batch_size, 128, batch_size, 64, and batch_size, 28, respectively.

[0148] In this embodiment, the accuracy of the trained model is basically above 99.5%.

[0149] S300: If the transformer tag is a fault tag, an alarm message is sent to the staff end.

[0150] In the embodiment of the present application, the transformer label includes a fault label and a normal operation label. When the transformer label output by the model is a fault label, it indicates that there is a small-turn short circuit in the transformer. At this time, an alarm message is sent to the staff end, and the staff can view the alarm message through the staff end and determine the specific location of the small-turn short circuit in the transformer based on the specific content of the fault label. For example, when the fault label is a U-phase high short circuit, it indicates that the high-voltage U-phase coil is short-circuited, and the staff can repair the high-voltage U-phase coil.

[0151] The method in the embodiment of the present application has the following advantages: 1. Improve the safety of the transformer. The method in the embodiment of the present application can detect the anomaly in a timely manner when a small-turn short-circuit fault occurs in the transformer. After the power grid staff obtains the anomaly, they will take timely action to effectively avoid the further expansion of the fault and ensure the safety of personnel and property of the substation. 2. Reduce the losses caused by the fault. The method in the embodiment of the present application can alarm at the first time of the fault and inform the fault location, provide maintenance personnel with accurate fault location information, improve fault handling efficiency, and reduce the power outage time caused by fault investigation. It reduces the cost of maintenance, reduces the power outage time caused by the fault for users, and reduces the losses caused by the fault to the power grid.

[0152] In the above embodiment, a method and device for diagnosing early-stage transformer interturn short-circuit faults based on leakage magnetic analysis are provided. This method performs fault diagnosis by monitoring magnetic field changes during transformer operation without disassembling or destroying the transformer, and has the advantage of being non-invasive. This method can also detect magnetic field changes during transformer operation in real time and quickly perform fault diagnosis, and has the advantage of strong real-time performance. Fault diagnosis is performed by combining signal processing technology and machine learning algorithms, and has high accuracy. In addition, this method is applicable to interturn short-circuit fault triage of transformers of various types and sizes, and has wide applicability.

[0153] An embodiment of the present application further provides a transformer fault diagnosis device based on a spatiotemporal sequence classification algorithm, wherein a plurality of magnetic flux sensors are arranged around the transformer. Figure 4 The schematic diagram of the structure of a transformer fault diagnosis device based on a spatiotemporal sequence classification algorithm according to some embodiments is exemplified. The device includes: a receiving unit 401 , an input unit 402 , and a sending unit 403 .

[0154] A receiving unit 401 is configured to receive a first leakage magnetic flux signal sent by the magnetic flux sensor;

[0155] An input unit 402 is configured to input the first leakage flux signal into a preset spatiotemporal sequence classification algorithm model to output a transformer label; wherein the preset spatiotemporal sequence classification algorithm model is trained based on the second leakage flux signal sent by the flux sensor in a fault state of a small-turn short circuit occurring at different locations of the transformer and in a normal operating state;

[0156] The sending unit 403 is configured to send an alarm message to a staff terminal if the transformer tag is a fault tag.

[0157] An embodiment of the present application further provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the transformer fault diagnosis method based on the spatiotemporal sequence classification algorithm are implemented.

[0158] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the transformer fault diagnosis method based on the spatiotemporal sequence classification algorithm are implemented.

[0159] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform, or by hardware.

[0160] Those skilled in the art will understand that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the units or processes in the accompanying drawings are not necessarily required to implement the present application. Those skilled in the art will understand that the units in the devices in the implementation scenario can be distributed in the devices of the implementation scenario according to the implementation scenario description, or can be changed accordingly and located in one or more devices different from the implementation scenario. The units of the above-mentioned implementation scenario can be combined into one unit, or can be further split into multiple sub-units.

[0161] The serial numbers of the above application are for descriptive purposes only and do not represent the advantages or disadvantages of the implementation scenarios. The above disclosure only discloses several specific implementation scenarios of the present application, but the present application is not limited thereto. Any changes that can be conceived by those skilled in the art should fall within the scope of protection of the present application.

Claims

1. A transformer fault diagnosis method based on spatiotemporal sequence classification algorithm, characterized in that: A plurality of magnetic flux sensors are arranged around the transformer; the method comprises: receiving a first leakage magnetic flux signal sent by the magnetic flux sensor; The first leakage flux signal is input into a preset spatiotemporal sequence classification algorithm model to output a transformer label; wherein the preset spatiotemporal sequence classification algorithm model is trained based on the second leakage flux signal sent by the flux sensor under the fault state of a small-turn short circuit at different positions of the transformer and the normal operating state; the transformer is a three-phase transformer; the number of the flux sensors is eight, and the flux sensors acquire leakage flux signals in the three directions of x, y and z; the preset spatiotemporal sequence classification algorithm model includes a shape transformation module, a first one-dimensional convolution module, a second one-dimensional convolution module, a message passing graph module and a fully connected layer module; the training steps of the preset spatiotemporal sequence classification algorithm model include: determining the working cycle of the three-phase transformer based on the working frequency of the three-phase transformer; in the fault state of a small-turn short circuit at different positions of the three-phase transformer and the normal operating state, respectively based on the preset number of acquisitions of the flux sensor in one working cycle, acquiring the second leakage flux signal sent by the flux sensor of the three-phase transformer within the preset working cycle; based on the The preset number of acquisitions and the preset working cycle are used to determine the time slot length; based on the number of magnetic flux sensors and the number of directions contained in the second leakage magnetic flux signal, the total number of sensing data in the second leakage magnetic flux signal collected by all magnetic flux sensors in one working cycle is determined; the second leakage magnetic flux signal is formed into data of a first target matrix shape based on the total number of sensing data and the time slot length, wherein the first target matrix shape is batch_size, 24, 128; the input data of the first target matrix shape is input into the shape transformation module to output data of a second target matrix shape, wherein the second target matrix shape is batch_size×24×8, 1, 16; the data of the second target matrix shape is passed through the first one-dimensional convolution module to output the first target data; the first target data is input into the second one-dimensional convolution module to output the second target data; the second target data is input into the message passing graph module to output the third target data; the third target data is input into the fully connected layer module; If the transformer tag is a fault tag, an alarm message is sent to the staff end.

2. The method according to claim 1, characterized in that The first one-dimensional convolution module includes a first one-dimensional convolution layer, a first one-dimensional batch normalization layer, a first activation layer, a first maximum pooling layer and a random dropout layer; The step of passing the data of the second target matrix shape through the first one-dimensional convolution module to output the first target data includes: Input the data of the second target matrix shape to the first one-dimensional convolutional layer to output data of the third target matrix shape, where the third target matrix shape is batch_size×24×8,64,16; Input the data of the third target matrix shape into the first one-dimensional batch normalization layer; Input the data output from the first one-dimensional batch normalization layer into the first activation layer; Input the data output from the first activation layer to the first maximum pooling layer to output data of a fourth target matrix shape, where the fourth target matrix shape is batch_size×24×8,64,9; The data in the fourth target matrix shape is input to the random deactivation layer to output the first target data.

3. The method according to claim 2, characterized in that The second one-dimensional convolution module includes a second one-dimensional convolution layer, a second one-dimensional batch normalization layer, a second activation layer, a second maximum pooling layer, a first shape transformation unit, a first linear layer, a third one-dimensional batch normalization layer and a second shape transformation unit; The step of inputting the first target data into the second one-dimensional convolution module to output second target data includes: Input the first target data into the second one-dimensional convolutional layer to output data of a fifth target matrix shape, where the fifth target matrix shape is batch_size×24×8,18,7; Input the data of the fifth target matrix shape into the second one-dimensional batch normalization layer; Input the data output by the second one-dimensional batch normalization layer into the second activation layer; Input the data output by the second activation layer to the second maximum pooling layer to output data of a sixth target matrix shape, where the sixth target matrix shape is batch_size×24×8,18,4; Inputting the data of the sixth target matrix shape into the first shape transformation unit to output data of a seventh target matrix shape, wherein the seventh target matrix shape is batch_size×24,576; Input the data of the seventh target matrix shape to the first linear layer to output data of an eighth target matrix shape, where the eighth target matrix shape is batch_size×24,32; Inputting the data of the eighth target matrix shape into the third one-dimensional batch normalization layer; The data output by the third one-dimensional batch normalization layer is input to the second shape transformation unit to output second target data, where the matrix shape of the second target data is batch_size, 24, 32.

4. The method according to claim 3, characterized in that The message passing graph module includes a relationship matrix unit, a second linear layer, a fourth one-dimensional batch normalization layer and a third activation layer; The step of inputting the second target data into the message passing graph module to output third target data comprises: inputting the second target data into the relationship matrix unit to generate a relationship matrix; Performing a matrix multiplication operation on the relationship matrix and the second target data to obtain data in a tenth target matrix shape, where the tenth target matrix shape is batch_size, 24, 32; Input the data of the tenth target matrix shape into the second linear layer to obtain data of the eleventh target matrix shape, where the eleventh target matrix shape is batch_size, 24, 16; Input the data of the eleventh target matrix shape into the fourth one-dimensional batch normalization layer; The data output by the fourth one-dimensional batch normalization layer is input into the third activation layer to obtain third target data.

5. The method according to claim 4, characterized in that The relationship matrix unit includes a fourth activation layer and a softmax layer; the step of inputting the second target data into the relationship matrix unit to generate a relationship matrix includes: Transpose the second target data, and perform matrix multiplication on the transposed second target data and the second target data to obtain data in a twelfth target matrix shape, where the twelfth target matrix shape is batch_size×24×24; Performing a dot product operation on the data of the twelfth target matrix shape and the attenuation matrix, and inputting the resultant data into the fourth activation layer; The data output by the fourth activation layer is input into the softmax layer to generate a relationship matrix.

6. The method according to claim 5, characterized in that The fully connected layer module includes a third shape change unit, a third linear layer, a fourth linear layer, a fifth linear layer, a fifth activation layer, a sixth activation layer and a seventh activation layer; The step of inputting the third target data into the fully connected layer module includes: Input the third target data to the third shape changing unit to output data of a thirteenth target matrix shape, wherein the thirteenth target matrix shape is batch_size, 384; Inputting the data of the thirteenth target matrix shape into the fifth activation layer through the third linear layer to obtain data of a fourteenth target matrix shape, wherein the fourteenth target matrix shape is batch_size,128; Inputting the data of the fourteenth target matrix shape into the sixth activation layer through the fourth linear layer to obtain data of a fifteenth target matrix shape, wherein the fifteenth target matrix shape is batch_size,64; The data of the fifteenth target matrix shape is input into the seventh activation layer through the fifth linear layer to obtain third target data, and the shape of the third target data is batch_size,28.

7. The method according to claim 6, characterized in that The fault labels include high-voltage U-phase short circuit at high point, high-voltage U-phase short circuit at middle point, high-voltage U-phase short circuit at low point, high-voltage V-phase short circuit at high point, high-voltage V-phase short circuit at middle point, high-voltage V-phase short circuit at low point, high-voltage W-phase short circuit at high point, high-voltage W-phase short circuit at middle point, high-voltage W-phase short circuit at low point, medium-voltage U-phase short circuit at high point, medium-voltage U-phase short circuit at middle point, medium-voltage U-phase short circuit at low point, medium-voltage V-phase short circuit at high point, medium-voltage V-phase short circuit at middle point, medium-voltage V-phase short circuit at low point, medium-voltage W-phase short circuit at high point, medium-voltage W-phase short circuit at middle point, medium-voltage W-phase short circuit at low point, low-voltage U-phase short circuit at high point, low-voltage U-phase short circuit at middle point, low-voltage U-phase short circuit at low point, low-voltage V-phase short circuit at high point, low-voltage V-phase short circuit at middle point, low-voltage V-phase short circuit at low point, low-voltage W-phase short circuit at high point, low-voltage W-phase short circuit at middle point or low-voltage W-phase short circuit at low point.

8. The method according to claim 7, characterized in that The four magnetic flux sensors are evenly distributed on one side of the three-phase transformer and are centrally symmetrical with the remaining four magnetic flux sensors about the center point of the coil of the three-phase transformer.

9. A transformer fault diagnosis device based on spatiotemporal sequence classification algorithm, characterized in that: A plurality of magnetic flux sensors are arranged around the transformer; the device comprises: a receiving unit, configured to receive a first leakage magnetic flux signal sent by the magnetic flux sensor; An input unit is used to input the first leakage flux signal into a preset spatiotemporal sequence classification algorithm model to output a transformer label; wherein the preset spatiotemporal sequence classification algorithm model is trained based on the second leakage flux signal sent by the flux sensor under the fault state of a small-turn short circuit occurring at different positions of the transformer and the normal operating state; the transformer is a three-phase transformer; the number of the flux sensors is eight, and the flux sensors acquire leakage flux signals in three directions of x, y and z; the preset spatiotemporal sequence classification algorithm model includes a shape transformation module, a first one-dimensional convolution module, a second one-dimensional convolution module, a message passing graph module and a fully connected layer module; the training steps of the preset spatiotemporal sequence classification algorithm model include: determining the working cycle of the three-phase transformer based on the working frequency of the three-phase transformer; acquiring the second leakage flux signal sent by the flux sensor of the three-phase transformer within the preset working cycle based on the preset number of acquisitions of the flux sensor in one working cycle in the fault state of a small-turn short circuit occurring at different positions of the three-phase transformer and the normal operating state; Based on the preset number of acquisitions and the preset working cycle, the time slot length is determined; based on the number of magnetic flux sensors and the number of directions contained in the second leakage magnetic flux signal, the total number of sensing data in the second leakage magnetic flux signal collected by all magnetic flux sensors in one working cycle is determined; the second leakage magnetic flux signal is formed into data of a first target matrix shape based on the total number of sensing data and the time slot length, wherein the first target matrix shape is batch_size, 24, 128; the input data of the first target matrix shape is input into the shape transformation module to output data of a second target matrix shape, wherein the second target matrix shape is batch_size×24×8, 1, 16; the data of the second target matrix shape is passed through the first one-dimensional convolution module to output the first target data; the first target data is input into the second one-dimensional convolution module to output the second target data; the second target data is input into the message passing graph module to output the third target data; the third target data is input into the fully connected layer module; The sending unit is used to send an alarm message to the staff end if the transformer tag is a fault tag.

Images