A method and device for judging fault indication of surge arrester

Abstract

This invention relates to the field of surge arrester fault detection technology, and particularly to a surge arrester fault indication and judgment method and indicating device. A surge arrester fault indication and judgment method includes S1, historical data acquisition; S2, feature vector generation; S3, a neural network with a multi-head attention mechanism; S4, model training and optimization; and S5, real-time monitoring. This invention establishes a multi-channel input neural network model with a multi-head attention judgment mechanism by collecting lightning current data, leakage current data, and temperature difference data between the surge arrester and the environment. When a sudden lightning strike occurs, the model reacts immediately and reflects the fault situation to the indicating device. For aging detection over a long period, it achieves accurate detection of aging-related factors. The surge arrester fault judgment method established through the above model can realize fault indication for the surge arrester.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lightning arrester fault detection, in particular to a lightning arrester fault indication method and device. BACKGROUND

[0002] The gapless zinc oxide lightning arrester is widely used in the power system and plays an important role in maintaining the stability of the power transmission and distribution line. The conventional lightning arrester detection method is to apply a direct current and detect the leakage current, that is, the detection personnel applies a direct current to the lightning arrester from the outside, detects the size of the leakage current inside the lightning arrester, and thus detects the lightning protection performance of the current lightning arrester. However, for the live lightning arrester actually installed in the power transmission line, it works in a dangerous alternating current environment, and it is difficult for the detection personnel to apply a direct voltage and current during its working process, so the difficulty of detecting whether it is faulty is greatly increased. Therefore, a method for indicating and judging the fault of the lightning arrester through the sensor data of the lightning arrester itself is considered to be developed. SUMMARY

[0003] Other features and advantages of the present application will be set forth in the following specification, and in part will be apparent from the specification, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and attained by the structure particularly pointed out in the specification as well as the appended drawings.

[0004] The present application aims to overcome the above-mentioned deficiencies, and provides a lightning arrester fault indication method and device. By collecting lightning current data, leakage current data and temperature difference data between the lightning arrester itself and the environment, a neural network model with multi-channel input and multi-head attention judgment mechanism is established. The lightning current data is input into the multi-head attention module after the lightning current features are extracted by the one-dimensional convolutional neural network. The attention module sets the highest weight for it. Therefore, when the lightning arrester has a lightning burst, the model reacts first and reflects the fault condition to the indication device. For the aging condition detection with a long time span, the leakage current data and the temperature difference data between the lightning arrester itself and the environment are combined and input into the LSTM neural network, and the weights of the data in the daytime and nighttime are set, so as to realize the accurate detection of the aging influencing factors. The lightning arrester fault judgment method established by the above-mentioned model can realize the fault indication of the lightning arrester.

[0005] The present application provides a lightning arrester fault indication method, comprising:

[0006] S1, historical data collection: for the existing lightning arresters, the lightning strike sensor in the lightning arrester collects lightning current data, the leakage current sensor collects leakage current data, the photovoltaic voltage sensor is arranged at the top of the lightning arrester to collect photovoltaic voltage data, the lightning arrester temperature sensor collects the temperature of the lightning arrester itself, and the environmental temperature sensor collects the external environmental temperature;

[0007] S2, feature vector generation:

[0008] S21, lightning current sequence is generated for the collected lightning current data, which is input into a one-dimensional convolutional neural network to generate lightning current feature data;

[0009] S22, according to the photovoltaic voltage data, it is determined whether the current weather is daytime or night, and a gating value is formed;

[0010] S23, the collected all-day leakage current data is input into the LSTM1 neural network to generate leakage current feature data with daytime and nighttime information respectively;

[0011] S24, the temperature difference data formed by the lightning arrester temperature and the environmental temperature all day is input into the LSTM2 neural network to generate temperature difference feature data with daytime and nighttime information respectively;

[0012] S25, after weighting, the gating value is fused with the leakage current feature data to form one-dimensional final leakage current feature; after weighting, the gating value is fused with the temperature difference feature data to form one-dimensional final temperature difference feature; the final leakage current feature and the final temperature difference feature are formed into aging influence feature data through feature superposition algorithm;

[0013] S3, neural network with multi-head attention mechanism: for the lightning current feature data and aging influence feature data generated in S2 step, a multi-head attention mechanism is added, a plurality of attention heads are set to focus on different parts of the sequence and produce corresponding weights, and the diagnostic result of the lightning arrester is obtained by fusing the classification through weighted summation of the outputs of the plurality of heads, the diagnostic result includes normal, lightning strike and aging;

[0014] S4, model training and optimization: the model established in S1-S3 steps is trained by setting test set and validation set to form a mature lightning arrester fault diagnosis model;

[0015] S5, real-time monitoring: for the lightning arrester that needs real-time monitoring, the data of the lightning arrester is input into the lightning arrester fault diagnosis model, and the diagnostic result is output to realize real-time monitoring.

[0016] In some embodiments, in the step S21, the one-dimensional convolutional neural network specifically comprises an input layer, a convolutional layer, an activation function, and a pooling layer, the input layer of the one-dimensional convolutional neural network is used to receive the lightning current data to generate a lightning current sequence, the convolutional layer is one-dimensional convolution, the activation function is ReLU, the pooling layer adopts global average one-dimensional pooling, and the lightning current feature data output by the pooling layer is input into the multi-head attention module.

[0017] In some embodiments, in the step S22, the collected photovoltaic voltage data is used as an identifier for distinguishing between day and night, and for filtering the influence of short cloud cover on the photovoltaic voltage data. When the photovoltaic voltage is less than 0.3V and lasts for 1h, it is determined as night. When the photovoltaic voltage is greater than or equal to 0.3V and lasts for 1h, it is determined as day. A sigmoid gating function is set according to the determination condition, and a gating value g between 0 and 1 is generated.

[0018] In some embodiments, in the step S25, the specific steps of forming the aging influence feature data by the feature superposition algorithm are as follows: aligning the one-dimensional final leakage current feature and the final temperature difference feature, adding the data at the corresponding positions to form new aging influence feature data, and calculating the proportion of the final leakage current feature and the final temperature difference feature in the new aging influence feature.

[0019] In some embodiments, for the proportion of the final leakage current feature and the final temperature difference feature in the new aging influence feature, if the final output of the arrester fault diagnosis model is aging, the proportion result is displayed together with the aging feature, and if the output result is normal or lightning, the proportion result is not displayed.

[0020] In some embodiments, in the step S3, the specific steps of setting the neural network with multi-head attention mechanism include:

[0021] S41, concatenating the lightning current feature data and the aging influence feature data as the input of the multi-head attention module;

[0022] S42, dividing the input sequence into 3 heads and adding position encoding;

[0023] S43, performing three different linear transformations on the input sequence with position encoding to obtain query Q, key K, and value V;

[0024] S44, normalizing the attention score using the Softmax function to obtain the attention weight;

[0025] S45, weighted sum of the attention weight and the value V;

[0026] S46, fusing features through a fully connected layer and outputting through a linear layer, and the output result is the diagnosis result of the arrester.

[0027] In some embodiments, during the training and optimization process in the S4 step, the mean square error is used as the loss function, the training optimization is realized by the optimizer, and the Adam optimizer is selected as the optimization method of the model parameters.

[0028] A lightning arrester fault indication device comprises:

[0029] A shell is mounted on the lightning arrester;

[0030] A control center is internally provided with a lightning arrester fault diagnosis model, and the control center is arranged in the shell;

[0031] A display screen is connected with the control center and arranged on the surface of the shell;

[0032] An indication fan is provided with three blades and has different colors;

[0033] The indication fan is electrically connected with the control center, and when the lightning arrester fault diagnosis model outputs a diagnosis result, the indication fan rotates to the color corresponding to the diagnosis result.

[0034] By adopting the technical scheme, the beneficial effects of the present application are as follows:

[0035] The present application collects lightning current data, leakage current data, and temperature difference data between the lightning arrester itself and the environment, thereby establishing a neural network model with multi-channel input and multi-head attention judgment mechanism, and the lightning current data is input into the multi-head attention module after being extracted by the one-dimensional convolutional neural network, and the attention module is set with the highest weight, so that when the lightning arrester suddenly strikes, the model reacts first and reflects the fault condition to the indication device, and for the long-time span aging detection, the leakage current data and the temperature difference data between the lightning arrester itself and the environment are combined and input into the LSTM neural network, and the data in the day and night is set with weights, thereby realizing accurate detection of aging influencing factors, and the lightning arrester fault judgment method established by the above model can realize the fault indication of the lightning arrester.

[0036] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure.

[0037] It is certain that such purposes and other purposes of the present application will become more apparent after the detailed description of the preferred embodiments described in the following various drawings and drawings.

[0038] In order to make the above and other purposes, features and advantages of the present application more apparent and easy to understand, one or more preferred embodiments are described in detail below, and the drawings are shown as follows. Attached Figure Description

[0039] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0040] In the accompanying drawings, the same parts use the same reference numerals, and the drawings are schematic and not necessarily drawn to actual scale.

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only one or more embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on such drawings without creative effort.

[0042] Figure 1 This is a schematic diagram of the overall process of the fault indication judgment method in some embodiments of the present invention;

[0043] Figure 2 This is a schematic diagram of the overall structure of the multi-channel input to multi-head attention mechanism neural network in some embodiments of the present invention;

[0044] Figure 3 This is a schematic diagram illustrating the process of extracting timing features of photovoltaic voltage, temperature difference sequence, and leakage current using an LSTM neural network in some embodiments of the present invention.

[0045] Figure 4 This is a schematic diagram of the LSTM neural network structure in some embodiments of the present invention;

[0046] Figure 5 This is a schematic diagram illustrating the weighted and superimposed process between temperature difference characteristics and leakage current characteristics in some embodiments of the present invention.

[0047] Figure 6 This is a schematic diagram of the internal structure of a surge arrester fault indication device in some embodiments of the present invention.

[0048] Explanation of key figure labels:

[0049] 1. Casing; 2. Control center; 3. Indicator fan. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0051] In addition, in the description of the present application, it needs to be understood that the terms "center", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "axial", "radial", "circumferential" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0052] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting", "fixing" and the like should be understood broadly, for example, can be fixed connection, can also be detachable connection, or integral; can be directly connected, or indirectly connected through an intermediate medium, can be the internal communication of two elements or the interaction relationship between two elements. However, it is noted that direct connection means that the connection between the two main bodies does not pass through a transition structure to establish a connection relationship, but only connects through a connecting structure to form a whole. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0053] In the present application, unless otherwise explicitly specified and limited, the first feature is "on" or "under" the second feature, which can be direct contact between the first and second features, or indirect contact between the first and second features through an intermediate medium. In the description of the present specification, the description referring to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0054] Referring to Figures 1-2 , Figure 1 The whole flowchart of the fault indication judgment method in some embodiments of the present application is shown in the figure. Figure 2 The overall structure schematic diagram of the multi-channel input to the multi-head attention mechanism neural network in some embodiments of the present application is shown in the figure.

[0055] According to some embodiments of the present application, the present application provides a lightning arrester fault indication judgment method, comprising:

[0056] S1, historical data collection: for the existing lightning arresters, the lightning strike sensor in the lightning arrester collects lightning current data, the leakage current sensor collects leakage current data, the photovoltaic voltage sensor is arranged at the top of the lightning arrester to collect photovoltaic voltage data, the lightning arrester temperature sensor collects the temperature of the lightning arrester itself, and the environmental temperature sensor collects the external environmental temperature;

[0057] The lightning strike sensor is arranged to collect lightning current data, and when the lightning current data increases suddenly, it indicates that the lightning arrester is currently struck by lightning. The conventional zinc oxide lightning arrester can withstand 20-30 times of nominal discharge current impact, but if the single current exceeds the rated value, the lightning arrester may be broken down and short-circuited, thereby prompting maintenance personnel to replace the lightning arrester in time by jumping to the lightning arrester breakdown state;

[0058] The leakage current data is arranged to detect the aging state of the lightning arrester. As can be seen from the lightning detection process, after each time of nominal discharge current impact, the lightning arrester will age over time. When the leakage current increases to a certain value, the lightning arrester can no longer be used, and maintenance personnel need to be reminded to replace the lightning arrester in time. The leakage current data of the lightning arrester itself will gradually increase over time, and the size of the leakage current will also change with the temperature difference. When the difference between the environmental temperature detected by the photovoltaic voltage sensor and the lightning arrester temperature sensor is large, the leakage current will increase accordingly. When the difference between the environmental temperature and the lightning arrester temperature is small, the leakage current will also decrease accordingly. Therefore, it is necessary to consider whether the lightning arrester has failed due to aging over time or due to a large temperature difference.

[0059] Reference Figures 3-4 , Figure 3 is a process diagram for extracting time sequence features of photovoltaic voltage, temperature difference sequence and leakage current in some embodiments of the application through an LSTM neural network; Figure 4 is an LSTM neural network structure diagram in some embodiments of the application.

[0060] S2, feature vector generation:

[0061] S21, lightning current sequence is generated for the collected lightning current data, and is input into a one-dimensional convolutional neural network to generate lightning current feature data;

[0062] The one-dimensional convolutional neural network specifically includes an input layer, a convolutional layer, an activation function and a pooling layer. The input layer of the one-dimensional convolutional neural network is used to receive lightning current data to generate lightning current sequence. The convolutional layer is one-dimensional convolution, the activation function is ReLU, and the pooling layer adopts global average one-dimensional pooling. The lightning current feature data output by the pooling layer is input into a multi-head attention module.

[0063] S22, determine the current weather as day or night according to photovoltaic voltage data, form a gating value;

[0064] The collected photovoltaic voltage data is used as an identifier to distinguish between day and night, and to filter the influence of short cloud cover on photovoltaic voltage data. When the photovoltaic voltage is less than 0.3V and lasts for 1h, it is determined to be night. When the photovoltaic voltage is greater than or equal to 0.3V and lasts for 1h, it is determined to be day. According to the determination condition, a sigmoid gating function is set to generate a gating value g between 0 and 1.

[0065] S23, the collected leakage current data throughout the day is input into the LSTM1 neural network to generate leakage current feature data with day and night information;

[0066] S24, the temperature difference data formed by the arrester temperature and the ambient temperature throughout the day is input into the LSTM2 neural network to generate temperature difference feature data with day and night information;

[0067] During the day, due to the influence of sunlight radiation, the temperature difference between the arrester and the environment is larger. The arrester and environmental temperature difference is collected, and the temperature difference sequence is input into the LSTM1 neural network to form the day temperature difference data feature. The leakage current under the day working condition is measured, and the temperature difference sequence is input into the LSTM2 neural network to form the day leakage current data feature. The LSTM neural network learns to make the leakage current data feature and the temperature difference data feature have the time sequence information of the day.

[0068] At night, due to the decrease of sunlight radiation, the temperature difference between the arrester and the environment is small under normal working condition. The arrester and environmental temperature difference is collected, and the temperature difference sequence is input into the LSTM1 neural network to form the night temperature difference data feature. The leakage current under the night working condition is measured, and the temperature difference sequence is input into the LSTM2 neural network to form the night leakage current data feature. The LSTM neural network learns to make the leakage current data feature and the temperature difference data feature have the time sequence information of the night.

[0069] The reason for separately learning the time sequence features of day and night is that in the day state, the leakage current is large due to the temperature difference. The same current will be determined as a fault state by the system in the night state. Therefore, the difference between day and night needs to be considered to make it fit the actual use state of the arrester.

[0070] As shown in Figure 4 , the specific steps of capturing time dependence of leakage current feature data or temperature difference feature data are as follows: the leakage current feature sequence or temperature difference feature sequence is input into the forget gate and input gate, respectively. The forget gate discards useless memory to generate cell state The input gate filters new information and generates candidate cell states; the cell states are updated by combining the results generated by the forget gate and the input gate. The output is sent to the hidden layer through the output gate; the activation functions of the hidden layer are the sigmoid function and the tanh function, and the number of units is 128 by default. This is the hidden state from the previous time step. The hidden state after cell renewal is output as leakage current characteristic data or temperature difference characteristic data with time sequence information after passing through a fully connected layer.

[0071] Reference Figure 5 , Figure 5 This is a schematic diagram illustrating the weighted and superimposed process between temperature difference characteristics and leakage current characteristics in some embodiments of the present invention;

[0072] S25. The weighted gate values ​​are fused with the leakage current feature data to form a one-dimensional final leakage current feature; the weighted gate values ​​are fused with the temperature difference feature data to form a one-dimensional final temperature difference feature; the final leakage current feature and the final temperature difference feature are combined using a feature superposition algorithm to form aging effect feature data.

[0073] Learning the day-night characteristics of temperature difference and leakage current through LSTM neural network requires a large amount of data and accurate learning takes a long time. The speed of model judgment of day-night state can be improved by adding prior conditions. That is, the gate value in step S22 is introduced. When the photovoltaic voltage is less than 0.3V and lasts for 1 hour, it is considered night. When the photovoltaic voltage is greater than or equal to 0.3V and lasts for 1 hour, it is considered day. The sigmoid gate function is set according to the judgment conditions.

[0074] Taking daytime and nighttime temperature differences as an example, since the testing personnel set a clear threshold for the photovoltaic voltage of the surge arrester based on actual experience, the sigmoid gate function can be set very steeply, making g very close to 0 (night) or 1 (day). The LSTM neural network first forms the initial daytime and nighttime features of temperature difference and leakage current, and then through weighted fusion with the gate value, the final temperature difference feature = g * daytime temperature difference feature + (1-g) * nighttime temperature difference feature, which makes the model easier to train, faster to draw conclusions, and more versatile, as it does not need to learn the daytime and nighttime patterns precisely.

[0075] The specific steps for forming aging impact feature data through feature overlay algorithm are as follows: Align the one-dimensional final leakage current feature and final temperature difference feature, add the data at corresponding positions of the two to form new aging impact feature data, and calculate the proportion of the final leakage current feature and final temperature difference feature in the new number of aging impact features. If the final diagnosis result of the surge arrester fault diagnosis model is aging, the proportion result is displayed together with the aging feature; if the output result is normal or lightning strike, the proportion result is not displayed.

[0076] When processing multi-source data, the feature overlay algorithm can integrate information from different sensors or datasets, improving the integrity and accuracy of the data. In this application, the feature overlay algorithm can fuse data from temperature sensors and leakage current sensors, enhancing the decision-making ability and accuracy of the detection results. At the same time, the feature overlay algorithm can effectively process and analyze large amounts of data, is adaptable to aging data detection with a long time span, and can improve the data processing capability of the model.

[0077] S3. A neural network with a multi-head attention mechanism: For the lightning current feature data and aging effect feature data generated in step S2, a multi-head attention mechanism is added. Multiple attention heads are set to focus on different parts of the sequence and generate corresponding weights. The outputs of multiple heads are weighted and summed, and the classification is fused to obtain the diagnosis result of the surge arrester. The diagnosis result includes normal, lightning strike, and aging.

[0078] The specific steps for setting up a neural network with a multi-head attention mechanism include:

[0079] S41. The lightning current characteristic data and aging effect characteristic data are concatenated and used as the input of the multi-head attention module. ;

[0080] One-dimensional concatenation can be used to combine feature vectors from different sources to form a comprehensive feature representation when processing multimodal data or multiple feature sources. This helps neural networks better capture the diversity and complexity of data. By concatenating multiple feature vectors, the dimensionality of the input features can be increased, thereby improving the expressive power of the model.

[0081] S42, Input sequence It is divided into 3 headers, and a position code is added;

[0082] Dimensions of each head ,in, Define the total dimension of the model and add the corresponding positional encoding to the input;

[0083]

[0084]

[0085] Where P is the position code, l is the temporal position corresponding to the position code, and k is from 0 to... Integer values ​​between;

[0086] For sequence length L and total model dimension Create an L* Position encoding matrix PE;

[0087] Add the position encoding matrix PE to the input sequence: X= +PE ensures that the input data at each time step... It contains location information, where X is the input sequence after adding location encoding;

[0088] S43. Perform three different linear transformations on the input sequence with added position encoding to obtain query Q, key K and value V;

[0089]

[0090] in, , , These are the weight matrices;

[0091] For each query Q, calculate its attention score with all keys K. ;

[0092] S44. Use the Softmax function to calculate the output of each attention head;

[0093]

[0094] in, For the i-th attention head, Let V be the value of the i-th attention head;

[0095] S45. Features are fused through a fully connected layer and output through three linear layers. The output result is the diagnostic result of the surge arrester.

[0096] The final output of the fused multi-head attention mechanism is:

[0097]

[0098] Here, head1 to head3 are the outputs of the three attention heads, respectively. Used to map the concatenated vectors to the final output space. This is the final output of the multi-head attention mechanism.

[0099] S4. Model Training and Optimization: The model established in steps S1-S3 is trained by setting up a test set and a validation set to form a mature surge arrester fault diagnosis model.

[0100] In the S4 step training and optimization process, the mean squared error is used as the loss function, and the training optimization is achieved through an optimizer. The Adam optimizer is selected as the optimization method for model parameters.

[0101] S5. Real-time monitoring: For surge arresters that require real-time monitoring, the surge arrester data is input into the surge arrester fault diagnosis model, and the diagnosis results are output to achieve real-time monitoring.

[0102] Reference Figure 6 , Figure 6 This is a schematic diagram of the internal structure of a surge arrester fault indication device in some embodiments of the present invention.

[0103] According to some embodiments of the present invention, the present invention provides a surge arrester fault indication device, comprising:

[0104] Casing 1, which is mounted on the surge arrester;

[0105] Control center 2, which is equipped with a surge arrester fault diagnosis model, is located inside the outer casing 1;

[0106] A display screen is connected to the control center 2 and is mounted on the surface of the housing.

[0107] Indicator fan 3, which has three blades of different colors;

[0108] The indicator fan 3 is electrically connected to the control center 2. When the surge arrester fault diagnosis model outputs the diagnosis result, the indicator fan 3 rotates to the color corresponding to the diagnosis result.

[0109] In actual use, lightning strike faults are indicated by red fan blades, aging faults by yellow fan blades, and normal conditions by green fan blades. The PCB circuit board with integrated control center 2 is located above the indicator fan 3 and blocks the two fan blades. When the current fault is detected, the red or yellow fan blades are rotated downwards so that they can be observed from the outside. If the surge arrester is working normally, the green fan blades are rotated downwards.

[0110] It should be understood that the embodiments disclosed herein are not limited to the specific processing steps or materials disclosed herein, but should be extended to equivalent substitutions of such features as understood by those skilled in the art. It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting.

[0111] The term "embodiment" in this specification refers to a specific feature or characteristic described in connection with an embodiment that is included in at least one embodiment of the invention. Therefore, phrases or "embodiments" appearing throughout this specification do not necessarily refer to the same embodiment. Furthermore, the described features or characteristics can be incorporated into one or more embodiments in any other suitable manner. In the above description, specific details, such as thickness and quantity, are provided to provide a comprehensive understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented without the above-described specific details or using other methods, components, materials, etc.

Claims

1. A method of judging a failure indication of a surge arrester, characterized by, Comprising S1, historical data collection: for the existing lightning arrester, the lightning strike sensor in the lightning arrester collects lightning current data, the leakage current sensor collects leakage current data, the photovoltaic voltage sensor is arranged at the top of the lightning arrester to collect photovoltaic voltage data, the lightning arrester temperature sensor collects the temperature of the lightning arrester itself, and the environmental temperature sensor collects the external environmental temperature; S2, feature vector generation: S21, the collected lightning current data is input into a one-dimensional convolutional neural network to generate lightning current sequence, and lightning current feature data is generated; S22, according to the photovoltaic voltage data, it is determined whether the current weather is daytime or night, and a gating value is formed; S23, the collected all-day leakage current data is input into an LSTM1 neural network to generate leakage current feature data with daytime and nighttime information respectively; S24, the temperature difference data formed by the lightning arrester temperature and the environmental temperature is input into an LSTM2 neural network to generate temperature difference feature data with daytime and nighttime information respectively; S25, after weighting, the gating value is fused with the leakage current feature data to form one-dimensional final leakage current feature; After weighting, the gating value is fused with the temperature difference feature data to form one-dimensional final temperature difference feature; the final leakage current feature and the final temperature difference feature are combined through a feature superposition algorithm to form aging influence feature data; S3, neural network with multi-head attention mechanism: for the lightning current feature data and the aging influence feature data generated in step S2, a multi-head attention mechanism is added, a plurality of attention heads are set to focus on different parts of the sequence and produce corresponding weights, and the outputs of the plurality of heads are weighted and summed to fuse the classification to obtain the diagnosis result of the lightning arrester, including normal, lightning strike and aging; S4, model training and optimization: the model established in steps S1-S3 is trained by setting a test set and a validation set to form a mature lightning arrester fault diagnosis model; S5, real-time monitoring: for lightning arresters that need real-time monitoring, the data of the lightning arrester is input into the lightning arrester fault diagnosis model, and the diagnosis result is output to realize real-time monitoring.

2. The method of claim 1, wherein In step S21, the one-dimensional convolutional neural network specifically includes an input layer, a convolutional layer, an activation function and a pooling layer, the input layer of the one-dimensional convolutional neural network is used to receive lightning current data to generate lightning current sequence, the convolutional layer is one-dimensional convolution, the activation function is ReLU, and the pooling layer adopts one-dimensional global average pooling, and the lightning current feature data output by the pooling layer is input into the multi-head attention module.

3. The method of claim 1, wherein the method further comprises: In step S22, the collected photovoltaic voltage data is used as an identifier to distinguish between daytime and nighttime, and is also used to filter the influence of short cloud cover on photovoltaic voltage data. When the photovoltaic voltage is less than 0.3V and lasts for 1h, it is determined as night, and when the photovoltaic voltage is greater than or equal to 0.3V and lasts for 1h, it is determined as daytime. According to the determination condition, a sigmoid gating function is set to generate a gating value g between 0 and 1.

4. The method of claim 1, wherein the method further comprises: In the S25 step, the specific steps of forming the aging influence feature data by the feature superposition algorithm are as follows: aligning the one-dimensional final leakage current feature and the final temperature difference feature, adding the data at the corresponding positions to form new aging influence feature data, and calculating the proportion of the final leakage current feature and the final temperature difference feature in the new aging influence feature data.

5. The method of claim 4, wherein the step of determining the failure of the surge arrester is performed by a method comprising: For the proportion of the final leakage current feature and the final temperature difference feature in the new aging influence feature data, if the final output of the lightning arrester fault diagnosis model is aging, the proportion result is displayed together with the aging feature; if the output result is normal or lightning strike, the proportion result is not displayed.

6. The method of claim 1, wherein In the S3 step, the specific steps of setting the neural network with multi-head attention mechanism include: S41, concatenating the lightning current feature data and the aging influence feature data as the input of the multi-head attention module; S42, dividing the input sequence into 3 heads and adding position encoding; S43, performing three different linear transformations on the input sequence with position encoding to obtain query Q, key K and value V; S44, normalizing the attention score using the Softmax function to obtain the attention weight; S45, weighted sum of the attention weight and the value V; S46, fusing the features through the full connection layer and outputting through the linear layer, and the output result is the diagnosis result of the lightning arrester.

7. The method of claim 1, wherein the method further comprises: In the S4 step of training and optimization, the mean square error is used as the loss function, the optimizer is used to realize the training and optimization, and the Adam optimizer is selected as the optimization method of the model parameters.

8. A surge arrester failure indicator device, characterized by, The device is used to realize the lightning arrester fault indication judgment method in any one of claims 1-7, and the device comprises: a shell mounted on the lightning arrester; a control center with a lightning arrester fault diagnosis model inside, which is arranged in the shell; a display screen connected with the control center and arranged on the surface of the shell; an indication fan with three blades and different colors; the indication fan is electrically connected with the control center, and when the lightning arrester fault diagnosis model outputs the diagnosis result, the indication fan rotates to the color corresponding to the diagnosis result.

Images