Dry-type transformer iron core fault identification method based on BGRU

Abstract

The invention discloses a dry-type transformer iron core fault identification method based on BGRU, and the method comprises the following steps: carrying out the feature extraction of collected voiceprint signals and infrared image sample data through employing a neural network, and calculating a multi-modal fusion feature; optimizing the momentum factor of the neural network by using a dung beetle optimization algorithm; training the optimized neural network by adopting multi-modal fusion features, and obtaining a dry-type transformer iron core loosening voiceprint recognition model after training is completed; inputting a voiceprint signal to be detected into the dry-type transformer iron core loosening voiceprint recognition model, and recognizing and judging whether the dry-type transformer iron core is loosened or not; according to the method, the momentum factors of the neural network are optimized through a dung beetle optimization algorithm, and a more appropriate value is found in different momentum factor combinations, so that the convergence speed of the model is improved, fluctuation in gradient descent can be reduced, and a training curve is smoother.

Description

Technical Field

[0001] The invention relates to the technical field of power equipment fault identification, and in particular to a dry-type transformer core fault identification method based on BGRU. Background Art

[0002] Transformers are one of the most important equipment in the power system, and ensuring their safe and stable operation is the key to the normal development of the entire society. Therefore, effectively identifying the fault type of the transformer through voiceprints and realizing intelligent operation and maintenance of the transformer are of great significance to the normal operation of the entire power system and the development of society.

[0003] The transformer's soundprint signal mainly comes from the vibration of the iron core. When the transformer core fails, the transformer's vibration characteristics will change significantly, thus generating different soundprint signals. For a running transformer, the continuous soundprint signal generated by the iron core contains rich transformer status information, so the transformer's soundprint signal can be used to achieve non-stop online monitoring of the transformer.

[0004] The infrared image of a dry-type transformer can directly reflect the temperature information on the transformer surface, but when the internal fault of the dry-type transformer is not obvious, it is difficult to obtain the internal situation of the transformer through the infrared image. At this time, fusing the voiceprint information with the temperature information for multimodal analysis is a way to solve the problem that it is difficult to accurately identify faults with a single state information. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a dry-type transformer core fault identification method based on BGRU, which aims to solve the problems mentioned in the background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a dry-type transformer core fault identification method based on BGRU, comprising the following steps:

[0007] Step S1: Collecting soundprint signals and infrared image sample data when the dry-type transformer is running normally and when the iron core is loose;

[0008] Step S2: using a bidirectional gated neural network to extract features from the voiceprint signal and infrared image sample data in step S1, and calculating multimodal fusion features;

[0009] Step S3: using the dung beetle optimization algorithm to optimize the momentum factor of the bidirectional gated neural network in step S2;

[0010] Step S4: using multimodal fusion features to train the optimized bidirectional gated neural network, and after the training is completed, a dry-type transformer core loosening soundprint recognition model is obtained;

[0011] Step S5: using the dry-type transformer core loosening soundprint recognition model to perform fault identification and judgment on the dry-type transformer to be tested;

[0012] In step S3, the specific process of optimizing the momentum factor of the neural network using the dung beetle optimization algorithm is as follows:

[0013] Step S301: optimizing the position of the dung beetle in the dung beetle optimization algorithm by using the momentum factor of the bidirectional gated neural network as the position of the dung beetle;

[0014] Step S302: simulating the dung beetle's ball rolling behavior;

[0015] Step S303: simulating the breeding behavior of dung beetles;

[0016] Step S304: simulating the stealing behavior of dung beetles, and introducing a strategy to dynamically change the acceptance of the thieving dung beetles to the local optimal position or the global optimal position according to the strength of the dung beetle's search ability for improvement;

[0017] Step S305: Determine whether the current number of iterations reaches the maximum number of iterations. If it reaches the maximum number of iterations, output the current position as the optimal momentum factor of the bidirectional gated neural network. If it does not reach the maximum number of iterations, return to step S302 to continue iterating.

[0018] Furthermore, in the dung beetle rolling ball behavior, the probability of the dung beetle encountering an obstacle while rolling the dung ball is expressed as:

[0019]

[0020] Where α is a parameter used to determine whether an obstacle is encountered, 1 indicates no obstacle is encountered, and -1 indicates an obstacle is encountered; η is a random number between (0, 1); λ represents the probability of a dung beetle encountering an obstacle, λ∈(0, 1).

[0021] Furthermore, in the dung beetle rolling ball behavior, when the dung beetle does not encounter obstacles when rolling the dung ball, the position update of the dung beetle is expressed as:

[0022]

[0023] Where k is the deviation coefficient, k∈(0, 0.02]; b is a constant, b∈(0, 1); ΔX represents the intensity of light; X ω represents the worst position in the dung beetle population; represents the position information of the i-th dung beetle at the t+1-th iteration; represents the position information of the i-th dung beetle at the t-1th iteration; t is the current iteration number; X i Represents the location information of the i-th dung beetle;

[0024] When the dung beetle encounters an obstacle while rolling the dung ball, the dung beetle's position update is expressed as:

[0025]

[0026] In the formula, θ∈(0,π] represents the deviation coefficient. When θ is 0, When π, the position of the dung beetle is not updated; Represents the position information of the i-th dung beetle at the t-th iteration.

[0027] Furthermore, in the simulation of dung beetle breeding behavior, dung beetles roll the dung balls to a safe area and hide them, providing a safe area for female dung beetles to lay eggs and protect their offspring; the definition of the safe area is expressed as:

[0028]

[0029] In the formula, X * Indicates the optimal position of the dung beetle colony in the current iteration; Lb * , Ub * are the lower and upper limits of the safety area respectively; represents the dynamic selection factor; Lb and Ub are the lower and upper bounds of the optimization problem respectively.

[0030] Furthermore, in the simulation of dung beetle breeding behavior, female dung beetles lay eggs in a safe area. Each female dung beetle will produce an egg ball, and the position update of the egg ball is expressed as:

[0031]

[0032] In the formula, represents the position of the i-th egg ball at the t-th iteration, are all independent random vectors of 1×D, where D is the dimension of the optimization problem;

[0033] When the egg ball position exceeds the safe area, the following processing is performed:

[0034]

[0035] In the formula, B i represents the position of the i-th egg ball.

[0036] Furthermore, the egg balls hatch into small dung beetles that crawl out of the ground to forage for food. The small dung beetles choose the best foraging area for foraging. The range of the best foraging area is defined as:

[0037]

[0038] Where, Lb b ,Ub b are the lower and upper limits of the optimal foraging area respectively; X b is the global best position;

[0039] After determining the best foraging area, the position update of the dung beetle during foraging is expressed as:

[0040]

[0041] In the formula, is a 1×D random vector that follows a normal distribution; is a random vector with values ​​in the range (0,1).

[0042] Furthermore, simulating the stealing behavior of dung beetles, the position update is expressed as:

[0043]

[0044] In the formula, S is a constant; is a 1×D random vector that follows a normal distribution.

[0045] Furthermore, a strategy is introduced to dynamically change the acceptance of the thieving dung beetles to the local optimal position or the global optimal position according to the strength of the dung beetle's search ability. The specific process is: sort the dung beetle population according to the search ability, and record the search ability ranking of the i-th dung beetle at the t iteration as The updated position of the stealing dung beetle based on its own search ability is expressed as:

[0046]

[0047] Where pop is the population size of dung beetles.

[0048] Furthermore, the specific process of step S4 is: randomly divide the multimodal fusion features corresponding to the collected data under the rated no-load voltage level into 75% training sets and 25% test sets, and train and test the optimized bidirectional gated neural network; use other multimodal fusion features except the multimodal fusion features corresponding to the data under the rated no-load voltage level as verification sets to verify the optimized bidirectional gated neural network, and after completion, obtain the dry-type transformer core loose soundprint recognition model.

[0049] Compared with the existing technology, the present invention has the following beneficial effects:

[0050] (1) The present invention optimizes the momentum factor of BGRU through the dung beetle optimization algorithm and finds a more suitable value among different momentum factor combinations, thereby improving the convergence speed of the model and reducing the fluctuation in the gradient descent, making the training curve smoother.

[0051] (2) The present invention uses a bidirectional gated neural network BGRU to extract features from voiceprint signals and infrared image samples. Compared with GRU, the feature extraction of text and image data is more efficient and accurate.

[0052] (3) In order to solve the problem that the dung beetle optimization algorithm is prone to stagnation and cannot converge to the global optimal solution, the present invention introduces a strategy that dynamically changes the acceptance of the thieving dung beetle to the local optimal position or the global optimal position according to the strength of the dung beetle's search ability, thereby improving the global optimization ability of the dung beetle optimization algorithm and enhancing the robustness of the dung beetle optimization algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION

[0054] like Figure 1 As shown, the present invention provides a technical solution: a dry-type transformer core fault identification method based on BGRU, comprising the following steps:

[0055] Step S1: Collecting soundprint signals and infrared image sample data when the dry-type transformer is running normally and when the core is loose.

[0056] In step S1, the specific process of collecting the soundprint signals when the dry-type transformer is running normally and when the core is loose is as follows:

[0057] Install the sensor on the B phase of the high voltage side of the dry-type transformer and loosen the transverse tie rod of the core to collect the dry-type transformer at 60% U N 、80%U N 、100%U N 、120%U N 、(U N It represents the soundprint signal of normal and loose core faults under rated no-load voltage and different rated no-load voltages. The sampling frequency is 20kHz, and a total of 8000 sets of data are collected.

[0058] Step S2: Use a bidirectional gated neural network (BGRU) to extract features from the voiceprint signal and infrared image sample data in step S1 respectively, and calculate the multimodal fusion features.

[0059] The specific process of step S2 is:

[0060] The voiceprint signal and infrared image data of the dry-type transformer are input into the bidirectional gated neural network (BGRU) for forward calculation and reverse calculation respectively to obtain multimodal fusion features.

[0061] The specific process of judging transformer core fault by bidirectional gated neural network (BGRU) is as follows:

[0062] The cross-attention mechanism is used to obtain the attention information of different modal data from the extracted multimodal fusion features. Finally, the concat fusion layer, convolution layer and fully connected layer are used to determine whether the dry-type transformer core is loose.

[0063] Step S3: Use the dung beetle optimization algorithm to optimize the momentum factor of the bidirectional gated neural network (BGRU) in step S2.

[0064] The specific process of step S3 is:

[0065] Step S301: optimizing the position of the dung beetle in the dung beetle optimization algorithm by using the momentum factor of the bidirectional gated neural network (BGRU).

[0066] Step S302: simulating the dung beetle's ball rolling behavior;

[0067] Dung beetles will continuously roll their dung balls. They need light to guide their navigation to ensure that they roll the dung balls in a straight line. However, there is a certain probability that they will encounter obstacles. The probability of encountering obstacles when rolling the dung balls is expressed as:

[0068]

[0069] Where α is a parameter used to determine whether an obstacle is encountered, 1 indicates no obstacle is encountered, and -1 indicates an obstacle is encountered; η is a random number between (0, 1); λ represents the probability of a dung beetle encountering an obstacle, λ∈(0, 1).

[0070] When the dung beetle does not encounter any obstacles while rolling the dung ball, the position update of the dung beetle is expressed as:

[0071]

[0072] Where k is the deviation coefficient, k∈(0, 0.02]; b is a constant, b∈(0, 1); ΔX represents the intensity of light; X ω represents the worst position in the dung beetle population; represents the position information of the i-th dung beetle at the t+1-th iteration; represents the position information of the i-th dung beetle at the t-1th iteration; t is the current iteration number; X i Represents the location information of the i-th dung beetle.

[0073] When the dung beetle encounters an obstacle while rolling the dung ball, it repositions itself by dancing on the dung ball to obtain a new rolling route. The position update of the dung beetle is expressed as:

[0074]

[0075] In the formula, θ∈(0,π] represents the deviation coefficient. When θ is 0, When π, the position of the dung beetle is not updated; Represents the position information of the i-th dung beetle at the t-th iteration.

[0076] Step S303: simulating the breeding behavior of dung beetles;

[0077] Dung beetles roll their dung balls to a safe area and hide them there, in order to provide a safe area for female dung beetles to lay eggs and protect their offspring; the definition of a safe area is expressed as:

[0078]

[0079] In the formula, X * Indicates the optimal position of the dung beetle colony in the current iteration; Lb * , Ub * are the lower and upper limits of the safety area respectively; represents the dynamic selection factor; Lb and Ub are the lower and upper bounds of the optimization problem respectively.

[0080] After finding a safe area, the female dung beetle will lay eggs in the safe area. Each female dung beetle will produce an egg ball, and the position update of the egg ball is expressed as:

[0081]

[0082] In the formula, represents the position of the i-th egg ball at the t-th iteration, are all independent random vectors of 1×D, where D is the dimension of the optimization problem.

[0083] When the egg ball position exceeds the safe area, the following processing is performed:

[0084]

[0085] In the formula, B i represents the position of the i-th egg ball.

[0086] The eggs will hatch into small dung beetles that crawl out of the ground to forage for food. The small dung beetles will choose an optimal foraging area for foraging. The range of the optimal foraging area is defined as:

[0087]

[0088] Where, Lb b ,Ub b are the lower and upper limits of the optimal foraging area respectively; X b is the global best position.

[0089] After determining the best foraging area, the position update of the dung beetle during foraging is expressed as:

[0090]

[0091] In the formula, is a 1×D random vector that follows a normal distribution; is a random vector with values ​​in the range (0,1).

[0092] Step S304: simulating the stealing behavior of dung beetles;

[0093] The stealing mantis will steal the dung balls of other dung beetles near the best stealing location, which is the global optimal position. The position update is expressed as:

[0094]

[0095] In the formula, S is a constant; is a 1×D random vector that follows a normal distribution.

[0096] The thieving dung beetle will update its position according to the global optimal position. It can be seen that the thieving dung beetle will keep approaching the local optimal position and the global optimal position. However, in the early stage of the algorithm, the search range of the group should be expanded as much as possible. Although quickly approaching the local optimal position or the global optimal position can make the algorithm converge quickly, there is also a certain probability that the algorithm will stagnate and fail to converge to the global optimal solution.

[0097] To address the above problem, a strategy is designed to dynamically change the acceptance of the thieving dung beetles to the local optimal position or the global optimal position according to the strength of the dung beetle's search ability. The specific method is as follows: sort the dung beetle population according to its search ability, and record the search ability ranking of the i-th dung beetle at the t iteration as The updated position of the stealing dung beetle based on its own search ability is expressed as:

[0098]

[0099] Where pop is the population size of dung beetles.

[0100] From the above, we can see that the stronger the search ability of the dung beetle, the less it is affected by the local optimal position or the global position when updating its position. This allows the dung beetles with strong search ability to continue to maintain their good search ability and continue searching in a small range to prevent skipping the optimal solution. At the same time, the dung beetles with weak search ability will take larger steps to approach the local optimal position and the global optimal position when updating their position, thereby improving the overall quality of the dung beetle population.

[0101] Step S305: Determine whether the current number of iterations has reached the maximum number of iterations. If so, output the current position as the optimal momentum factor of the bidirectional gated neural network (BGRU). If not, return to step S302 to continue iterating.

[0102] Step S4: Use multimodal fusion features to train the optimized bidirectional gated neural network (BGRU), and after training, obtain a dry-type transformer core loose soundprint recognition model.

[0103] The specific process of step S4 is: randomly divide the multimodal fusion features corresponding to the collected data under the rated no-load voltage level into 75% training sets and 25% test sets, and train and test the optimized bidirectional gated neural network (BGRU); use other multimodal fusion features except the multimodal fusion features corresponding to the data under the rated no-load voltage level as the verification set to verify the optimized bidirectional gated neural network (BGRU), and after completion, obtain the dry-type transformer core loose soundprint recognition model.

[0104] The dry-type transformer core loosening soundprint recognition model obtained after training is evaluated. The evaluation indicators include Accuracy (Ac), Precision (Pc), Recall (R), and F1score (F1), which are expressed as:

[0105]

[0106] Throughout, TN is a true negative sample; FN is a false negative sample; TP is a true sample; FP is a false positive sample; Ac is the accuracy rate, which reflects the overall correctness of the prediction of the dry-type transformer core loose voiceprint recognition model; Pc is the precision rate, which measures the accuracy of the dry-type transformer core loose voiceprint recognition model when predicting a positive sample; Recall is the recall rate, which measures the coverage ability of the dry-type transformer core loose voiceprint recognition model for positive samples; F1 is the F1 score. The higher the F1 score, the more reliable the prediction of the dry-type transformer core loose voiceprint recognition model for positive samples.

[0107] Step S5: Fault identification and judgment of the dry-type transformer to be tested is performed through the dry-type transformer core loosening soundprint recognition model.

[0108] The specific process of step S5 is:

[0109] The voiceprint signal and infrared image data of the dry-type transformer to be tested are input into the dry-type transformer core loosening voiceprint recognition model for forward calculation and reverse calculation respectively to obtain multimodal fusion features.

[0110] The cross-attention mechanism is used to obtain the attention information of different modal data from the extracted multimodal fusion features. Finally, the concat fusion layer, convolution layer and fully connected layer are used to determine whether the dry-type transformer core is loose.

[0111] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dry-type transformer core fault identification method based on BGRU, characterized in that: The steps include: Step S1: Collecting soundprint signals and infrared image sample data when the dry-type transformer is running normally and when the iron core is loose; Step S2: using a bidirectional gated neural network to extract features from the voiceprint signal and infrared image sample data in step S1, and calculating multimodal fusion features; Step S3: using the dung beetle optimization algorithm to optimize the momentum factor of the bidirectional gated neural network in step S2; Step S4: using multimodal fusion features to train the optimized bidirectional gated neural network, and after the training is completed, a dry-type transformer core loosening soundprint recognition model is obtained; Step S5: using the dry-type transformer core loosening soundprint recognition model to perform fault identification and judgment on the dry-type transformer to be tested; In step S3, the specific process of optimizing the momentum factor of the neural network using the dung beetle optimization algorithm is as follows: Step S301: optimizing the position of the dung beetle in the dung beetle optimization algorithm by using the momentum factor of the bidirectional gated neural network as the position of the dung beetle; Step S302: simulating the dung beetle's ball rolling behavior; Step S303: simulating the breeding behavior of dung beetles; Step S304: simulating the stealing behavior of dung beetles, and introducing a strategy to dynamically change the acceptance of the thieving dung beetles to the local optimal position or the global optimal position according to the strength of the dung beetle's search ability for improvement; Step S305: Determine whether the current number of iterations reaches the maximum number of iterations. If it reaches the maximum number of iterations, output the current position as the optimal momentum factor of the bidirectional gated neural network. If it does not reach the maximum number of iterations, return to step S302 to continue iterating.

2. According to the BGRU-based dry-type transformer core fault identification method of claim 1, it is characterized by: In the dung beetle rolling ball behavior, the probability of the dung beetle encountering an obstacle while rolling the dung ball is expressed as: In the formula, α represents a parameter for judging whether an obstacle is encountered, 1 means no obstacle is encountered, and -1 means an obstacle is encountered; η is a random number between (0,1); λ represents the probability of a dung beetle encountering an obstacle, λ∈(0,1).

3. The dry-type transformer core fault identification method based on BGRU according to claim 2 is characterized in that: In the dung beetle rolling ball behavior, when the dung beetle does not encounter obstacles when rolling the dung ball, the dung beetle's position update is expressed as: Where k is the deviation coefficient, k∈(0, 0.02]; b is a constant, b∈(0, 1); αX represents the intensity of light; X ω represents the worst position in the dung beetle population; represents the position information of the i-th dung beetle at the t+1-th iteration; represents the position information of the i-th dung beetle at the t-1th iteration; t is the current iteration number; X i Represents the location information of the i-th dung beetle; When the dung beetle encounters an obstacle while rolling the dung ball, the dung beetle's position update is expressed as: In the formula, θ∈(0,π] represents the deviation coefficient. When , the position of the dung beetle is not updated; Represents the position information of the i-th dung beetle at the t-th iteration.

4. The dry-type transformer core fault identification method based on BGRU according to claim 3 is characterized in that: In the simulation of dung beetle breeding behavior, dung beetles roll dung balls to a safe area and hide them, providing a safe area for female dung beetles to lay eggs and protect their offspring; the definition of the safe area is expressed as: In the formula, X * Indicates the optimal position of the dung beetle colony in the current iteration; Lb * , Ub * are the lower and upper limits of the safety area respectively; represents the dynamic selection factor; Lb and Ub are the lower and upper bounds of the optimization problem respectively.

5. The dry-type transformer core fault identification method based on BGRU according to claim 4 is characterized in that: In the simulation of dung beetle breeding behavior, female dung beetles lay eggs in a safe area. Each female dung beetle will produce an egg ball, and the position update of the egg ball is expressed as: In the formula, represents the position of the i-th egg ball at the t-th iteration, are all independent random vectors of 1×D, where D is the dimension of the optimization problem; When the egg ball position exceeds the safe area, the following processing is performed: In the formula, B i represents the position of the i-th egg ball.

6. The dry-type transformer core fault identification method based on BGRU according to claim 5 is characterized in that: The egg balls hatch into small dung beetles that crawl out of the ground to forage for food. The small dung beetles choose the best foraging area for foraging. The range of the best foraging area is defined as: Where, Lb b , Ub b are the lower and upper limits of the optimal foraging area respectively; X b is the global best position; After determining the best foraging area, the position update of the dung beetle during foraging is expressed as: In the formula, is a 1×D random vector that follows a normal distribution; is a random vector with values ​​in the range (0,1).

7. The dry-type transformer core fault identification method based on BGRU according to claim 6 is characterized in that: Simulating the stealing behavior of dung beetles, the position update is expressed as: In the formula, S is a constant; is a 1×D random vector that follows a normal distribution.

8. The dry-type transformer core fault identification method based on BGRU according to claim 7 is characterized in that: A strategy is introduced to dynamically change the acceptance of the local optimal position or the global optimal position of the thieving dung beetle according to the strength of the dung beetle's search ability. The specific process is: sort the dung beetle population according to the search ability, and record the search ability ranking of the i-th dung beetle at the tth iteration as The updated position of the stealing dung beetle based on its own search ability is expressed as: Where pop is the population size of dung beetles.

9. The dry-type transformer core fault identification method based on BGRU according to claim 8 is characterized in that: The specific process of step S4 is: randomly divide the multimodal fusion features corresponding to the collected data under the rated no-load voltage level into 75% training sets and 25% test sets, and train and test the optimized bidirectional gated neural network; use other multimodal fusion features except the multimodal fusion features corresponding to the data under the rated no-load voltage level as the verification set to verify the optimized bidirectional gated neural network, and after completion, obtain the dry-type transformer core loose soundprint recognition model.

Images