A Transformer Partial Discharge Detection Method Based on Neural Network
By combining ultrasonic signals and ultra-high frequency electromagnetic wave signals in transformer partial discharge detection, and using the Kuyu optimization algorithm to optimize the probabilistic neural network model, the problem of inaccurate identification in transformer partial discharge detection is solved, the accuracy and efficiency of detection are improved, and the stable operation of the power system is supported.
Patent Information
- Application Number
- CN202411066900.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-08-06
AI Technical Summary
Existing technologies for detecting partial discharge in transformers suffer from inaccurate identification, particularly due to inconsistent skill levels among maintenance personnel and the use of a single signal acquisition method, leading to misjudgments and missed detections, which affect the assessment of transformer operating status.
A neural network-based approach is adopted, combining ultrasonic signals and ultra-high frequency electromagnetic wave signals. The probabilistic neural network model is optimized using the bitter fish optimization algorithm to identify and assess the risk of partial discharge types. The smoothing factor of the probabilistic neural network model is used to adjust the model sensitivity and improve the detection accuracy.
By combining multiple signal acquisition methods and improving the algorithm optimization model, false judgments are reduced, the accuracy and efficiency of transformer partial discharge detection are improved, and more reliable technical support is provided for the safe and stable operation of the power system.
Smart Images

Figure CN118656731B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of online detection of electric power equipment, and in particular relates to a transformer partial discharge detection method based on a neural network. Background Art
[0002] In the power system, oil-immersed power transformers play a vital role. Any failure of the transformer may cause a large-scale power outage. Therefore, it is particularly important to detect the status of the transformer and ensure its stable operation for the stable operation of the power system. Due to the air gap inside the insulation, aging wear and overvoltage, the transformer will have partial discharge and other faults that are difficult to be directly observed by the naked eye. Partial discharge will add to the deterioration of the insulation parts. Over time, the vicious cycle will eventually lead to insulation failure of the power equipment.
[0003] When partial discharge occurs in the transformer, a series of ultrasonic signals will be generated. These ultrasonic signals are transmitted inside the equipment in the form of sound wave vibrations, and the ultrasonic probe installed in the equipment will capture these vibration signals in real time and transmit them to the integrated processing module for subsequent analysis. At the same time, in order to capture the ultra-high frequency electromagnetic waves (frequency range is usually between 500MHz and 2000MHz) released with partial discharge, an ultra-high frequency electromagnetic wave signal collector is usually used. It can effectively collect the ultra-high frequency electromagnetic wave signals generated during partial discharge and transmit these signals to the integrated processing module. The integrated processing module performs feature processing on the signals of each module.
[0004] At present, many experts and scholars have conducted in-depth research and analysis on the partial discharge phenomenon of transformers, and found that common types of partial discharge in transformers include: suspended potential discharge, tip corona discharge, cavitation discharge, and discharge caused by metal particles. Although the discharge laws of these common types of partial discharge have been mastered, there are still some challenges in practical applications. On the one hand, due to the uneven professional skills of operation and maintenance personnel, they may not be able to fully master the technology of analyzing, processing, and judging the collected partial discharge signals. This may lead to inaccurate identification of partial discharge phenomena and affect the evaluation of the operating status of the transformer. On the other hand, the currently commonly used partial discharge signal acquisition method is relatively single, usually relying only on ultrasonic signals or ultra-high frequency signals for partial discharge research and judgment. This single acquisition method may not be able to fully reflect the complex situation of partial discharge inside the transformer, resulting in poor judgment, missed judgment, misjudgment, etc. Therefore, a more accurate partial discharge analysis and judgment method is urgently needed to support the safe, stable and efficient operation of the power system. This method should be able to comprehensively consider a variety of signal acquisition methods and combine advanced signal processing technology to accurately and reliably identify and evaluate partial discharge phenomena inside the transformer, thereby providing more powerful technical support for the operation of the power system.
[0005] The bitter fish optimization algorithm establishes an algorithm model by simulating the mating behavior of male and female bitter fish with the help of oysters. The algorithm enriches the method of feature selection for data sets. Optimizing and applying it to online detection of partial discharge of transformers can improve detection efficiency and accuracy. Summary of the invention
[0006] In response to the current challenges in transformer partial discharge detection, the present invention collects signal data generated by transformer partial discharge and preprocesses the signal data; after the preprocessing, a feature data set that can characterize the partial discharge characteristics is extracted; the feature data set is sent to a probabilistic neural network model optimized by the bitter fish algorithm to realize online discrimination of the risk type of partial discharge.
[0007] A transformer partial discharge detection method based on a neural network comprises the following steps:
[0008] Step S1: Collect signal data of partial discharge during the historical operation of the transformer, mark the signal data according to the risk type, construct a data set, and pre-process the data set to obtain a feature data set;
[0009] Step S2: construct a probabilistic neural network model, import the feature data set in step S1 into the probabilistic neural network model for training, and during the training process, optimize the smoothing factor of the probabilistic neural network model by using the bitter fish algorithm to obtain the optimal smoothing factor;
[0010] Step S3: collecting signal data of partial discharge when the transformer is running in real time and preprocessing it, importing the preprocessed signal data into the trained probabilistic neural network model for detection, and outputting the risk type of partial discharge of the transformer in real time;
[0011] In step S2, the process of optimizing the smoothing factor of the probabilistic neural network model by the bitter fish algorithm is as follows:
[0012] Step S21: Initialize the bitter fish population in the search space;
[0013] Step S22: Searching for oysters;
[0014] Step S23: discarding oysters;
[0015] Step S24: generating new bitter fish;
[0016] Step S25: hunting bitter fish;
[0017] Step S26: Determine whether the maximum number of iterations has been reached. When the maximum number of iterations has been reached, calculate the fitness of the bittern at the current number of iterations, and select the bittern with the maximum fitness as the optimal smoothing factor. When the maximum number of iterations has not been reached, continue with steps S22 to S26.
[0018] Further, the signal data in step S1 includes ultrasonic signals and ultra-high frequency electromagnetic wave signals;
[0019] The data set in step S1 is represented as:
[0020] P = [ sc 1 , 1 ⋯ sc 1 , n ⋯ sd 1 , 1 ⋯ sd 1 , n ⋮ ⋱ ⋮ ⋮ ⋮ ⋱ ⋮ scm , 1 ⋯ scm , n ⋯ sdm , 1 ⋯ sdm , n ] ;
[0021] In the formula, represents a dataset, Represents the ultrasonic signal of the first sampling point in the first sampling interval; represents the ultrasonic signal of the first sampling point in the mth sampling interval, Represents the ultrasonic signal of the nth sampling point in the first sampling interval; represents the ultrasonic signal of the nth sampling point in the mth sampling interval, Represents the UHF electromagnetic wave signal of the first sampling point in the first sampling interval, Represents the ultra-high frequency electromagnetic wave signal of the first sampling point in the mth sampling interval, Represents the ultra-high frequency electromagnetic wave signal of the nth sampling point in the first sampling interval, Represents the ultra-high frequency electromagnetic wave signal of the nth sampling point in the mth sampling interval;
[0022] The step S1 of preprocessing the data set to obtain the feature data set is as follows:
[0023] Step S11: normalizing the data set, thereby adjusting the value range of the data set to the interval between 0 and 1, and obtaining a normalized data set;
[0024] Step S12: Calculate the difference between each signal data of the normalized data set and the average value of the signal data of the normalized data set, compare the absolute value of the obtained difference with the preset abnormal threshold, and when the absolute value of the difference is equal to or greater than the abnormal threshold, retain the signal data; when the absolute value of the difference is less than the abnormal threshold, use the average value of the signal data of the normalized data set to replace the signal data, thereby forming a feature data set.
[0025] Furthermore, the risk types in step S1 include: not discharged, low risk, medium risk and high risk.
[0026] Further, in step S2, the probabilistic neural network model includes an input layer, a pattern layer, a summation layer and an output layer;
[0027] The input layer receives the real-time signal data of partial discharge during transformer operation in step S1 as input and imports it into the pattern layer. The pattern layer calculates the degree of matching between the real-time signal data and the preset feature data set and outputs it as a probability density. The summation layer accumulates the probability density of each risk type of partial discharge and calculates the comprehensive probability according to the output of the pattern layer. The output layer outputs the predicted risk type based on the comprehensive probability calculated by the summation layer.
[0028] Among them, the smoothing factor is used to adjust the probability density of the pattern layer output and thus adjust the sensitivity of the probabilistic neural network model classification.
[0029] Furthermore, step S21 is specifically as follows: creating an initial population, which is expressed as:
[0030] F 0 = [ F 1 1 F 1 2 ⋯ F 1 D F 2 1 F 2 2 ⋯ F 2 D ⋮ ⋮ ⋮ ⋮ F n 1 F n 2 ⋯ F n D ] ;
[0031] In the formula, The matrix representing the initial population of bitter fish, represents a matrix element in the matrix, It means a bitter fish. and Respectively represent the row and column position of each matrix element in the matrix, that is, the position of the bitter fish;
[0032] Each bittern in the initial population is evaluated as:
[0033] Fitness = [ f ( F 1 1 ) f ( F 1 2 ) ⋯ f ( F 1 D ) f ( F 2 1 ) f ( F 2 2 ) ⋯ f ( F 2 D ) ⋮ ⋮ ⋮ ⋮ f ( F n 1 ) f ( F n 2 ) ⋯ f ( F n D ) ] ;
[0034] in, represents the fitness value matrix, represents the fitness function;
[0035] The fitness function is expressed as:
[0036] ;
[0037] In the formula, represents fitness, Represents the output of the probabilistic neural network model The predicted risk type, Indicates The risk type is marked. is the number of predicted values.
[0038] Furthermore, the oyster search in step S22 is specifically as follows:
[0039] Step S221: The occupancy status of the oyster is as follows:
[0040] ;
[0041] In the formula, represents the number of iterations, Indicated in In the iteration The location of the bittern that caught oysters, Indicated in In the iteration The location of the bittern that caught oysters, The best oysters that bitter fish seek, represents one of the randomly selected oyster populations, and Each represents a random number between 0 and 1. Indicates the current location of the oyster. indicates the rate at which bitterlings escape from or approach oysters;
[0042] Step S222: Generate a random number that conforms to the Cauchy distribution, so as to increase the rate at which the bittern escapes from or approaches the oyster. The probability density function of the Cauchy distribution is It is expressed as:
[0043] ;
[0044] in, Represents positional parameters, represents the scale parameter, Input representing the probability density function of the Cauchy distribution;
[0045] Use the random number generated by Cauchy distribution as the random factor, expressed as:
[0046] ;
[0047] In the formula, represents the random factor, Represents a random number uniformly distributed between 0 and 1;
[0048] The random factor Multiply by the current speed of the bitter fish , get the updated speed , expressed as:
[0049] ;
[0050] Step S223: During the simulation, the rate at which the bitter fish escape from or approach the oyster will decrease as the number of iterations increases. At this time, the global search becomes a local search. The expression of the rate at which the bitter fish escapes from or approaches the oyster in each iteration is as follows:
[0051] ;
[0052] in, represents the rate at which bitterlings escape from or approach oysters in the first iteration, represents the maximum number of iterations, Indicates that bitter fish is in The rate of escaping or approaching oysters in iterations, Represents a random function, used to generate a random sequence;
[0053] The process of generating random sequences by random functions is expressed as:
[0054] ;
[0055] Presets Equal to 1, the formula is:
[0056] ;
[0057] Step S224: As the number of iterations increases, the oyster occupancy state formula in step S221 In the condition, reduce The value of to reduce the spatial duplication around the optimal solution is as follows:
[0058] ;
[0059] in, Indicates the area where the oyster currently exists. represents the adjustment parameter; the smaller the probability of deviation when performing two types of searches in the oyster occupancy state expression, represents a random real function.
[0060] Furthermore, the discarding of oysters in step S23 is specifically as follows:
[0061] Calculate the position of bitterlings that have not caught oysters, expressed as:
[0062] ;
[0063] in, Indicated in In the iteration The location of the bittern that caught no oysters, Indicated in In the iteration The location of the bittern that caught no oysters, represents the average position of the bitter fish population, and All represent preset adjustment parameters;
[0064] The average position of bitterlings that did not catch oysters is expressed as follows:
[0065] ;
[0066] in, represents the number of times the optimal solution is evaluated by the algorithm, .
[0067] The average position of bitterlings that failed to capture oysters is updated by random numbers to continue searching for oysters. The formula is as follows:
[0068] ;
[0069] In the formula, is the updated position of the average position of bitterlings that did not catch oysters, represents random numbers from a standard normal distribution, represents the strength coefficient;
[0070] examine Whether it exceeds the boundary of the search space, Adjustments are made to ensure that they are within the valid search space, as shown below:
[0071] ;
[0072] in, Indicates the adjusted position of the bitter fish. and denote the minimum and maximum bounds of the search space respectively.
[0073] Furthermore, the production of new bitter fish in step S24 is specifically as follows;
[0074] Generate new bitter fish around the current bitter fish position, as follows:
[0075] ;
[0076] in, Indicated in In the iteration The location of the new bitter fish generated, Indicated in In the iteration The location of the new bitter fish generated, Indicates the update radius for spawning new bitter fish.
[0077] Further, the process of hunting bitter fish in step S25 is expressed as follows:
[0078] ;
[0079] in, Indicated in The first iteration The location of the hunted bitter fish, Indicated in The first iteration The probability of a hunted bitter fish, Indicates the total number of bitter fish.
[0080] Furthermore, in step S2, the number of neurons in the input layer of the probabilistic neural network model is equal to the dimension of the signal data, and the model layer calculates the output of each risk type according to the central Euclidean distance between the input signal data and the feature data set:
[0081] fA ( X ) = 1 ( 2 π ) d / 2 δ d 1 m 1 ∑ exp [ − ( X − X ai ) T ( X − X ai ) 2 δ 2 ] ;
[0082] Where: represents the probability density function, The feature vector representing the input signal data, represents the dimension of the training sample, Represents the total number of samples in the feature dataset, For the The central vector of each risk type, represents the smoothing factor, represents transpose, represents the activation function;
[0083] The summation layer takes a weighted average of the outputs of neurons of the same category in the pattern layer and calculates the Probability density of risk class , expressed as:
[0084] ;
[0085] In the formula, To indicate the The total number of samples of the risk type, Indicates Samples of risk-type classes;
[0086] The output layer adopts the Bayesian classification criterion and selects the risk type with the maximum posterior probability density for output.
[0087] Beneficial effects of the present invention: The present invention extracts features from signal data of partial discharge of transformers and uses the improved bitter fish algorithm to optimize the probabilistic neural network model to achieve online detection of the risk type of partial discharge signals of transformers. Compared with existing solutions, the present invention combines the two signals for processing, which reduces misjudgments caused by accidental situations to a certain extent. In addition, the present invention optimizes the probabilistic neural network by using the improved bitter fish algorithm, which greatly improves the accuracy of detection and provides a new solution to the problem of the difficulty in selecting the smoothing factor of the probabilistic neural network. BRIEF DESCRIPTION OF THE DRAWINGS
[0088] Figure 1 The flowchart of a transformer partial discharge detection method based on a neural network is shown in FIG. DETAILED DESCRIPTION
[0089] The following describes the embodiments of the present invention through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention.
[0090] Reference Figure 1 A transformer partial discharge detection method based on a neural network, in one example, comprises the following steps:
[0091] A transformer partial discharge detection method based on a neural network comprises the following steps:
[0092] Step S1: Collect signal data of partial discharge during the historical operation of the transformer, mark the signal data according to the risk type, construct a data set, and pre-process the data set to obtain a feature data set;
[0093] Step S2: construct a probabilistic neural network model, import the feature data set in step S1 into the probabilistic neural network model for training, and during the training process, optimize the smoothing factor of the probabilistic neural network model by using the bitter fish algorithm to obtain the optimal smoothing factor;
[0094] Step S3: collecting signal data of partial discharge when the transformer is running in real time and preprocessing it, importing the preprocessed signal data into the trained probabilistic neural network model for detection, and outputting the risk type of partial discharge of the transformer in real time;
[0095] In step S2, the process of optimizing the smoothing factor of the probabilistic neural network model by the bitter fish algorithm is as follows:
[0096] Step S21: Initialize the bitter fish population in the search space;
[0097] Step S22: searching for oysters to find the optimal solution space of the smoothing factor;
[0098] Step S23: discarding oysters to retain the found optimal solution space and further find a new optimal solution space;
[0099] Step S24: generating a new bitter fish for further searching for a better solution in the optimal solution space;
[0100] Step S25: hunting bitter fish to eliminate the difference;
[0101] Step S26: Determine whether the maximum number of iterations has been reached. When the maximum number of iterations has been reached, calculate the fitness of the bittern at the current number of iterations, and select the bittern with the maximum fitness as the optimal smoothing factor. When the maximum number of iterations has not been reached, continue with steps S22 to S26.
[0102] In one example, the signal data in step S1 includes an ultrasonic signal and an ultra-high frequency electromagnetic wave signal;
[0103] The data set in step S1 is represented as:
[0104] P = [ sc 1 , 1 ⋯ sc 1 , n ⋯ sd 1 , 1 ⋯ sd 1 , n ⋮ ⋱ ⋮ ⋮ ⋮ ⋱ ⋮ scm , 1 ⋯ scm , n ⋯ sdm , 1 ⋯ sdm , n ] ;
[0105] In the formula, represents a dataset, Represents the ultrasonic signal of the first sampling point in the first sampling interval; represents the ultrasonic signal of the first sampling point in the mth sampling interval, Represents the ultrasonic signal of the nth sampling point in the first sampling interval; represents the ultrasonic signal of the nth sampling point in the mth sampling interval, Represents the UHF electromagnetic wave signal of the first sampling point in the first sampling interval, Represents the ultra-high frequency electromagnetic wave signal of the first sampling point in the mth sampling interval, Represents the ultra-high frequency electromagnetic wave signal of the nth sampling point in the first sampling interval, Represents the ultra-high frequency electromagnetic wave signal of the nth sampling point in the mth sampling interval;
[0106] The step S1 of preprocessing the data set to obtain the feature data set is as follows:
[0107] Step S11: normalizing the data set, thereby adjusting the value range of the data set to the interval between 0 and 1, and obtaining a normalized data set;
[0108] Specifically, the normalization process is performed by applying the maximum and minimum methods, which are as follows:
[0109] ;
[0110] In the formula, represents the normalized value, Represents signal data, Indicates the maximum value of signal data. Indicates the minimum value of signal data;
[0111] Step S12: Calculate the difference between each signal data of the normalized data set and the average value of the signal data of the normalized data set, compare the absolute value of the obtained difference with the preset abnormal threshold, and when the absolute value of the difference is equal to or greater than the abnormal threshold, retain the signal data; when the absolute value of the difference is less than the abnormal threshold, use the average value of the signal data of the normalized data set to replace the signal data, thereby forming a feature data set.
[0112] In one example, the risk types in step S1 include: not discharged, low risk, medium risk, and high risk.
[0113] In one example, in step S2, the probabilistic neural network model includes an input layer, a pattern layer, a summation layer, and an output layer;
[0114] The input layer receives the real-time signal data of partial discharge during transformer operation in step S1 as input and imports it into the pattern layer. The pattern layer calculates the degree of matching between the real-time signal data and the preset feature data set and outputs it as a probability density. The summation layer accumulates the probability density of each risk type of partial discharge and calculates the comprehensive probability according to the output of the pattern layer. The output layer outputs the predicted risk type based on the comprehensive probability calculated by the summation layer.
[0115] Among them, the smoothing factor is used to adjust the probability density of the pattern layer output and thus adjust the sensitivity of the probabilistic neural network model classification.
[0116] In one example, step S21 is specifically: creating an initial population, expressed as:
[0117] F 0 = [ F 1 1 F 1 2 ⋯ F 1 D F 2 1 F 2 2 ⋯ F 2 D ⋮ ⋮ ⋮ ⋮ F n 1 F n 2 ⋯ F n D ] ;
[0118] In the formula, The matrix representing the initial population of bitter fish, represents a matrix element in the matrix, It means a bitter fish. and Respectively represent the row and column position of each matrix element in the matrix, that is, the position of the bitter fish;
[0119] To simulate the reproduction stage of bittern, male bittern seeks oysters to attract female bittern. Oysters are used to simulate the optimal area for bittern, and each bittern is evaluated, which is expressed as:
[0120] Fitness = [ f ( F 1 1 ) f ( F 1 2 ) ⋯ f ( F 1 D ) f ( F 2 1 ) f ( F 2 2 ) ⋯ f ( F 2 D ) ⋮ ⋮ ⋮ ⋮ f ( F n 1 ) f ( F n 2 ) ⋯ f ( F n D ) ] ;
[0121] in, represents the fitness value matrix, represents the fitness function;
[0122] The fitness function is expressed as:
[0123] ;
[0124] In the formula, represents fitness, Represents the output of the probabilistic neural network model The predicted risk type, Indicates The risk type is marked. is the number of predicted values.
[0125] Further, in one example, the oyster search in step S22 is specifically as follows:
[0126] Step S221: The bitter fish will target the oyster and move closer to or away from it, and the target oyster will not be noticed and captured by another male bitter fish. The oyster possession state is as follows:
[0127] ;
[0128] In the formula, represents the number of iterations, Indicated in In the iteration The location of the bittern that caught oysters, Indicated in In the iteration The location of the bittern that caught oysters, The best oysters that bitter fish seek, represents one of the randomly selected oyster populations, and Each represents a random number between 0 and 1. Indicates the current location of the oyster. indicates the rate at which bitterlings escape from or approach oysters;
[0129] Step S222: Generate a random number that conforms to the Cauchy distribution, so as to increase the rate at which the bittern escapes from or approaches the oyster. The probability density function of the Cauchy distribution is It is expressed as:
[0130] ;
[0131] in, Represents positional parameters, represents the scale parameter, Input representing the probability density function of the Cauchy distribution;
[0132] The random number generated by Cauchy distribution is used as the random factor, which is expressed as
[0133] ;
[0134] In the formula, represents the random factor, Represents a random number uniformly distributed between 0 and 1;
[0135] The random factor Multiply by the current speed of the bitter fish , get the updated speed , expressed as:
[0136] ;
[0137] Step S223: During the simulation, the rate at which the bitter fish escape from or approach the oyster will decrease as the number of iterations increases. At this time, the global search becomes a local search. The expression of the rate at which the bitter fish escapes from or approaches the oyster in each iteration is as follows:
[0138] ;
[0139] in, represents the rate at which bitterlings escape from or approach oysters in the first iteration, and The values of represent the current number of iterations and the maximum number of iterations respectively. represents the rate at which the bitter fish escape from or approach the oyster in the current iteration. Represents a random function, used to generate a random sequence;
[0140] The process of generating random sequences by random functions is expressed as:
[0141] ;
[0142] Presets Equal to 1, the formula is:
[0143] ;
[0144] Step S224: As the number of iterations increases, the oyster occupancy state formula in step S221 In the condition, reduce The value of to reduce the spatial duplication around the optimal solution is as follows:
[0145] ;
[0146] in, Indicates the area where the oyster currently exists. represents the adjustment parameter; the smaller the probability of deviation when performing two types of searches in the oyster occupancy state expression, represents a random real function.
[0147] Step S23: Abandoning oysters. In the simulation of the behavior of bitter fish abandoning oysters, one of the bitter fish moves the oysters to another location and guards the oysters to prevent other bitter fish from approaching the oysters. The following formula represents the bitter fish that did not catch the oysters:
[0148] ;
[0149] in, Indicated in In the iteration The location of the bittern that caught no oysters, Indicated in In the iteration The location of the bittern that caught no oysters, represents the average position of the bitter fish population, and All represent preset adjustment parameters;
[0150] The average position of bitterlings that did not catch oysters is expressed as follows:
[0151] ;
[0152] in, represents the number of times the optimal solution is evaluated by the algorithm, .
[0153] The average position of bitterlings that failed to capture oysters is updated by random numbers to continue searching for oysters. The formula is as follows:
[0154] ;
[0155] In the formula, is the updated position of the average position of bitterlings that did not catch oysters, represents random numbers from a standard normal distribution, represents the strength coefficient;
[0156] examine Whether it exceeds the boundary of the search space, Adjustments are made to ensure that they are within the valid search space, as shown below:
[0157] ;
[0158] in, Indicates the adjusted position of the bitter fish. and denote the minimum and maximum bounds of the search space respectively.
[0159] Further, in one example, step S24 of producing new bitter fish is specifically as follows: after catching oysters, the male bitter fish can attract the female bitter fish and lay eggs in the oysters, and the male can fertilize the eggs, and produce new bitter fish around the current bitter fish position, as shown in the following formula:
[0160] ;
[0161] in, Indicated in In the iteration The location of the new bitter fish generated, Indicated in In the iteration The location of the new bitter fish generated, Indicates the update radius for spawning new bitter fish.
[0162] Further, in one example, step S25 of hunting bittern is specifically as follows: the number of bittern that can be accommodated by oysters is limited, and weak bittern will be hunted by oysters or predators. The process is expressed as follows:
[0163] ;
[0164] in, Indicated in The first iteration The location of the hunted bitter fish, Indicated in The first iteration The probability of a hunted bitter fish, Indicates the total number of bitter fish.
[0165] Further, in one example, the number of neurons in the input layer of the probabilistic neural network model in step S2 is equal to the dimension of the signal data, and the model layer calculates the output of each risk type according to the central Euclidean distance between the input signal data and the feature data set:
[0166] fA ( X ) = 1 ( 2 π ) d / 2 δ d 1 m 1 ∑ exp [ − ( X − X ai ) T ( X − X ai ) 2 δ 2 ] ;
[0167] Where: represents the probability density function, The feature vector representing the input signal data, represents the dimension of the training sample, Represents the total number of samples in the feature dataset, For the The central vector of each risk type, represents the smoothing factor, represents transpose, represents the activation function;
[0168] The summation layer takes a weighted average of the outputs of neurons of the same category in the pattern layer and calculates the Probability density of risk class , expressed as:
[0169] ;
[0170] In the formula, To indicate the The total number of samples of the risk type, Indicates Samples of risk-type classes;
[0171] The output layer adopts the Bayesian classification criterion and selects the risk type with the maximum posterior probability density for output.
[0172] The present invention is described in detail above in conjunction with the embodiments of the accompanying drawings. A person skilled in the art can make various variations of the present invention according to the above description. Therefore, some details in the embodiments should not be construed as limiting the present invention, and the present invention shall be protected by the scope defined by the attached claims.
Claims
1. A transformer partial discharge detection method based on neural network, characterized in that: The following steps are involved: Step S1: Collect signal data of partial discharge during the historical operation of the transformer, mark the signal data according to the risk type, construct a data set, and pre-process the data set to obtain a feature data set; Step S2: construct a probabilistic neural network model, import the feature data set in step S1 into the probabilistic neural network model for training, and during the training process, optimize the smoothing factor of the probabilistic neural network model by using the bitter fish algorithm to obtain the optimal smoothing factor; Step S3: collecting signal data of partial discharge when the transformer is running in real time and preprocessing it, importing the preprocessed signal data into the trained probabilistic neural network model for detection, and outputting the risk type of partial discharge of the transformer in real time; In step S2, the process of optimizing the smoothing factor of the probabilistic neural network model by the bitter fish algorithm is as follows: Step S21: Initialize the bitter fish population in the search space; Step S22: Searching for oysters; Step S23: discarding oysters; Step S24: generating new bitter fish; Step S25: hunting bitter fish; Step S26: Determine whether the maximum number of iterations has been reached. If the maximum number of iterations has been reached, calculate the fitness of the bittern at the current number of iterations, and select the bittern with the maximum fitness as the optimal smoothing factor. If the maximum number of iterations has not been reached, continue with steps S22 to S26; In step S2, the probabilistic neural network model includes an input layer, a pattern layer, a summation layer and an output layer; The input layer receives the real-time signal data of partial discharge during transformer operation in step S1 as input and imports it into the pattern layer. The pattern layer calculates the degree of matching between the real-time signal data and the preset feature data set and outputs it as a probability density. The summation layer accumulates the probability density of each risk type of partial discharge and calculates the comprehensive probability according to the output of the pattern layer. The output layer outputs the predicted risk type based on the comprehensive probability calculated by the summation layer. Among them, the smoothing factor adjusts the probability density of the pattern layer output to adjust the sensitivity of the probabilistic neural network model classification; Step S21 is specifically: creating an initial population, expressed as: ; In the formula, The matrix representing the initial population of bitter fish, represents a matrix element in the matrix, It means a bitter fish. and Respectively represent the row and column position of each matrix element in the matrix, that is, the position of the bitter fish; Each bittern in the initial population is evaluated as: ; in, represents the fitness value matrix, represents the fitness function; The fitness function is expressed as: ; In the formula, represents fitness, Represents the output of the probabilistic neural network model The predicted risk type, Indicates The risk type is marked. is the number of predicted values; The specific steps of searching for oysters in step S22 are as follows: Step S221: The occupancy status of the oyster is as follows: ; In the formula, represents the number of iterations, Indicated in In the iteration The location of the bittern that caught oysters, Indicated in In the iteration The location of the bittern that caught oysters, The best oysters for bitter fish. represents one of the randomly selected oyster populations, and Both represent a random number between 0 and 1. Indicates the current location of the oyster. indicates the rate at which bitterlings escape from or approach oysters; Step S222: Generate a random number that conforms to the Cauchy distribution, so as to increase the rate at which the bittern escapes from or approaches the oyster. The probability density function of the Cauchy distribution is It is expressed as: ; in, Represents positional parameters, represents the scale parameter, Input representing the probability density function of the Cauchy distribution; The random number generated by Cauchy distribution is used as the random factor, which is expressed as ; In the formula, represents the random factor, Represents a random number uniformly distributed between 0 and 1; The random factor Multiply by the current speed of the bitter fish , get the updated speed , expressed as: ; Step S223: During the simulation, the rate at which the bitter fish escape from or approach the oyster will decrease as the number of iterations increases. At this time, the global search becomes a local search. The expression of the rate at which the bitter fish escapes from or approaches the oyster in each iteration is as follows: ; in, represents the rate at which bitterlings escape from or approach oysters in the first iteration, represents the maximum number of iterations, Indicates that bitter fish is in The rate of escaping or approaching oysters in iterations, Represents a random function, used to generate a random sequence; The process of generating random sequences by random functions is expressed as: ; Presets Equal to 1, the formula is: ; Step S224: As the number of iterations increases, the oyster occupancy state formula in step S221 In the condition, reduce The value of to reduce the spatial duplication around the optimal solution is as follows: ; in, Indicates the area where the oyster currently exists. represents the adjustment parameter; the smaller the probability of deviation when performing two types of searches in the oyster occupancy state expression, represents a random real function; The specific steps of discarding oysters in step S23 are as follows: Calculate the position of bitterlings that have not caught oysters, expressed as: ; in, Indicated in In the iteration The location of the bittern that caught no oysters, Indicated in In the iteration The location of the bittern that caught no oysters, represents the average position of the bitter fish population, and All represent preset adjustment parameters; The average position of bitterlings that did not catch oysters is expressed as follows: ; in, represents the number of times the optimal solution is evaluated by the algorithm, ; The average position of bitterlings that failed to capture oysters is updated by random numbers to continue searching for oysters. The formula is as follows: ; In the formula, is the updated position of the average position of bitterlings that did not catch oysters, represents random numbers from a standard normal distribution, represents the strength coefficient; examine Whether it exceeds the boundary of the search space, Adjustments are made to ensure that they are within the valid search space, as shown below: ; in, Indicates the adjusted position of the bitter fish. and Respectively represent the minimum and maximum boundaries of the search space; The production of new bitter fish in step S24 is specifically as follows: Generate new bitter fish around the current bitter fish position, as follows: ; in, Indicated in In the iteration The location of the new bitter fish generated, Indicated in In the iteration The location of the new bitter fish generated, Indicates the update radius for generating new bitter fish; The process of hunting bitter fish in step S25 is expressed as follows: ; in, Indicated in The iteration The location of the hunted bitter fish, Indicated in The first iteration The probability of a hunted bitter fish, Indicates the total number of bitter fish.
2. A transformer partial discharge detection method based on neural network as claimed in claim 1, characterized in that: The signal data in step S1 includes ultrasonic signals and ultra-high frequency electromagnetic wave signals; The data set in step S1 is represented as: ; In the formula, represents a dataset, Represents the ultrasonic signal of the first sampling point in the first sampling interval; represents the ultrasonic signal of the first sampling point in the mth sampling interval, Represents the ultrasonic signal of the nth sampling point in the first sampling interval; represents the ultrasonic signal of the nth sampling point in the mth sampling interval, Represents the UHF electromagnetic wave signal of the first sampling point in the first sampling interval, Represents the ultra-high frequency electromagnetic wave signal of the first sampling point in the mth sampling interval, Represents the ultra-high frequency electromagnetic wave signal of the nth sampling point in the first sampling interval, Represents the ultra-high frequency electromagnetic wave signal of the nth sampling point in the mth sampling interval; The step S1 of preprocessing the data set to obtain the feature data set is as follows: Step S11: normalizing the data set, thereby adjusting the value range of the data set to the interval between 0 and 1, and obtaining a normalized data set; Step S12: Calculate the difference between each signal data of the normalized data set and the average value of the signal data of the normalized data set, compare the absolute value of the obtained difference with the preset abnormal threshold, and when the absolute value of the difference is equal to or greater than the abnormal threshold, retain the signal data; when the absolute value of the difference is less than the abnormal threshold, use the average value of the signal data of the normalized data set to replace the signal data, thereby forming a feature data set.
3. A transformer partial discharge detection method based on neural network as claimed in claim 1, characterized in that: The risk types in step S1 include: not discharged, low risk, medium risk and high risk.
4. A transformer partial discharge detection method based on neural network as claimed in claim 1, characterized in that: In step S2, the number of neurons in the input layer of the probabilistic neural network model is equal to the dimension of the signal data. The model layer calculates the output of each risk type based on the central Euclidean distance between the input signal data and the feature data set: ; Where: represents the probability density function, The feature vector representing the input signal data, represents the dimension of the training sample, Represents the total number of samples in the feature dataset, For the The central vector of each risk type, represents the smoothing factor, represents transpose, represents the activation function; The summation layer takes a weighted average of the outputs of neurons of the same category in the pattern layer and calculates the Probability density of risk class , expressed as: ; In the formula, To indicate the The total number of samples of the risk type, Indicates Samples of risk-type classes; The output layer adopts the Bayesian classification criterion and selects the risk type with the maximum posterior probability density for output.
