A high frequency partial discharge type identification method and system based on gradient descent method

Abstract

The present invention discloses a method and system for identifying high-frequency partial discharge types based on a gradient descent method, the method comprising: reading data of determined partial discharge types, solving each coefficient of a fitting polynomial based on a gradient descent method, and establishing a database of characteristic data and discharge types; reading data of partial discharge types of the equipment under test, calculating the total discharge amplitude based on the polynomial, identifying and classifying the partial discharge types of errors; defining new types of partial discharge, solving each coefficient of the polynomial based on self-learning, and updating the newly added discharge type database; the system comprises a database module, an identification and classification module, and an update module. The present invention realizes online monitoring of the current operating status of power equipment, and realizes early warning and prediction of possible faults according to historical development trends, taking precautions before they occur, avoiding further evolution into power grid accidents, and improving the level of automation.

Description

Technical Field

[0001] The present invention relates to the field of high-frequency partial discharge identification, and in particular to a high-frequency partial discharge type identification method and system based on a gradient descent method. Background Art

[0002] High-frequency partial discharge is a common type of partial discharge in electrical equipment such as transformers, GIS (Gas Insulated Switchgear), GIL (Gas Insulated Metal Enclosed Transmission Line), circuit breakers, and switch cabinets. High-frequency partial discharge has the following characteristics: (1) Strong intermittency. Partial discharge is sometimes very intermittent, and it may be difficult for live detection personnel to capture partial discharge signals. (2) Small discharge volume. Many partial discharges are caused by the destruction of a small part of the insulation. Therefore, the discharge volume is small, and it is difficult to capture the signal through traditional transformers. (3) Wide frequency band. The frequency range of high-frequency partial discharge detection is usually between 3MHz and 30MHz. The frequency is high and the frequency band is wide. However, due to the fixed filter parameters, it is difficult to ensure filtering in the wide range of 3MHz to 30MHz through hardware or software filters. Some high-frequency signals may be lost, and even many signals cannot be preserved by Fourier filtering. In order to realize online monitoring of the current operating status of power equipment and to achieve early warning and prediction of possible faults based on historical development trends, an effective high-frequency partial discharge type identification method is needed. Summary of the invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a method and system for identifying high-frequency partial discharge types based on the gradient descent method.

[0004] Technical solution: The high-frequency partial discharge type identification method based on the gradient descent method described in the present invention comprises the following steps:

[0005] (1) Reading the data of the determined partial discharge type, solving the coefficients of the fitting polynomial based on the gradient descent method, and establishing a database of characteristic data and discharge types;

[0006] (2) Read the data of the partial discharge type of the equipment under test, calculate the total discharge amplitude based on the polynomial, and identify and classify the partial discharge type of the error;

[0007] (3) Define new types of partial discharge, solve the coefficients of the polynomial based on self-learning, and update the database of new discharge types.

[0008] Furthermore, the step (1) comprises:

[0009] (1.1) Label different groups of data according to the type of partial discharge;

[0010] (1.2) Read the data of partial discharge times and discharge amount, and perform normalization processing on the per unit value according to the total amplitude of discharge amount and the total number of discharge amounts;

[0011] (1.3) Fitting polynomials to set initial values ​​of coefficients;

[0012] (1.4) Calculate the error value;

[0013] (1.5) Complete the adjustment and correction of each polynomial coefficient based on the gradient descent method;

[0014] (1.6) Determine whether the gradient descent distance of each coefficient is less than the allowable error judgment;

[0015] (1.7) Determine the coefficients of the polynomials of discharge capacity and discharge times of the discharge type;

[0016] (1.8) Establish the mapping relationship between y and θ corresponding to different local discharge types.

[0017] Furthermore, the step (1.3) of fitting the polynomial to set the initial values ​​of each coefficient includes:

[0018] The partial discharge data from 0 to 360° are divided into n groups according to the set angle step, which is expressed as:

[0019] Data x l , x2, ..., x n ,

[0020] in p is the angle step, n is the number of data in one cycle; y is the total discharge amplitude,

[0021] From the actual values ​​read from the file, a polynomial equation is established for the process prediction value y′ and x:

[0022] y′=θ T X=a0+a1x1+……+a n x n ,

[0023] where θ=[a0,a1,……,a n ],θ T is the transposed matrix of θ, θ is the discharge coefficient at different angles, X = [x0, x1, ..., x n ], X is the number of discharges at different angles, x0 is set to 1, and the discharge coefficients at different angles are a1, a2, ..., a n The initial value is set to l.

[0024] Furthermore, the step (1.4) of calculating the error value includes:

[0025] Substitute the x and a values ​​into the polynomial y′=θ T X=a0+a1x1+……+a n x n The process prediction value y′ is obtained, and the prediction error is calculated based on the actual value y of the total discharge amplitude and the process prediction value.

[0026] Through m groups of X calculations, we can get m groups of y', and through m groups of y' and y data, we can calculate the prediction error. Calculate the error value of the loss function.

[0027] Furthermore, the adjustment and correction of the coefficients of the polynomial in step (1.5) includes:

[0028] The gradient of the loss function is calculated according to the gradient descent method, that is,

[0029] Iteratively update a according to the calculation results i Each coefficient, that is Among them, L is the step size of the gradient descent method. The step size is multiplied by the gradient of the loss function to get the distance the current position has dropped.

[0030] Furthermore, the step (1.6) of determining the gradient descent distance of each coefficient includes:

[0031] When there is a i The gradient descent distance is greater than or equal to ε, and the iterative calculation continues Until all a i The gradient descent distance is less than ε at the end of a i End of iterative calculation.

[0032] Furthermore, the coefficients of the polynomial of the discharge amount and the discharge number of the discharge type determined in step (1.7) include:

[0033] Determine the polynomial y′=θ T X=a0+a1x1+……+a n x n The θ value of the discharge coefficient at different angles, that is, [a0, a1, ..., a n ], it is approximately considered that y≈y′, that is, polynomial expressions for different discharge types are established:

[0034] y=θ T X=a0+a1x1+……+a n x n .

[0035] Furthermore, the step (2) comprises:

[0036] (2.1) Read the partial discharge times and discharge amount data of the equipment under test;

[0037] (2.2) According to the different polynomial coefficients of different partial discharge types existing in the database, Q(x) = a0+a1x1+…+a n x n Different values ​​of the total discharge amplitude q are calculated;

[0038] (2.3) Calculate the error between the actual discharge amount and the predicted value according to the q value of different discharge types, that is, calculate the error value of yq, and analyze the error between the actual discharge amount and the predicted value;

[0039] (2.4) Identification and classification of partial discharge types, the error value is set within the allowable ε q The range is the type of partial discharge, where ε q Enter or set a fixed value parameter.

[0040] Furthermore, the step (3) comprises:

[0041] (3.1) New partial discharge types are labeled, and all error values ​​are greater than the set ε q , labeling new types of partial discharges through manual intervention;

[0042] (3.2) Solving the coefficients of the polynomial based on self-learning;

[0043] (3.3) Determine the coefficients of the polynomial between the discharge amount and the number of discharges of the discharge type, i.e., the value of θ of the discharge coefficient;

[0044] (3.4) Update the current new type to the existing database or discharge type model.

[0045] The high-frequency partial discharge type identification system based on the gradient descent method of the present invention comprises:

[0046] A database module is used to read the data of the determined partial discharge type, solve the coefficients of the fitting polynomial based on the gradient descent method, and establish a database of characteristic data and discharge types;

[0047] The identification and classification module is used to read the data of the partial discharge type of the equipment under test, calculate the total discharge amplitude based on the polynomial, and identify and classify the partial discharge type of the error;

[0048] The update module is used to define new types of partial discharges, solve the coefficients of the polynomial based on self-learning, and update the database of newly added discharge types.

[0049] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the gradient descent method is used to adjust the polynomial coefficient parameters, polynomial curve fitting, polynomial weight result calculation and error analysis to complete the classification and identification of high-frequency partial discharge types; it is suitable for online monitoring of high-frequency partial discharge and partial discharge type identification of transformers, GIS, GIL, circuit breakers, and switch cabinets, and is also suitable for the self-learning high-frequency partial discharge database update of the above-mentioned equipment; before a minor or serious fault occurs, the partial discharge collection and discharge type identification are identified through high-frequency current, thereby realizing online monitoring of the current operating status of the power equipment, and realizing early warning and prediction of possible faults based on historical development trends, taking precautions before they occur, avoiding further evolution into power grid accidents, and improving the level of automation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Flowchart for the establishment of a database of partial discharge characteristics and types;

[0051] Figure 2 Flow chart for identification and classification of partial discharge types;

[0052] Figure 3 Flowchart for self-learning and database update of new types of partial discharge. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0054] The high-frequency partial discharge type identification method based on the gradient descent method described in the present invention includes the establishment of partial discharge characteristics and type database based on the gradient descent method; identification and classification of partial discharge types based on polynomial calculation and error; self-learning of new partial discharge types and database update.

[0055] (1) Establishment of partial discharge characteristics and type database based on gradient descent method

[0056] like Figure 1 As shown, the establishment of a database of partial discharge characteristics and types based on the gradient descent method includes: reading data of determined partial discharge types, solving coefficients of fitting polynomials based on the gradient descent method, and establishing a database of characteristic data and discharge types, which specifically includes the following steps:

[0057] (1.1) Labeling of partial discharge types: Several groups of partial discharge types of existing data are obtained offline. The same discharge type is classified into the same group. Each group of the same discharge type contains several data. Each different group of data is labeled with a partial discharge type according to the partial discharge type.

[0058] (1.2) Reading of partial discharge times and discharge amount data: reading the partial discharge times and discharge amount amplitude information of each group of each data according to the storage format, and calculating the total discharge amplitude or total discharge times and other information, and performing normalization processing on the per unit value according to the total discharge amplitude and the total discharge times.

[0059] (1.3) The fitting polynomial sets the initial values ​​of each coefficient, reads the data of partial discharge times and discharge amount, and performs normalization processing on the per-unit value. Then, the partial discharge times data from 0 to 360° is divided into n groups according to the set angle step, i.e., data x1, x2, ..., x n ,in p is the angle step, n is the number of data in one cycle; y is the total discharge amplitude, the real value read from the file, and the polynomial equation of the process prediction value y′ and x is established, that is, y′=θ T X=a0+a1x1+……+a n x n , where θ=[a0,a1,……,a n ],θ T is the transposed matrix of θ, θ is the discharge coefficient at different angles, X = [x0, x1, ..., x n ], X is the number of discharges at different angles. For the convenience of calculation, x0 is set to 1, and the discharge coefficients at different angles are a1, a2, ..., a n The initial value is set to 1.

[0060] (1.4) Error calculation, substitute the values ​​of x and a into the polynomial y′=θ T x=a0+a1x1+……+a n x n The process prediction value y′ is obtained, and the prediction error is calculated according to the actual value y of the total discharge amplitude and the process prediction value. In practical applications, m groups of y′ are obtained by calculating m groups of X, and the prediction error is calculated by m groups of y′ and y data, that is, by Calculate the error value of the loss function.

[0061] (1.5) Based on the gradient descent method, the adjustment and correction of each coefficient of the polynomial are completed, and the gradient of the loss function is calculated according to the gradient descent method, that is, Then iteratively update a according to the calculation results i Each coefficient, that is Where L is the step size of the gradient descent method. Multiply the step size by the gradient of the loss function to get the distance the current position has dropped.

[0062] (1.6) Whether the gradient descent distance of each coefficient is less than the allowable error judgment process, when there is a i The gradient descent distance is greater than or equal to ε, and the iterative calculation continues Until all ai The gradient descent distance is less than ε at the end of a i End of iterative calculation.

[0063] (1.7) The coefficients of the polynomial of the discharge capacity and discharge times of the discharge type are determined. The calculation result [a0, a1, ..., a n ] is the final result of θ, which determines the polynomial y′=θ T X=a0+a1x1+……+a n x n The θ value of the discharge coefficient at different angles, that is, [a0, a1, ..., a n ], then we can approximately consider y≈y′, that is, y=θ for different discharge types is established. T X=a0+a1x1+……+a n x n The polynomial expression of .

[0064] (1.8) A database of characteristic data and discharge types is established, and a mapping relationship between y and θ corresponding to different local discharge types is established.

[0065] The entire process of establishing the partial discharge characteristics and type database based on the gradient descent method is completed.

[0066] (2) Identification and classification of partial discharge types based on polynomial calculation and error

[0067] like Figure 2 As shown, the identification and classification of partial discharge types based on polynomial calculation and error include: data reading of partial discharge type of the equipment under test, calculation of the total discharge amplitude based on polynomials, identification and classification of partial discharge types based on errors, specifically including the following steps:

[0068] (2.1) Reading of the number of partial discharges and the amount of partial discharges of the equipment under test. For an online partial discharge monitoring device, the number of partial discharges x and the total amount of partial discharge amplitude y in the same time period are read online through the buffer, or the number of partial discharges x and the total amount of partial discharge amplitude y of the equipment under test from the PRPD and PRPS files are read online; for an offline partial discharge analysis device or offline analysis system, the number of partial discharges x and the total amount of partial discharge amplitude y of the equipment under test are read offline from the PRPD and PRPS files.

[0069] (2.2) The total discharge amplitude prediction calculation is based on the different polynomial coefficients of different partial discharge types in the database Q(x) = a0+a1x1+…+a n x n Different total discharge amplitude q values ​​are calculated.

[0070] (2.3) The error analysis between the actual discharge amount and the predicted value is to calculate the error between the actual total discharge amount and the predicted value according to the q value of different discharge types, that is, to calculate the error value of yq.

[0071] (2.4) Identification and classification of partial discharge types, the error value is set within the allowable ε q The range is the type of partial discharge, where ε q Can be entered or set as a fixed value / parameter.

[0072] The whole process of identifying and classifying partial discharge types based on polynomial calculation and error is completed.

[0073] (3) Self-learning of new types of partial discharge and database update

[0074] like Figure 3 As shown, the self-learning and database update of the new type of partial discharge includes: the definition of the new type of partial discharge, the solution of each coefficient of the self-learning polynomial, and the update of the database of the newly added discharge type, which specifically includes the following steps:

[0075] (3.1) New partial discharge types are labeled, and all error values ​​are greater than the set ε q That is, this type of partial discharge does not belong to any known type in the existing database. The new type of partial discharge is labeled through manual intervention. At the same time, several pieces of data of this type need to be saved. The number of discharges and the discharge amplitude of each piece need to be collected and calculated in the same time period.

[0076] (3.2) Self-learning solution of the polynomial coefficients is the same as steps (1.3) to (1.6) of module 1 "Establishment of partial discharge characteristics and type database based on gradient descent method".

[0077] (3.3) The coefficients of the polynomial of discharge type and discharge number are determined, and the value of θ of the discharge coefficient of the polynomial is determined, that is, [a0, a1, ..., a n ] value.

[0078] (3.4) Added discharge type database update, updating the current new type to the existing database or discharge type model.

[0079] The whole process of self-learning of the new type of partial discharge and database update is completed.

[0080] The high-frequency partial discharge type identification system based on the gradient descent method of the present invention comprises:

[0081] A database module is used to read the data of the determined partial discharge type, solve the coefficients of the fitting polynomial based on the gradient descent method, and establish a database of characteristic data and discharge types;

[0082] The identification and classification module is used to read the data of the partial discharge type of the equipment under test, calculate the total discharge amplitude based on the polynomial, and identify and classify the partial discharge type of the error;

[0083] The update module is used to define new types of partial discharges, solve the coefficients of the polynomial based on self-learning, and update the database of newly added discharge types.

[0084] The present invention aims at the characteristics of strong intermittency, small discharge amount and wide frequency band of partial discharge. For the strong intermittency, the present invention adopts online monitoring to solve it, for the small discharge amount, the present invention adopts high-frequency current transformer collection, and for the wide frequency band, the present invention adopts polynomial curve fitting based on gradient descent method. In summary, the present invention completes the classification and identification of high-frequency partial discharge types by acquiring the data of partial discharge times and discharge amplitude, adjusting the polynomial coefficient parameters, polynomial curve fitting, polynomial weight result calculation and error analysis by using the gradient descent method.

[0085] This method is applicable to the online monitoring of partial discharge and identification of partial discharge types of transformers, GIS, GIL, circuit breakers and switch cabinets, and is also applicable to the self-learning update of the high-frequency partial discharge database of the above-mentioned equipment. Before a minor or serious fault occurs, the partial discharge collection and discharge type identification are identified through high-frequency current, so as to realize the online monitoring of the current operating status of the power equipment, and realize the early warning and prediction of the possible faults according to the historical development trend, so as to prevent it from further evolving into a power grid accident, thereby improving the automation level and having the characteristics of high automation level.

Claims

1. A high-frequency partial discharge type identification method based on gradient descent method, characterized in that: The steps include: (1) Reading the data of the determined partial discharge type, solving the coefficients of the fitting polynomial based on the gradient descent method, and establishing a database of characteristic data and discharge types; (2) Read the data of the partial discharge type of the equipment under test, calculate the total discharge amplitude based on the polynomial, and identify and classify the partial discharge type of the error; (3) Define new types of partial discharge, solve the coefficients of the polynomial based on self-learning, and update the database of new discharge types. The step (1) comprises: (1.1) Label different groups of data according to the type of partial discharge; (1.2) Read the data of partial discharge times and discharge amount, and perform normalization processing on the per unit value according to the total amplitude of discharge amount and the total number of discharge amounts; (1.3) Fitting polynomials to set initial values ​​of coefficients; (1.4) Calculate the error value; (1.5) Complete the adjustment and correction of each polynomial coefficient based on the gradient descent method; (1.6) Determine whether the gradient descent distance of each coefficient is less than the allowable error judgment; (1.7) Determine the coefficients of the polynomials of discharge capacity and discharge times of the discharge type; (1.8) Establish the mapping relationship between y and θ corresponding to different partial discharge types, The step (1.3) of fitting the polynomial and setting the initial values ​​of the coefficients includes: The partial discharge data from 0 to 360° are divided into n groups according to the set angle step, which is expressed as: Data x1, x2, ..., xn, in p is the angle step, n is the number of data in one cycle; y is the total discharge amplitude, From the actual values ​​read from the file, a polynomial equation is established for the process prediction value y′ and x: y′=θ T X=a0+a1x1+L L+a n x n , where θ = [a0, a1, LL, a n ],θ T is the transposed matrix of θ, θ is the discharge coefficient at different angles, X = [x0, x1, LL, x n ], X is the number of discharges at different angles, x0 is set to 1, and the initial values ​​of the discharge coefficients a1, a2, ..., an at different angles are set to 1.

2. The high-frequency partial discharge type identification method based on the gradient descent method according to claim 1 is characterized in that: The step (1.4) of calculating the error value comprises: Substitute the x and a values ​​into the polynomial y′=θ T X=a0+a1x1+L L+a n x n The process prediction value y′ is obtained, and the prediction error is calculated based on the actual value y of the total discharge amplitude and the process prediction value. Through m groups of X calculations, we can get m groups of y', and through m groups of y' and y data, we can calculate the prediction error. Calculate the error value of the loss function.

3. The high-frequency partial discharge type identification method based on the gradient descent method according to claim 1 is characterized in that: The adjustment and correction of each coefficient of the polynomial in step (1.5) includes: The gradient of the loss function is calculated according to the gradient descent method, that is, Iteratively update a according to the calculation results i Each coefficient, that is Among them, L is the step size of the gradient descent method. The step size is multiplied by the gradient of the loss function to get the distance the current position has dropped.

4. The high-frequency partial discharge type identification method based on the gradient descent method according to claim 1 is characterized in that: The step (1.6) of determining the gradient descent distance of each coefficient includes: When there is a i The gradient descent distance is greater than or equal to ε, and the iterative calculation continues Until all a i The gradient descent distance is less than ε at the end of a i End of iterative calculation.

5. The high-frequency partial discharge type identification method based on the gradient descent method according to claim 1 is characterized in that: The coefficients of the polynomial of the discharge amount and discharge number of the discharge type determined in step (1.7) include: Determine the polynomial y′=θ T X=a0+a1x1+L L+a n x n The θ value of the discharge coefficient at different angles, that is, [a0, a1, LL, a n ], it is approximately considered that y≈y′, that is, polynomial expressions for different discharge types are established: y=θ T X=a0+a1x1+L L+a n x n 。 6. The high-frequency partial discharge type identification method based on the gradient descent method according to claim 1 is characterized in that: The step (2) comprises: (2.1) Read the partial discharge times and discharge amount data of the equipment under test; (2.2) According to the different polynomial coefficients of different partial discharge types already existing in the database, Q(x) = a n +a1x1+L L+a n x n Different values ​​of the total discharge amplitude q are calculated; (2.3) According to the q value of different discharge types, the error between the actual discharge amount and the predicted value is calculated, that is, the error value of yq is calculated, and the error between the actual discharge amount and the predicted value is analyzed; (2.4) Identification and classification of partial discharge types, the error value is set within the allowable ε q The range is the type of partial discharge, where ε q Enter or set a fixed value parameter.

7. The high-frequency partial discharge type identification method based on the gradient descent method according to claim 1 is characterized in that: The step (3) comprises: (3.1) New partial discharge types are labeled, and all error values ​​are greater than the set ε q , labeling new types of partial discharges through manual intervention; (3.2) Solving the coefficients of the polynomial based on self-learning; (3.3) Determine the coefficients of the polynomial between the discharge amount and the number of discharges of the discharge type, i.e., the value of θ of the discharge coefficient; (3.4) Update the current new type to the existing database or discharge type model.

8. A high-frequency partial discharge type identification system based on gradient descent method, characterized in that: include: The database module is used to read the data of the determined partial discharge type, solve the coefficients of the fitting polynomial based on the gradient descent method, and establish a database of characteristic data and discharge types; it includes the following steps: (1.1) Label different groups of data according to the type of partial discharge; (1.2) Read the data of partial discharge times and discharge amount, and perform normalization processing on the per unit value according to the total amplitude of discharge amount and the total number of discharge amounts; (1.3) Fitting polynomials to set initial values ​​of coefficients; (1.4) Calculate the error value; (1.5) Complete the adjustment and correction of each polynomial coefficient based on the gradient descent method; (1.6) Determine whether the gradient descent distance of each coefficient is less than the allowable error judgment; (1.7) Determine the coefficients of the polynomials of discharge capacity and discharge times of the discharge type; (1.8) Establish the mapping relationship between y and θ corresponding to different partial discharge types, The step (1.3) of fitting the polynomial and setting the initial values ​​of the coefficients includes: The partial discharge data from 0 to 360° are divided into n groups according to the set angle step, which is expressed as: Data x1, x2, ..., x n , in p is the angle step, n is the number of data in one cycle; y is the total discharge amplitude, From the actual values ​​read from the file, a polynomial equation is established for the process prediction value y′ and x: y′=θ T X=a0+a1x1+L L+a n x n , where θ = [a0, a1, LL, a n ],θ λ is the transposed matrix of θ, θ is the discharge coefficient at different angles, X = [x0, x1, LL, x n ], X is the number of discharges at different angles, x0 is set to 1, and the discharge coefficients at different angles are a1, a2, ..., a n The initial value is set to 1. The identification and classification module is used to read the data of the partial discharge type of the equipment under test, calculate the total discharge amplitude based on the polynomial, and identify and classify the partial discharge type of the error; The update module is used to define new types of partial discharges, solve the coefficients of the polynomial based on self-learning, and update the database of newly added discharge types.

Images