Method for broadband ultrasonic detection of micro water in transformer oil

Abstract

The invention discloses a method for broadband ultrasonic detection of micro water in transformer oil. The method comprises the following steps: S1, acquiring broadband ultrasonic wave data passing through the transformer oil; s2, detecting the transformer oil obtained with the ultrasonic data to obtain pre-detection data corresponding to the ultrasonic data; s3, dividing the ultrasonic data into sample data and reference data, selecting the sample data, processing the sample data, and inputting the processed sample data into a neural network; s4, linking the training result with the micro-water content of the pre-detected data, and training the neural network to obtain the relationship between the ultrasonic data and the industrial process parameters of the transformer oil; and S5, detecting the micro-water content of the transformer oil by adopting the trained neural network, and comparing the micro-water content with reference data to verify the accuracy. According to the invention, online intelligent detection of the micro water in the transformer insulating oil can be realized.

Description

Technical Field

[0001] The invention relates to a method for broadband ultrasonic detection of micro-water in transformer oil, which is used in the technical field of transformer oil detection. Background Art

[0002] At present, the main means of operation safety detection of power transformers in my country is to carry out power outage maintenance and overhaul of transformers on schedule as a preventive measure. The "Guidelines for Maintenance of Power Transformers" stipulates that newly commissioned transformers need to undergo minor repairs once a year. Except for the first major overhaul five years after commissioning, the remaining major overhauls can be separated by a maximum of 10 years. Although the regular maintenance strategy has played a certain effective role in the safe and stable operation of my country's power equipment for a long time, with the continuous acceleration of the pace of social development, the continuous expansion of the scale of the power grid, and the continuous updating of power components, the disadvantages of traditional maintenance strategies have become increasingly prominent. According to the provisions in the maintenance manual, normally operating transformers must be withdrawn from operation and maintained on schedule, which will increase the frequency of transformer outages, increase the maintenance workload, increase the maintenance cost, and cause the problem of "excessive maintenance". In addition, if an abnormality occurs during the normal operation of the transformer, it may not be repaired in time, resulting in a decrease in the interruption of equipment operation, bringing many hidden dangers to the operation of the power grid, and causing the phenomenon of "insufficient maintenance" of equipment. The disadvantages of traditional maintenance methods are becoming increasingly prominent, and a reasonable operating condition maintenance mechanism has become the inevitable development direction of the power industry at home and abroad.

[0003] Maintaining a good insulation state is the key to a large power transformer having a good operating condition. At present, all power transformers in the power industry at home and abroad use transformer insulating oil as the main material for insulation and cooling. The aging of the transformer, the deterioration of the insulation performance, and the decomposition of its internal insulating materials will affect the change of the water content in the transformer oil. With the increase of the water content in the oil, the insulation performance of the transformer will decrease. When it reaches a certain level, the transformer may even have major faults such as burning of the insulating material and breakdown accidents, and this is an irreversible process. Therefore, real-time monitoring of the water content in the transformer oil is of great significance to ensure the safe operation of the transformer. Transformer oil is a liquid insulating medium. If the water content in the oil is too high, it will cause a series of faults that need to be solved in the transformer operation, such as increasing the loss of the insulating medium and reducing the breakdown voltage. When the power transformer is working normally, a strong magnetic field will be generated. If the water content in the transformer oil exceeds the standard when the transformer is working, under its complex strong magnetic field, the water and oil will be arranged in polarity order and arranged into impurity "bridges" in the direction of the electric field, so that the leakage current flowing through the impurity bridge increases and the breakdown voltage of the oil is reduced.

[0004] At present, the existing mature technical methods include:

[0005] (1) Karl Fischer method. Advantages: high precision and high sensitivity, can accurately measure trace moisture. Disadvantages: complex operation, high requirements for the detection environment, not suitable for on-site detection.

[0006] (2) Infrared spectroscopy. Advantages: non-contact, non-destructive testing. Disadvantages: relatively low detection accuracy, sensitive to interference from other impurities in the oil

[0007] (3) Humidity sensor method. Advantages: Online monitoring can be achieved. Disadvantages: Low detection accuracy and sensitivity, gas-liquid separation is required, which may affect the accuracy of the measurement results.

[0008] Currently, there is a lack of stable, reliable, convenient and efficient methods for online detection of transformer oil. Summary of the invention

[0009] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for broadband ultrasonic detection of trace water in transformer oil, which can realize online intelligent detection of trace water in transformer insulating oil.

[0010] A technical solution to achieve the above object is: a method for broadband ultrasonic detection of trace water in transformer oil, comprising the following steps:

[0011] S1, acquiring ultrasonic data of broadband ultrasound passing through transformer oil;

[0012] S2, testing the transformer oil for acquiring the ultrasonic data, and obtaining pre-test data corresponding to the ultrasonic data;

[0013] S3, dividing the ultrasonic data into sample data and reference data, selecting the sample data and processing it before inputting it into the neural network;

[0014] S4, linking the training results with the micro-water content of the pre-inspection data, training the neural network, and obtaining the relationship between the ultrasonic data and the industrial process parameters of the transformer oil;

[0015] S5, uses a trained neural network to detect the moisture content of transformer oil and compares it with reference data to verify accuracy.

[0016] Furthermore, in S1, the types of ultrasonic data include attenuation coefficient, amplitude, phase, flight time and flight speed.

[0017] Furthermore, ultrasonic data are acquired using ultrasonic waves of 200kHz, 300kHz, 400kHz, and 500kHz, respectively.

[0018] Furthermore, in S2, the method for detecting the transformer oil for obtaining ultrasonic data is the Karl Fischer method.

[0019] Furthermore, in S3, the method for processing the sample data includes reducing the dimension of the data.

[0020] Furthermore, the specific method of dimensionality reduction is to use kernel principal component analysis to reduce the dimensionality of the data.

[0021] Furthermore, the specific method of kernel principal component analysis is:

[0022] S3.1, given input sample;

[0023] S3.2, select the kernel function and calculate the kernel matrix;

[0024] S3.3, center the matrix;

[0025] S3.4, perform eigenvalue decomposition on the centered matrix;

[0026] S3.5, select the feature town corresponding to the first k largest eigenvalues ​​λ;

[0027] S3.6, dimensionality reduction.

[0028] Furthermore, in S4, the neural network used is a GA-BP neural network.

[0029] Furthermore, the specific process of realizing the GA-BP neural network is as follows:

[0030] S4.1, determine the neural network topology;

[0031] S4.2, weight threshold is encoded to obtain the initial population;

[0032] S4.3, decoding to obtain weights and thresholds;

[0033] S4.4, assigning the weight threshold to the newly created BP neural network;

[0034] S4.5, train and test the network;

[0035] S4.6, test error;

[0036] S4.7, calculate fitness;

[0037] S4.8, select chromosomes with high fitness for replication;

[0038] S4.9, crossover;

[0039] S4.10, variation;

[0040] S4.11, get a new population;

[0041] If the termination condition is met, stop training; if not, return to step S4.3.

[0042] Furthermore, in S4, the industrial process parameters are specifically viscosity, density, temperature, particle size and particle number.

[0043] The advantages of the method for broadband ultrasonic detection of trace water in transformer oil of the present invention are:

[0044] (1) The present invention uses an improved data preprocessing technology, which allows the data to retain more original features, has a direct impact on the prediction of the final result, and makes the budget result more accurate.

[0045] (2) The present invention uses improved KPCA to reduce the dimension of the micro-water parameters in transformer oil, so that the reduced-dimensional data is closer to the original data, reducing the amount of model calculation to a certain extent.

[0046] (3) The input amplitude and phase as well as the flight time and flight speed of the present invention more accurately summarize the broadband ultrasonic data set of micro-water in transformer oil, and the prediction accuracy is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Flowchart for implementing KPCA dimensionality reduction;

[0048] Figure 2 Flowchart for implementing GA-BP neural network. DETAILED DESCRIPTION

[0049] In order to better understand the technical solution of the present invention, the following is a detailed description through specific embodiments:

[0050] A method for broadband ultrasonic detection of trace water in transformer oil of the present invention comprises the following steps:

[0051] S1, obtains ultrasonic data of broadband ultrasound passing through transformer oil.

[0052] The types of ultrasonic data include but are not limited to attenuation coefficient, amplitude, phase, flight time and flight speed. Since the use of ultrasonic waves of different frequencies to comprehensively detect the water content of the measured material is more accurate than a single frequency, ultrasonic waves of 200kHz, 300kHz, 400kHz and 500kHz are used to obtain ultrasonic data. In the actual measurement of this implementation, a total of 120-dimensional data is obtained.

[0053] S2, testing the transformer oil from which the ultrasonic data is obtained, and obtaining pre-test data corresponding to the ultrasonic data. In order to ensure the measurement accuracy and make the matching degree between the neural algorithm and the actual result more accurate, the Karl Fischer method is used to test the transformer oil from which the ultrasonic data is obtained.

[0054] S3, divide the ultrasonic data into sample data and reference data, select the sample data and process it before inputting it into the neural network. The number of transformer oil samples taken is 100 groups, so the total number of input samples of the neural network is 120*100. Since the large number of dimensions affects the calculation speed and causes data redundancy, it is necessary to reduce the dimension of the data. KPCA is used for data dimension reduction.

[0055] The basic process of KPCA is to calculate the sample point X∈i in the original sample space with dimension d. d , introduce the mapping φ to project it into a high-dimensional feature space z=φ(x)X∈i with dimension d' d , and then implement the PCA method in the feature space to obtain the dimension reduction result y = Wz∈i k The derivation formula is as follows:

[0056] KPCA can be expressed as a linear model y=Wz∈i in a high-dimensional feature space k According to the representation theorem, the parameters W∈i of the linear model k×d (W represents the parameter matrix of the linear model, and its dimension is k*d, that is, the projection matrix from high-dimensional space to low-dimensional space) The optimal solution can be expressed as a linear combination of all samples:

[0057]

[0058] Among them, a i Represents the weight coefficient, which is used to represent the coefficient of linear combination of samples; Z i ∈i d represents the eigenvector in the low-dimensional space; φ(x i ) represents the mapping function φ to transform the sample x i The result after projection to high-dimensional space; then its j-th dimension projection vector w j ∈i d (w j Representing the j-th projection vector in the low-dimensional space) can be expressed as:

[0059]

[0060] in, represents the weight corresponding to the eigenvalue in principal component analysis, α j =(α1 j , α2 j , …α N j )∈R N .

[0061] According to the results of principal component analysis, the projection vector w j Should meet:

[0062]

[0063] Among them, λ j represents the eigenvalue of the jth principal component;

[0064] because The above formula can be expressed as:

[0065] ZZ T Zα j =λ j Zα j (4)

[0066] Z is a sample matrix, each row is a feature vector of a sample; in general, we do not know the specific form of the mapping φ, so we introduce the kernel function:

[0067] κ(x i ,x j )=φ(x i ) T φ(x j ),K=Z T Z (5)

[0068] Among them, κ(x i ,x j ) is used to represent two samples x i and x j Similarity in high-dimensional space, the kernel function maps φ(x i )·φ(x j ) is realized; the kernel matrix K represents the similarity between samples in the low-dimensional space; the original formula is further expressed as:

[0069] ZKα j = Zλ j α j (6)

[0070] If the following equation is satisfied, the above equation is automatically satisfied:

[0071] Kα j =λ j α j (7)

[0072] That is a j is the eigenvector of K, and we take the k eigenvectors a with the largest K. 1 ,a 2 ,....,a k (The eigenvector of the kernel matrix represents the direction of the kth principal component in the kernel principal component analysis), then the jth dimension coordinate y after the original sample x is projected j (representing the projection of sample x in low-dimensional space, based on KPCA) is:

[0073]

[0074] The process of implementing KPCA is as follows Figure 1 shown.

[0075] Specifically:

[0076] S3.1, given input sample;

[0077] S3.2, select the kernel function and calculate the kernel matrix;

[0078] S3.3, center the matrix;

[0079] S3.4, perform eigenvalue decomposition on the centered matrix;

[0080] S3.5, select the feature town corresponding to the first k largest eigenvalues ​​λ;

[0081] S3.6, dimensionality reduction.

[0082] S4, linking the training results with the micro-water content of the pre-inspection data, training the neural network, and obtaining the relationship between the ultrasonic data and the industrial process parameters of the transformer oil.

[0083] GA-BP neural network is a hybrid optimization model that combines Genetic Algorithm (GA) and Backpropagation Neural Network (BP). GA is used to optimize the initial weight threshold of BP neural network, thereby avoiding the common local optimal trap of BP neural network and accelerating training convergence.

[0084] The use of optimization algorithms can greatly affect the results. The training samples and test samples randomly generate initial weights and thresholds. If you want a stable and robust neural network, you need to optimize. After confirmation, the weights and thresholds will not be randomly generated, making the results more accurate. Implementing the GA-BP process is as follows: Figure 2 As shown, specifically:

[0085] S4.1, determine the neural network topology;

[0086] S4.2, weight threshold is encoded to obtain the initial population;

[0087] S4.3, decoding to obtain weights and thresholds;

[0088] S4.4, assigning the weight threshold to the newly created BP neural network;

[0089] S4.5, train and test the network;

[0090] S4.6, test error;

[0091] S4.7, calculate fitness;

[0092] S4.8, select chromosomes with high fitness for replication;

[0093] S4.9, crossover;

[0094] S4.10, variation;

[0095] S4.11, get a new population;

[0096] If the termination condition is met, stop training; if not, return to step S4.3.

[0097] In S4, the industrial process parameters are specifically viscosity, density, temperature, particle size and particle number.

[0098] S5, uses a trained neural network to detect the moisture content of transformer oil and compares it with reference data to verify accuracy.

[0099] Those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention, and are not intended to limit the present invention. As long as they are within the spirit of the present invention, any changes or modifications to the above embodiments will fall within the scope of the claims of the present invention.

Claims

1. A method for detecting trace water in transformer oil using broadband ultrasound, characterized in that: The steps include: S1, acquiring ultrasonic data of broadband ultrasound passing through transformer oil; S2, testing the transformer oil for acquiring the ultrasonic data, and obtaining pre-test data corresponding to the ultrasonic data; S3, dividing the ultrasonic data into sample data and reference data, selecting the sample data and processing it before inputting it into the neural network; S4, linking the training results with the micro-water content of the pre-inspection data, training the neural network, and obtaining the relationship between the ultrasonic data and the industrial process parameters of the transformer oil; S5, uses a trained neural network to detect the moisture content of transformer oil and compares it with reference data to verify accuracy.

2. The method for detecting trace water in transformer oil by broadband ultrasound according to claim 1, characterized in that: In S1, the types of ultrasonic data include attenuation coefficient, amplitude, phase, flight time and flight speed.

3. A method for detecting trace water in transformer oil using broadband ultrasound according to claim 2, characterized in that: Ultrasonic data were acquired using ultrasonic waves of 200kHz, 300kHz, 400kHz, and 500kHz, respectively.

4. The method for detecting trace water in transformer oil by broadband ultrasound according to claim 1, characterized in that: In S2, the method for detecting the transformer oil for obtaining ultrasonic data is the Karl Fischer method.

5. The method for detecting trace water in transformer oil by broadband ultrasound according to claim 1, characterized in that: In S3, the method for processing sample data includes reducing the dimension of the data.

6. A method for broadband ultrasonic detection of trace water in transformer oil according to claim 5, characterized in that: The specific method of dimensionality reduction is to use kernel principal component analysis to reduce the dimensionality of the data.

7. A method for broadband ultrasonic detection of trace water in transformer oil according to claim 6, characterized in that: The specific method of kernel principal component analysis is: S3.1, given input sample; S3.2, select the kernel function and calculate the kernel matrix; S3.3, center the matrix; S3.4, perform eigenvalue decomposition on the centered matrix; S3.5, select the feature town corresponding to the first k largest eigenvalues ​​λ; S3.6, dimensionality reduction.

8. The method for broadband ultrasonic detection of trace water in transformer oil according to claim 1, characterized in that: In S4, the neural network used is GA-BP neural network.

9. A method for broadband ultrasonic detection of trace water in transformer oil according to claim 8, characterized in that: The specific process of implementing the GA-BP neural network is as follows: S4.1, determine the neural network topology; S4.2, weight threshold is encoded to obtain the initial population; S4.3, decoding to obtain weights and thresholds; S4.4, assigning the weight threshold to the newly created BP neural network; S4.5, train and test the network; S4.6, test error; S4.7, calculate fitness; S4.8, select chromosomes with high fitness for replication; S4.9, crossover; S4.10, variation; S4.11, get a new population; If the termination condition is met, stop training; if not, return to step S4.

3.

10. The method for detecting trace water in transformer oil by broadband ultrasound according to claim 1, characterized in that: In S4, the industrial process parameters are specifically viscosity, density, temperature, particle size and particle number.

Images