Transformer winding state judgment method based on frequency response residual analysis

Abstract

The invention discloses a transformer winding state judgment method based on frequency response residual analysis, which comprises the following steps of: controlling a digital signal generator by using a computer to send a standard excitation signal to obtain a response signal, performing Fourier transform to obtain frequency response data # imgabs0 # and # imgabs1 #, and calculating frequency response curves Xk and X, and further performing linear regression processing on the data to obtain a residual sequence Dk, then calculating a standardized Euclidean distance # imgabs2 # and a distance threshold lambda i, but performing classification through the current distance threshold is not optimal, further calculation needs to be performed, calculating entropy E (Hi), then calculating information gain Gi, sequentially calculating and taking the group with the maximum information gain as an optimal segmentation point, and finally performing classification through the current distance threshold. At the moment, when # imgabs3 is at the break point, it is indicated that the winding state is normal, and when # imgabs4 is at the break point, it is indicated that the winding state is faulty.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer winding fault determination, and in particular to a transformer winding state determination method based on frequency response residual analysis. Background Art

[0002] With the rapid development of power systems, transformers, as key equipment for power transmission and distribution, have a direct impact on the safety and reliability of power systems. Transformer windings are core components of transformers, and their condition directly affects their performance and lifespan. However, due to factors such as long-term operation, mechanical vibration, overload, and short-circuit shock, transformer windings may deform, shift, or loosen, leading to insulation degradation and even failure. Therefore, real-time monitoring and intelligent determination of transformer winding conditions are crucial. Traditional transformer winding condition monitoring methods rely primarily on manual experience and offline detection methods, such as frequency response analysis (FRA). While these methods can reflect winding condition changes to a certain extent, they have the following limitations: The data acquisition and processing process is complex, requiring professional personnel and difficult to automate; the determination of winding condition relies on manual experience, which is highly subjective and has a high error rate; the inability to monitor winding condition in real time makes it difficult to detect potential faults in a timely manner; and the lack of intelligent analysis capabilities prevents efficient processing and classification of complex data.

[0003] In recent years, with the rapid development of artificial intelligence and signal processing technologies, intelligent monitoring methods have gradually become a research hotspot. However, existing methods often ignore the potential value of residual analysis when processing frequency response data, and lack scientific basis for determining distance thresholds and state classification, resulting in a high rate of misjudgment. Therefore, developing an intelligent method for determining transformer winding status based on frequency response residual analysis is of great theoretical and practical significance. Summary of the Invention

[0004] This application provides a method for determining the state of a transformer winding based on frequency response residual analysis. The proposed method can accurately and effectively determine the state of the transformer winding. The specific method includes the following steps:

[0005] 1. A method for determining the state of a transformer winding based on frequency response residual analysis, characterized in that an experimental platform comprises: a signal generator (1), a grounding electrode (2), a capacitor (3), a resistor (4), an inductor (5), a low-voltage bushing (6), a medium-voltage bushing (7), a high-voltage bushing (8), a low-voltage winding (9), a medium-voltage winding (10), a high-voltage winding (11), a box (12), an iron core (16), a noise reduction filter module (13), a signal acquisition module (14), a computer (15), a grounding electrode (2a), a capacitor (3a), a resistor (4a), and an inductor (5a). The specific testing method comprises the following steps:

[0006] Step 1: Obtain frequency response data

[0007] First, the response signal must be measured. A signal generator is used to generate a standard excitation signal. The generated excitation signal flows through the low-voltage bushing and is injected into the transformer winding. It then flows through the low-voltage winding, medium-voltage winding, and high-voltage winding, and then flows out of the transformer through the high-voltage bushing. It then flows through the noise reduction filter module and the signal acquisition module. Finally, the computer measures the output response. The measured input voltage is U in (t)=[U in1 U in2 …U inn ], the output voltage is U out (t)=[U out_1 U out_2 … U out_n ], repeat the above steps to obtain k sets of input voltage and output voltage. Then perform Fourier transform on the obtained data to calculate the frequency response data of the transformer winding. The calculation formula is as follows:

[0008]

[0009] in, and They represent the frequency domain information obtained by Fourier transform of output voltage and input voltage, respectively. k Represents the frequency response curve of the kth group, intercepting X k The frequency response data of the mid-frequency range of 1kHz-1MHz is recorded as Expressed as the amplitude of the i-th point of the k-th group frequency response curve, f i It is expressed as the frequency of the i-th point, n represents the number, and 1≤i≤n. Let the frequency response data under normal winding state be recorded as X=[(f1 x1)(f2 x2)···(f i x i )···(f n x n )],x i It is expressed as the amplitude of the i-th point of the frequency response curve under normal conditions, and n represents the number of points;

[0010] Step 2: Data processing

[0011] Perform linear autoregression on the data:

[0012]

[0013] D k =1.32(X k -M k )

[0014] Among them, p represents the past p moments, Mk It is expressed as predicting the value at the current k moment through the past p moments, and is calculated by the least squares method or maximum likelihood estimation method Φ i , Φ i is the weight parameter, D k Represented as the kth group of residual sequence;

[0015] Step 3: Determine the distance threshold

[0016] The residual sequence obtained from step 2 is recorded as It is expressed as the i-th subsequence under the k-th group of residual sequence, m represents the number. The residual sequence D under the normal winding state can be obtained from step 2. The residual sequence under the normal winding state is divided into several subsequences, which are recorded as D = [S1 S2···S i ···S m ], S i It is represented as the i-th subsequence of the residual sequence under normal winding conditions;

[0017] The residual sequence under normal conditions is used as the benchmark sequence, and the standardized Euclidean distance between each subsequence of the residual sequence under normal conditions and the residual sequence under unknown conditions is calculated. Here we assume that S1 is used as an example:

[0018]

[0019] in It is expressed as the normalized Euclidean distance between the first subsequence of the residual sequence under normal conditions and the kth group of residual sequences under unknown conditions. σ k It is expressed as the standard deviation of the residual sequence of the kth group when it is unknown whether it is normal;

[0020] Calculate the distance threshold:

[0021]

[0022] Among them, λ1 represents the distance threshold at this time, Expressed as in the 25th percentile;

[0023] The specific calculation process is as follows:

[0024] Calculate position:

[0025]

[0026] calculate

[0027] Calculated here There are two results. The first is when the calculated W value is an integer:

[0028]

[0029] The second is when the calculated W value is a non-integer:

[0030] Take the integer part of W and record it as

[0031]

[0032] Where W represents the normalized Euclidean distance between the first subsequence of the residual sequence and the Wth residual sequence under unknown normality, and k represents the total number of residual sequences under unknown normality.

[0033] Step 4: Determine the winding status

[0034] The distance threshold obtained in step 3 can be used to divide the residual sequence into two categories when it is unknown whether it is normal or not. When it is Class A, it means the winding state is normal. When it is Class B, it means the winding status is faulty. However, this is just taking S1 as an example. To determine whether it is the best split point based on the distance threshold here, the following process is required:

[0035] Calculate entropy:

[0036] E(H 1 )=-0.98p(d1)logp(d1)-0.86p(d2)logp(d2)

[0037] Among them, p(d1) and p(d2) are the probability of the residual sequence being a normal sequence and the probability of being an unknown sequence, respectively, and E(H 1 ) is expressed as the entropy of the sequence;

[0038] Calculate information gain:

[0039]

[0040] Among them, ||H 1 ||、 They represent the total number of samples in the sample sequence, the number of class A samples, and the number of class B samples, respectively. They are respectively expressed as the entropy of class A in the sample and the entropy of class B in the sample;

[0041] Finally, we get the information gain when taking S1 as an example, then calculate the information gain when S2, compare them, discard the one with the smaller information gain, and so on until S m , and finally take the maximum information gain, which is the best split point. When When , it indicates that the winding status is faulty. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of the method of the present invention

[0043] Figure 2 This is a system diagram of the method of the present invention Specific implementation methods

[0044] The present invention will be described in further detail below with reference to the accompanying drawings:

[0045] Figure 1 The flowchart of the transformer winding state determination method based on frequency response residual analysis is shown, which is characterized by analyzing the residual sequence to determine the winding state, and specifically includes the following steps:

[0046] 1. A method for determining the state of a transformer winding based on frequency response residual analysis, characterized in that an experimental platform comprises: a signal generator (1), a grounding electrode (2), a capacitor (3), a resistor (4), an inductor (5), a low-voltage bushing (6), a medium-voltage bushing (7), a high-voltage bushing (8), a low-voltage winding (9), a medium-voltage winding (10), a high-voltage winding (11), a box (12), an iron core (16), a noise reduction filter module (13), a signal acquisition module (14), a computer (15), a grounding electrode (2a), a capacitor (3a), a resistor (4a), and an inductor (5a). The specific testing method comprises the following steps:

[0047] Step 1: Obtain frequency response data

[0048] First, the response signal must be measured. A signal generator is used to generate a standard excitation signal. The generated excitation signal flows through the low-voltage bushing and is injected into the transformer winding. It then flows through the low-voltage winding, medium-voltage winding, and high-voltage winding, and then flows out of the transformer through the high-voltage bushing. It then flows through the noise reduction filter module and the signal acquisition module. Finally, the computer measures the output response. The measured input voltage is U in (t)=[U in_1 U in_2 … U in_n ], the output voltage is U out (t)=[U out_1 U out_2 … U out_n ], repeat the above steps to obtain k sets of input voltage and output voltage. Then perform Fourier transform on the obtained data to calculate the frequency response data of the transformer winding. The calculation formula is as follows:

[0049]

[0050] in, and They represent the frequency domain information obtained by Fourier transform of output voltage and input voltage, respectively. k Represents the frequency response curve of the kth group, intercepting X k The frequency response data of the mid-frequency range of 1kHz-1MHz is recorded as Expressed as the amplitude of the i-th point of the k-th group frequency response curve, f i It is expressed as the frequency of the i-th point, n represents the number, and 1≤i≤n. Let the frequency response data under normal winding state be recorded as X=[(f1 x1)(f2 x2)···(f i x i )···(f n x n )],x i It is expressed as the amplitude of the i-th point of the frequency response curve under normal conditions, and n represents the number of points;

[0051] Step 2: Data processing

[0052] Perform linear autoregression on the data:

[0053]

[0054] D k =1.32(X k -M k )

[0055] Among them, p represents the past p moments, M k It is expressed as predicting the value at the current k moment through the past p moments, and is calculated by the least squares method or maximum likelihood estimation method Φ i , Φ i is the weight parameter, D k Represented as the kth group of residual sequence;

[0056] Step 3: Determine the distance threshold

[0057] The residual sequence obtained from step 2 is recorded as It is expressed as the i-th subsequence under the k-th group of residual sequence, m represents the number. The residual sequence D under the normal winding state can be obtained from step 2. The residual sequence under the normal winding state is divided into several subsequences, which are recorded as D = [S1 S2··· S i ··· S m ], S i It is represented as the i-th subsequence of the residual sequence under normal winding conditions;

[0058] The residual sequence under normal conditions is used as the benchmark sequence, and the standardized Euclidean distance between each subsequence of the residual sequence under normal conditions and the residual sequence under unknown conditions is calculated. Here we assume that S1 is used as an example:

[0059]

[0060] in It is expressed as the normalized Euclidean distance between the first subsequence of the residual sequence under normal conditions and the kth group of residual sequences under unknown conditions. σ k It is expressed as the standard deviation of the residual sequence of the kth group when it is unknown whether it is normal;

[0061] Calculate the distance threshold:

[0062]

[0063] Among them, λ1 represents the distance threshold at this time, Expressed as in the 25th percentile;

[0064] The specific calculation process is as follows:

[0065] Calculate position:

[0066]

[0067] calculate

[0068] Calculated here There are two results. The first is when the calculated W value is an integer:

[0069]

[0070] The second is when the calculated W value is a non-integer:

[0071] Take the integer part of W and record it as

[0072]

[0073] Where W represents the normalized Euclidean distance between the first subsequence of the residual sequence and the Wth residual sequence under unknown normality, and k represents the total number of residual sequences under unknown normality.

[0074] Step 4: Determine the winding status

[0075] The distance threshold obtained in step 3 can be used to divide the residual sequence into two categories when it is unknown whether it is normal or not. When it is Class A, it means the winding state is normal. When it is Class B, it means the winding status is faulty. However, this is just taking S1 as an example. To determine whether it is the best split point based on the distance threshold here, the following process is required:

[0076] Calculate entropy:

[0077] E(H 1 )=-0.98p(d1)logp(d1)-0.86p(d2)logp(d2)

[0078] in, p (d1) and p(d2) are the probability of the residual sequence being a normal sequence and the probability of being an unknown sequence, respectively. 1 ) is expressed as the entropy of the sequence;

[0079] Calculate information gain:

[0080]

[0081] Among them, ||H 1 ||、 They represent the total number of samples in the sample sequence, the number of class A samples, and the number of class B samples, respectively. They are respectively expressed as the entropy of class A in the sample and the entropy of class B in the sample;

[0082] Finally, we get the information gain when taking S1 as an example, then calculate the information gain when S2, compare them, discard the one with the smaller information gain, and so on until S m , and finally take the maximum information gain, which is the best split point. When When , it indicates that the winding status is faulty.

Claims

1. A method for determining transformer winding status based on frequency response residual analysis, characterized in that: The experimental platform includes: a signal generator (1), a grounding electrode (2), a capacitor (3), a resistor (4), an inductor (5), a low-voltage bushing (6), a medium-voltage bushing (7), a high-voltage bushing (8), a low-voltage winding (9), a medium-voltage winding (10), a high-voltage winding (11), a box (12), an iron core (16), a noise reduction filter module (13), a signal acquisition module (14), a computer (15), a grounding electrode (2a), a capacitor (3a), a resistor (4a), and an inductor (5a). The specific testing method includes the following steps: Step 1: Obtain frequency response data First, the response signal must be measured. A signal generator is used to generate a standard excitation signal. The generated excitation signal flows through the low-voltage bushing and is injected into the transformer winding. It then flows through the low-voltage winding, medium-voltage winding, and high-voltage winding, and then flows out of the transformer through the high-voltage bushing. It then flows through the noise reduction filter module and the signal acquisition module. Finally, the computer measures the output response. The measured input voltage is U in (t)=[U in_1 U in_2 … U in_n ], the output voltage is U out (t)=[U out_1 U out_2 … U out_n ], repeat the above steps to obtain k sets of input voltage and output voltage. Then perform Fourier transform on the obtained data to calculate the frequency response data of the transformer winding. The calculation formula is as follows: in, and They represent the frequency domain information obtained by Fourier transform of output voltage and input voltage, respectively. k Represents the frequency response curve of the kth group, intercepting X k The frequency response data of the mid-frequency range of 1kHz-1MHz is recorded as Expressed as the amplitude of the i-th point of the k-th group frequency response curve, f i It is expressed as the frequency of the i-th point, n represents the number, and 1≤i≤n. Let the frequency response data under normal winding state be recorded as X=[(f1 x1) (f2 x2) ··· (f i x i ) ··· (f n x n )],x i It is expressed as the amplitude of the i-th point of the frequency response curve under normal conditions, and n represents the number of points; Step 2: Data processing Perform linear autoregression on the data: D k =1.32(X k -M k ) Among them, p represents the past p moments, M k It is expressed as predicting the value at the current k moment through the past p moments, and is calculated by the least squares method or maximum likelihood estimation method Φ i , Φ i is the weight parameter, D k Represented as the kth group of residual sequence; Step 3: Determine the distance threshold The residual sequence obtained from step 2 is recorded as It is expressed as the i-th subsequence under the k-th group of residual sequence, m represents the number. The residual sequence D under the normal winding state can be obtained from step 2. The residual sequence under the normal winding state is divided into several subsequences, which are recorded as D = [S1 S2···S i ···S m ], S i It is represented as the i-th subsequence of the residual sequence under normal winding conditions; The residual sequence under normal conditions is used as the benchmark sequence, and the standardized Euclidean distance between each subsequence of the residual sequence under normal conditions and the residual sequence under unknown conditions is calculated. Here we assume that S1 is used as an example: in It is expressed as the normalized Euclidean distance between the first subsequence of the residual sequence under normal conditions and the kth group of residual sequences under unknown conditions. σ k It is expressed as the standard deviation of the residual sequence of the kth group when it is unknown whether it is normal; Calculate the distance threshold: Among them, λ1 represents the distance threshold at this time, Expressed as in the 25th percentile; The specific calculation process is as follows: Calculate position: calculate Calculated here There are two results. The first is when the calculated W value is an integer: The second is when the calculated W value is a non-integer: Take the integer part of W and record it as Where W represents the normalized Euclidean distance between the first subsequence of the residual sequence and the Wth residual sequence under unknown normality, and k represents the total number of residual sequences under unknown normality. Step 4: Determine the winding status The distance threshold obtained in step 3 can be used to divide the residual sequence into two categories when it is unknown whether it is normal or not. When it is Class A, it means the winding state is normal. When it is Class B, it means the winding status is faulty. However, this is just taking S1 as an example. To determine whether it is the best split point based on the distance threshold here, the following process is required: Calculate entropy: E(H 1 )=-0.98p(d1)logp(d1)-0.86p(d2)logp(d2) Among them, p(d1) and p(d2) are the probability of the residual sequence being a normal sequence and the probability of being an unknown sequence, respectively, and E(H 1 ) is expressed as the entropy of the sequence; Calculate information gain: Among them, ||H 1 ||、 They represent the total number of samples in the sample sequence, the number of class A samples, and the number of class B samples, respectively. They are respectively expressed as the entropy of class A in the sample and the entropy of class B in the sample; Finally, we get the information gain when taking S1 as an example, then calculate the information gain when S2, compare them, discard the one with the smaller information gain, and so on until S m , and finally take the maximum information gain, which is the best split point. When When , it indicates that the winding status is faulty.

Images