Switch cabinet partial discharge detection and identification algorithm based on TinyMaix

Abstract

A TinyMaix-based switch cabinet partial discharge detection and identification algorithm comprises the following steps: designing four typical insulation defect models, and respectively simulating corona discharge, suspension discharge, surface discharge and internal discharge; 200 samples are collected for each insulation defect, a sample set of 800 samples is formed in total, and a phase-resolved partial discharge (PRPD) graph is constructed according to a phase value in combination with a power frequency synchronization signal; the method comprises the following steps: dividing a power frequency sine wave into 360 phase windows, calculating a discharge frequency and a discharge amplitude average value in each phase window, constructing pulse counting distribution # imgabs0 # and average pulse amplitude distribution # imgabs1 #, extracting a probability pwin, a mean value mu and a variance sigma, and further calculating 19 statistical characteristics such as skewness Sk; a mean square error is selected as a loss function in each iteration of model solution, model parameters are updated according to the gradient of the loss function, and model convergence is accelerated; and inputting a partial discharge signal to be detected into the trained model, and outputting a classification result of the signal by the model to realize identification and classification of the partial discharge signal.

Description

Technical Field

[0001] The present invention discloses a partial discharge detection and recognition algorithm for switchgear based on TinyMaix, belonging to the field of partial discharge detection of power equipment. Background Art

[0002] The switchgear is an important device in the power system, and its safe and stable operation is crucial for the reliability of the power system. Partial discharge is an important sign of the insulation deterioration of the switchgear. Timely and accurately detecting and recognizing partial discharge is of great significance for preventing switchgear failures. For partial discharge recognition, traditional recognition algorithms such as support vector machines have achieved certain results in the partial discharge detection of switchgear. However, their model complexities are relatively high, the recognition accuracy is low, and when dealing with a large amount of real-time data, the running speed is slow, making it difficult to meet the requirements of real-time detection. Therefore, finding an identification algorithm with simple code and low model complexity is of great significance for the partial discharge detection of switchgear. Summary of the Invention

[0003] The purpose of the present invention is to design a complete set of partial discharge detection and recognition algorithms, and propose a partial discharge detection and recognition algorithm for switchgear based on TinyMaix, which is characterized by including the following steps:

[0004] S1: Combining the actual operation and detection conditions of the switchgear, designing 4 typical insulation defect models to respectively simulate corona discharge, floating discharge, surface discharge and internal discharge, and building an experimental platform to couple partial discharge pulses by using a bushing sensor and provide a phase synchronization signal;

[0005] S2: Using the sampling system to collect 200 samples for each insulation defect, forming a sample set of 800 samples in total, and combining the power frequency synchronization signal to construct a phase-resolved partial discharge (PRPD) map according to the phase value;

[0006] S3: Dividing the power frequency sine wave into 360 phase windows, calculating the number of discharges and the average discharge amplitude within each phase window, constructing a pulse count distribution and an average pulse amplitude distribution extracting the probability p win , mean μ, variance σ, and further calculating the skewness S k , kurtosis K u , cross-correlation coefficient cc and discharge coefficient QF and other 19 statistical features for subsequent model training;

[0007] S4: Selecting the mean square error as the loss function for each iteration of the model solution, and updating the model parameters through the backpropagation algorithm according to the gradient of the loss function to accelerate the model convergence;

[0008] S5: Using a decision tree as a classifier, input the partial discharge signal to be detected into the trained model, and the model outputs the classification result of the signal to achieve the recognition and classification of the partial discharge signal.

[0009] In step S1, it is necessary to design 4 typical insulation defect models, with their upper ends connected to high voltage, lower ends grounded, and connected to the experimental bench. The power supply is boosted by a step-up transformer and then connected in series with a current-limiting resistance protection circuit. The bushing sensor is used to couple the partial discharge pulse and provide a phase synchronization signal. Finally, the sampling device collects the partial discharge pulse through the bushing sensor.

[0010] In step S2, the pulse signal is processed by circuits such as filtering and amplification and then sampled by the ZYNQ7000 system. 2000 pulses can be collected at a time to form a discharge sample. 200 samples are collected for each insulation defect, and a total of 800 samples are formed into a sample set. Combining with the power frequency synchronization signal, a phase-resolved partial discharge (PRPD) map is constructed according to the phase value. The distribution of discharges in the positive and negative half-cycles of the power frequency is different under different insulation defects.

[0011] In step S3, count the number of discharges of each partial discharge sample in different amplitude intervals to obtain the partial discharge amplitude probability distribution H n (A), which is fitted with a two-parameter Weibull distribution, and different partial discharge types are distinguished by the shape parameter β.

[0012]

[0013] Among them, A is the amplitude of the partial discharge pulse, α is the scale parameter, β is the shape parameter, and β can be used to distinguish different partial discharge types. The corona discharge is 0.381, the floating discharge is 17.901, the surface discharge is 0.896, and the internal discharge is 1.403. Based on the normal distribution, the following statistical features can be obtained: skewness S k is the third moment of the signal, which is used to reflect the deviation degree of the discharge signal from the normal distribution. S k >0 indicates that the signal is skewed to the left. S k <0 indicates that the signal is skewed to the right.

[0014] Based on the normal distribution, the following statistical features can be obtained: skewness S k is the third moment of the signal, which is used to reflect the deviation degree of the discharge signal from the normal distribution. S k >0 indicates that the signal is skewed to the left. S k <0 indicates that the signal is skewed to the right.

[0015]

[0016] Kurtosis K uis the fourth moment of the signal, which is used to reflect the uniformity of the discharge signal relative to the normal distribution. K u > 0 indicates that the signal is clearer, K u < 0 indicates that the signal is flatter.

[0017]

[0018] The cross - correlation coefficient CC represents the similarity of the distribution shapes of the positive and negative half - cycles. CC = 0 indicates that the shapes are completely different. CC = 1 indicates that the shapes are the same.

[0019]

[0020] Among them and are the vertical coordinate values of the positive and negative half - cycles. The discharge coefficient QF is introduced to describe the difference between the positive - half - cycle discharge and the negative - half - cycle discharge. Then, the corresponding CC can also be corrected to mcc.

[0021]

[0022] Among them, A + and A - are the sum of the average discharge amplitudes in the positive and negative half - cycles, M + and M - are the number of discharge times in the positive and negative half - cycles, R A is the ratio of A + to A - and R M is the ratio of M + to M - of.

[0023] In the step S4, in each iteration of the model solution, the mean square error is selected as the loss function, and the model parameters are updated through the backpropagation algorithm according to the gradient of the loss function to accelerate the model convergence.

[0024]

[0025] where n is the number of training data, y i is the true output of each training data x i and is the predicted value of the model for x i of.

[0026] In the step S5, using the decision tree as the classifier, the local discharge signal to be detected is input into the trained model, and the model outputs the classification result of the signal to realize the identification and classification of the local discharge signal. Description of the Drawings

[0027] Figure 1Flowchart of the partial discharge detection and recognition algorithm for switchgear based on TinyMaix.

[0028] Figure 2 They are 4 typical insulation defect models designed.

[0029] Figure 3 It is the experimental platform built.

[0030] Figure 4 They are the phase-resolved partial discharge (PRPD) patterns of four typical partial discharge models.

[0031] Figure 5 It is a photo of the switchgear in a substation in Tianjin.

[0032] Figure 6 It is Figure 5 the corresponding PRPD pattern.

[0033] Figure 7 It is a photo of another switchgear in a substation in Tianjin.

[0034] Figure 8 It is Figure 7 the corresponding PRPD pattern. Specific implementation mode

[0035] The present invention is not limited to the implementation modes described below. The non-innovative labor achievements made by researchers in the field based on the present invention all belong to the protection scope of the present invention. The present invention will be further described below with reference to the accompanying drawings:

[0036] Figure 1 As shown, it is the flowchart of the partial discharge detection and recognition algorithm for switchgear based on TinyMaix. It includes the steps:

[0037] S1: Combining the actual operation and detection conditions of the switchgear, design 4 typical insulation defect models to respectively simulate corona discharge, floating discharge, surface discharge and internal discharge, and build an experimental platform to couple partial discharge pulses and provide phase synchronization signals by using a bushing sensor;

[0038] S2: Use the sampling system to collect 200 samples for each insulation defect, forming a sample set of 800 samples in total. Combine the power frequency synchronization signal and construct a phase-resolved partial discharge (PRPD) map according to the phase value;

[0039] S3: Divide the power frequency sine wave into 360 phase windows, calculate the discharge times and the average discharge amplitude in each phase window, and construct the pulse count distribution and the average pulse amplitude distribution Extract the probability p win 、the mean μ, the variance σ, and further calculate the skewness S k 、the kurtosis Ku Nineteen statistical features such as the cross - correlation coefficient cc and the discharge coefficient QF are used for subsequent model training;

[0040] S4: In each iteration of model solving, the mean square error is selected as the loss function, and according to the gradient of the loss function, the model parameters are updated through the backpropagation algorithm to accelerate model convergence;

[0041] S5: Using a decision tree as a classifier, the partial discharge signal to be detected is input into the trained model, and the model outputs the classification result of the signal to realize the identification and classification of the partial discharge signal.

[0042] The specific content of each process is as follows:

[0043] 1. Design the insulation defect model and build the experimental platform

[0044] In the actual operation and detection of switchgear, there are mainly four typical discharges. Therefore, four insulation defect models are designed, as Figure 2 shown, namely internal discharge, floating discharge, surface discharge and corona discharge. It is found that the bushing sensor embedded in the switchgear can couple partial discharge pulses and provide corresponding synchronous signals. Therefore, this sensor is used to collect signals. For different discharge insulation defects, the initial discharge voltages are slightly different: 5 KV for internal discharge, 4 KV for floating discharge, 3.5 KV for surface discharge, and 4.5 KV for corona discharge. The built experimental platform is as Figure 3 shown. A 220V power frequency AC power supply is used and connected to a 1:1000 step - up transformer. Then a current - limiting resistance protection circuit is connected in series, and the insulation defect is connected to the circuit. Finally, the sampling device collects partial discharge pulses through the bushing sensor, and the construction of the experimental platform conforms to the IEC60270 standard.

[0045] 2. Collection of partial discharge samples

[0046] The ZYNQ7000 sampling system can collect 2000 pulses at a time to form a discharge sample. 200 samples are collected for each insulation defect, and a total of 800 samples are formed into a sample set. Combining with the power frequency synchronous signal, a phase - resolved partial discharge (PRPD) pattern is constructed according to the phase value, as Figure 4 shown. The distributions of power frequency positive and negative half - cycle discharges under different insulation defects are different.

[0047] The positive and negative half-cycle discharges of internal discharges are concentrated at the rising edge of the power frequency voltage and have obvious symmetry, which is determined by the electric field distribution and discharge mechanism of internal defects in insulating materials; the positive and negative half-cycle discharge pulses of floating discharges are relatively symmetric and have large amplitudes. The discharges are concentrated in the first and third quadrants and are significantly different from other discharges, which is related to the force and charge distribution characteristics of the floating metal body in the electric field; the pulse amplitudes of surface discharges are concentrated in a lower range, and the positive and negative half-cycle discharges show a clear strip or band distribution; the positive half-cycle discharge pulses of corona discharges are higher and sparser, concentrated near 90 degrees, and the negative half-cycle discharge pulses are lower and denser, concentrated near 270 degrees.

[0048] 3. Statistical Feature Extraction

[0049] Divide the power frequency sine wave into 360 phase windows, calculate the number of discharges and the average discharge amplitude within each phase window, and construct the pulse count distribution and the average pulse amplitude distribution The probability p of the phase in the distribution of different discharge types win , mean μ, and variance σ are as follows:

[0050]

[0051] where N is the number of phase windows in the positive and negative half-cycles, and y i is the vertical coordinate value. Based on the normal distribution, the following statistical features can be obtained: skewness S k is the third moment of the signal, used to reflect the deviation degree of the discharge signal from the normal distribution. S k >0 indicates that the signal is skewed to the left. S k <0 indicates that the signal is skewed to the right.

[0052]

[0053] Kurtosis K u is the fourth moment of the signal, used to reflect the uniformity degree of the discharge signal relative to the normal distribution. K u >0 indicates that the signal is clearer, and K u <0 indicates that the signal is flatter.

[0054]

[0055] The cross-correlation coefficient CC represents the similarity of the distribution shapes of the positive and negative half-cycles. CC = 0 indicates completely different shapes. CC = 1 indicates the same shape.

[0056]

[0057] where and are the vertical coordinate values in the positive and negative half - cycles. The discharge coefficient QF is introduced to describe the difference between the positive - half - cycle discharge and the negative - half - cycle discharge. Then, the corresponding CC can also be corrected to mcc.

[0058]

[0059] Among them, A + and A - is the sum of the average discharge amplitudes in the positive and negative half - cycles, M + and M - is the number of discharge times in the positive and negative half - cycles, R A is the ratio of A + to A - and R M is the ratio of M + to M - Finally, 19 statistical features are extracted to train the model for partial discharge detection.

[0060] 4. Model Training

[0061] In each iteration of model solution, the mean square error is selected as the loss function, and according to the gradient of the loss function, the model parameters are updated through the back - propagation algorithm to accelerate model convergence.

[0062]

[0063] where n is the number of training data, y i is the true output of each training data x i , is the predicted value of the model for x i .

[0064] 5. Partial Discharge Signal Recognition

[0065] Using the decision tree as the classifier, the partial discharge signal to be detected is input into the trained model, and the classification result of the signal is output by the model to realize the recognition and classification of the partial discharge signal.

[0066] 6. Analysis of Measured Data

[0067] Detection is carried out on two switchgear cabinets in a substation in Tianjin, as shown in Figure 5 . It is found that there are corona discharges and internal discharges respectively.

[0068] Detection is carried out on switchgear cabinet 312 in the substation. After the data is sampled, the number of discharge times of each partial discharge sample in different amplitude intervals is counted to obtain the partial discharge amplitude probability distribution H n(A), it is fitted with a two-parameter Weibull distribution, and different partial discharge types are distinguished by the shape parameter β. The obtained β is 0.342, and it is preliminarily judged as corona discharge. The pulse count distribution is constructed and the average pulse amplitude distribution The probability p is extracted win , the mean μ, the variance σ, and the skewness S is further calculated k , the kurtosis K u , 19 statistical features such as the cross-correlation coefficient cc and the discharge coefficient QF are selected. The mean squared error is used as the loss function to accelerate the model convergence. Finally, a decision tree is used as the classifier. The partial discharge signal to be detected is input into the model, and the classification result of the model output signal is obtained. The detected three-phase PRPD patterns are as Figure 6 shown. It can be clearly seen that there are obvious phase offsets in the three-phase PRPD patterns. The discharges in phase A are concentrated in the first quadrant, and the discharges are concentrated in the voltage rising period. Moreover, the discharge amplitudes of adjacent cabinets decay in turn. Therefore, the result of partial discharge diagnosis is: there is corona discharge in phase A.

[0069] The 2021 switchgear cabinets in the substation as Figure 7 shown are detected. After the data is sampled, the discharge times of each partial discharge sample in different amplitude intervals are counted, and the partial discharge amplitude probability distribution H n (A) is obtained. It is fitted with a two-parameter Weibull distribution, and different partial discharge types are distinguished by the shape parameter β. The obtained β is 1.212, and it is preliminarily judged as internal discharge. The pulse count distribution is constructed and the average pulse amplitude distribution The probability p is extracted win , the mean μ, the variance σ, and the skewness S is further calculated k , the kurtosis K u , 19 statistical features such as the cross-correlation coefficient cc and the discharge coefficient QF are selected. The mean squared error is used as the loss function to accelerate the model convergence. Finally, a decision tree is used as the classifier. The partial discharge signal to be detected is input into the model, and the classification result of the model output signal is obtained. The detected three-phase PRPD patterns are as Figure 8 shown. It can be seen that the phases of phase C are concentrated in the first and third quadrants. Affected by the voltage, the discharges gradually expand to the second and fourth quadrants. The amplitudes of adjacent switchgear cabinets decay in turn. Therefore, the result of partial discharge diagnosis is: there is internal discharge in phase C.

Claims

1. A switch cabinet partial discharge detection and identification algorithm based on TinyMaix, characterized in that: The following steps are involved: S1: Combined with the actual operation and detection of switchgear, four typical insulation defect models are designed to simulate corona discharge, suspension discharge, surface discharge and internal discharge respectively, and an experimental platform is built to use bushing sensors to couple partial discharge pulses and provide phase synchronization signals; S2: 200 samples are collected for each insulation defect using the sampling system to form a sample set of 800 samples in total, and a phase-resolved partial discharge (PRPD) diagram is constructed based on the phase value in combination with the power frequency synchronization signal; S3: Divide the power frequency sine wave into 360 phase windows, calculate the number of discharges and the average discharge amplitude in each phase window, and construct the pulse count distribution and average pulse amplitude distribution Extraction probability p win , mean μ, variance σ, and further calculate the skewness S k , Kurtosis K u , cross-correlation coefficient cc and discharge coefficient QF and other 19 statistical features are used for subsequent model training; S4: The mean square error is used as the loss function for each iteration of the model solution, and the model parameters are updated through the back propagation algorithm according to the gradient of the loss function to accelerate the convergence of the model; S5: Using the decision tree as a classifier, the partial discharge signal to be detected is input into the trained model, and the model outputs the classification result of the signal to realize the recognition and classification of the partial discharge signal.

2. A switch cabinet partial discharge detection and identification algorithm based on TinyMaix according to claim 1, characterized in that: Four typical insulation defect models are designed, and bushing sensors are used to collect partial discharge signals. The process is as follows: Four typical insulation defect models are designed. The upper end is connected to high voltage and the lower end is grounded. They are directly connected to the constructed experimental bench. The power supply is boosted by a step-up transformer and then connected in series with a current-limiting resistor to protect the circuit. The bushing sensor is used to couple the partial discharge pulse and provide a phase synchronization signal. Finally, the sampling device collects the partial discharge pulse through the bushing sensor. The construction of the experimental bench complies with the IEC60270 standard.

3. A switch cabinet partial discharge detection and identification algorithm based on TinyMaix according to claim 1, characterized in that: The sampling system is used to collect 200 samples for each insulation defect, and the phase value is used to construct a phase-resolved partial discharge (PRPD) diagram. The process is as follows: After the pulse signal is processed by filtering, amplification and other circuits, it is sampled by the ZYNQ7000 system. 2000 pulses can be collected at a time to form a discharge sample. 200 samples are collected for each insulation defect, forming a sample set of 800 samples in total. Combined with the power frequency synchronization signal, a phase-resolved partial discharge (PRPD) diagram is constructed according to the phase value. The distribution of power frequency positive and negative half-cycle discharges is different under different insulation defects.

4. The switch cabinet partial discharge detection and identification algorithm based on TinyMaix according to claim 1 is characterized in that: Various statistical features are extracted, and the process is as follows: Count the number of discharges of each partial discharge sample in different amplitude ranges to obtain the partial discharge amplitude probability distribution H n (A) Fitting with two-parameter Weibull distribution, different PD types are distinguished by shape parameter β. Where A is the amplitude of the partial discharge pulse, α is the scale parameter, and β is the shape parameter. β can be used to distinguish different types of partial discharges. Corona discharge is 0.381, suspension discharge is 17.901, surface discharge is 0.896, and internal discharge is 1.

403. Based on the normal distribution, the following statistical characteristics can be obtained: skew S k It is the third-order moment of the signal, which is used to reflect the degree of deviation of the discharge signal from the normal distribution. k >0 means the signal is shifted to the left. k <0 means the signal is shifted to the right. Kurtosis K u It is the fourth-order moment of the signal, which is used to reflect the uniformity of the discharge signal relative to the normal distribution. u >0 means the signal is clearer, K u <0 means the signal is flatter. The cross-correlation coefficient CC indicates the similarity of the positive and negative half-cycle distribution shapes. CC = 0 indicates that the shapes are completely different. CC = 1 indicates that the shapes are the same. where y i + and i - is the vertical coordinate value of the positive and negative half-cycles. The discharge coefficient QF is introduced to describe the difference between the positive half-cycle discharge and the negative half-cycle discharge. Then, the corresponding CC can also be corrected to mcc. Among them, A + and A - is the sum of the average discharge amplitudes in the positive half cycle and the negative half cycle, M + and M - is the number of discharges in the positive and negative half cycles, R A Yes A + With A - The ratio, R M It is M + With M - ratio.

5. The switch cabinet partial discharge detection and identification algorithm based on TinyMaix according to claim 1 is characterized in that: The mean square error is used as the loss function for each iteration of the model solution to speed up the convergence of the model. The process is as follows: In each iteration of the model solution, the mean square error is selected as the loss function, and the model parameters are updated through the back propagation algorithm according to the gradient of the loss function to accelerate the convergence of the model. Where n is the number of training data, y i For each training data x i The real output is For the model pair x i The predicted value of .

6. The switch cabinet partial discharge detection and identification algorithm based on TinyMaix according to claim 1 is characterized in that: The process of identifying partial discharge signals is as follows: Using the decision tree as a classifier, the partial discharge signal to be detected is input into the trained model, and the model outputs the classification result of the signal to realize the recognition and classification of the partial discharge signal.

Images