Secondary AC loop insulation monitoring method and system based on multi-modal machine learning

Abstract

The invention belongs to the technical field of circuit automatic detection, and particularly provides a secondary alternating current circuit insulation monitoring method and system based on multi-mode machine learning, and the method comprises the steps: carrying out the processing of a high-frequency pulse current signal generated by a secondary circuit insulation defect fault through multivariate variational mode decomposition, time-frequency domain characteristics of the fault signals in different modes are obtained; feature parameters of fault diagnosis are selected, dimensionality reduction is carried out on the feature parameters of the fault time domain reconstruction signals and the fault frequency domain reconstruction signals through a local linear embedding algorithm, historical fault time-frequency feature vectors and real-time fault time-frequency feature vectors after dimensionality reduction serve as a training set and a test set respectively, classification is carried out through a support vector machine algorithm, and fault diagnosis is carried out. And fault types and fault rates of the fault types are obtained after classification. According to the method and the system, the insulation state of the secondary circuit is analyzed by extracting fault signal characteristics according to the fact that partial discharge can occur when insulation defects occur in the secondary circuit, and the partial discharge can generate high-frequency pulse current signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of circuit automation detection, and in particular, relates to a secondary AC circuit insulation monitoring method and system based on multimodal machine learning. Background Art

[0002] The secondary circuit of the power system is used to control, protect, adjust and monitor the working conditions of each component in the primary circuit, and plays a vital role in the safe and stable operation of the power system. However, due to the complex wiring of the secondary circuit and the wide range involved, the defects that exist are often hidden. Some defects are only discovered when the fault is serious, which seriously affects the correct operation of the protection device and even causes major power grid safety accidents.

[0003] At present, the main method for monitoring the insulation status of the secondary circuit is to use an infrared temperature measuring device to monitor the temperature of the secondary circuit terminal block to determine whether the terminal block is loose or falling off; and to use a current transformer to monitor the change of zero-sequence current to determine whether a wire break fault occurs in the secondary circuit.

[0004] CN106199340A discloses a method and device for determining a CT secondary circuit disconnection, which mainly connects the three-phase currents A, B, and C in parallel, and then collects the zero-sequence current in real time by connecting an external zero-sequence CT. At the same time, the three-phase currents A, B, and C and the external zero-sequence current are all connected to the AD sampling channel, and the zero-sequence current is calculated in real time. The difference between the collected zero-sequence current and the calculated zero-sequence current is compared to determine whether the CT secondary circuit is disconnected. This method can only determine whether a disconnection fault occurs in the secondary circuit, and cannot reflect the type and degree of insulation defect faults such as aging and damage of the insulation layer of the secondary circuit, loosening and falling off of the terminal, etc.

[0005] CN115436865A discloses a current transformer secondary circuit monitoring system based on infrared thermal imaging, and provides a current transformer secondary circuit monitoring system and method based on infrared thermal imaging. For accidents such as unplanned power outages caused by a sharp rise in the temperature of the current terminal due to high-voltage discharge when the secondary current circuit is open, and fires in the secondary current circuit cable and terminal box, the current transformer secondary circuit monitoring system collects infrared thermal imaging data, visible light image data, and temperature and humidity data in real time, and uses an Internet of Things wireless router to send the collected data to an intelligent analysis platform for analysis and calculation, identifies the operating status of the current transformer secondary circuit within the current monitoring range, and promptly gives reasonable operating prompt information, thereby solving the current problem of heavy manual monitoring tasks and easy omission of temperature measurement points in the current circuit transformer secondary circuit monitoring, and the temperature measurement results are greatly affected by different ambient temperatures, so it is difficult to find faults in a timely and accurate manner. The monitoring method of measuring the temperature of the terminal block can only reflect the loosening and falling off of the secondary circuit terminal, and cannot reflect the type and degree of insulation defect faults such as aging and damage of the secondary circuit insulation layer, disconnection and grounding, and it is difficult to comprehensively and reliably evaluate the insulation status of the secondary circuit. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a secondary AC circuit insulation monitoring method and system based on multimodal machine learning. When insulation defects occur in the secondary circuit, local discharge will occur, and local discharge will generate high-frequency pulse current signals. The insulation status of the secondary circuit is analyzed by extracting the fault signal characteristics.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: a secondary AC circuit insulation monitoring method based on multimodal machine learning, comprising the following steps: S1. The high-frequency pulse current signal generated by the secondary circuit insulation defect fault is processed by multivariate variational mode decomposition to obtain the time-frequency domain characteristics of the fault signal under different modes; S2, selecting the skewness and kurtosis of the time domain signal of the high-frequency pulse current and the amplitude and frequency distribution of the frequency domain signal as characteristic parameters of fault diagnosis, and using the local linear embedding algorithm to reduce the dimension of the characteristic parameters of the fault time domain reconstructed signal and the fault frequency domain reconstructed signal in step S1; S3. Use the reduced-dimensional historical fault time-frequency feature vector as a training set, and use the reduced-dimensional real-time fault time-frequency feature vector as a test set. Classify them using the support vector machine algorithm, and obtain the fault type and the failure rate of each fault type after classification.

[0008] In a preferred solution, in step S1, the time domain signal of the high-frequency pulse current signal is first subjected to multivariate variational modal decomposition, and then the different modal time domain signals generated after the decomposition are subjected to fast Fourier transform processing to obtain the corresponding frequency domain mode.

[0009] In a preferred embodiment, step S1 includes the following steps: S101, taking the secondary circuit fault time domain signal X' as the input signal, setting the initial conditions , ,in is the modal parameter, is the penalty factor; S102, perform multivariate variational mode decomposition on the fault time domain signal X' to obtain K The eigenmode functions are superimposed to form the recombined signal Y'; S103, calculate the Pearson correlation coefficient between the fault time domain signal X' and the reconstructed signal Y', and determine the size of the Pearson correlation coefficient and 0.95. The expression of the Pearson correlation coefficient is: ; in, represents the Pearson correlation coefficient, and Respectively represent the standard deviation of the fault time domain signal X' and the reconstructed signal Y', is the covariance of the fault time domain signal X' and the reconstructed signal Y'; S104, when When the fault time domain reconstruction signal Y1 is output, each intrinsic mode function of the reconstructed signal Y1 is transformed by fast Fourier transform to obtain the corresponding fault frequency domain reconstruction signal Y2; like , re-process the fault time domain signal X' by multivariate variational modal decomposition until .

[0010] In a preferred embodiment, step S2 includes the following steps: S201. Select the time-frequency characteristic parameters of the secondary AC circuit insulation defect: the skewness of the time domain signal Sk The expression is: ; Kurtosis of time domain signal Ku The expression is: ; in, is the time domain signal sample, is the mean value of the time domain signal samples, n is the number of time domain signal samples; S202, using a local linear embedding algorithm to reduce the dimension of the fault characteristic parameters: N The initial eigenvector matrix composed of the initial time-frequency characteristic parameters of each mode , the fault time-frequency feature vector composed of D feature parameters is obtained by dimensionality reduction processing through local linear embedding algorithm ,in, express N The time-frequency characteristic parameters of each initial mode, Indicates that the dimension reduction process is obtained D Time-frequency characteristic parameters; The calculation process is: 1) Calculate neighboring points: Calculate the Euclidean distance between any two sample points according to the metric space distance formula, and select the calculated closest distance to the sample point. K points as neighboring points, the expression of the Euclidean distance between any two sample points is: ; in, is the Euclidean distance between two sample points, , is any sample point selected, D for D A collection of sample points in dimensional Euclidean space, p= 2 is a constant; 2) Sample points are mapped to low-dimensional space: The calculation is performed according to the minimum reconstruction error formula, and the expression of the minimum reconstruction error formula is: ; in, It is a sample of K neighboring points, for and The weight between Z i For the i The local covariance matrix of sample points is w i For the i The local reconstruction weight vector of samples; Z i The expression is: ; w i The expression is: ; 3) Calculate the reconstruction weight matrix: The following conditions must be met when mapping all sample points to a low-dimensional space: ; in, is the output function, For sample The output vector of for of k neighboring points, Satisfy the conditions , , for N×N A sparse matrix used to store , for No. i List, for N×N The identity matrix i List, , represents the output feature vector, It is the final fault time-frequency feature parameter extracted after dimensionality reduction. In a preferred solution, in step three, classification is performed using a support vector machine algorithm, comprising the following steps: S301, feature mapping: introduce kernel function to map linearly inseparable fault feature data into high-dimensional space to achieve linear classification. The kernel function satisfies the following conditions: ; in x , z is a point in the input space, k is the kernel function, is the mapping function; S302, calculating a hyperplane: finding a hyperplane in a high-dimensional space so that the distance between each type of fault time-frequency characteristic data point and the hyperplane is maximized; S303, feature classification: obtain classification results of various fault feature data, and obtain the fault rate of real-time faults by comparing historical fault feature data.

[0011] The present invention also provides a secondary AC circuit insulation monitoring system based on multimodal machine learning, comprising a high-frequency current transformer, a collection unit, a convergence terminal and a secondary circuit insulation analysis information system, wherein the high-frequency current transformer is installed on phases A, B and C of the secondary circuit, and is used to collect high-frequency pulse current signals generated when insulation defects occur in the secondary AC circuit of a power plant; after receiving the pulse current high-frequency signal, the collection unit wirelessly transmits the data signal to the convergence terminal; the convergence terminal transmits the collected collection unit data of multiple unit cabinets to the secondary circuit insulation analysis information system; the secondary circuit insulation analysis information system executes a secondary AC circuit insulation monitoring method based on multimodal machine learning as described in claims 1 to 5, and performs data processing and fault diagnosis.

[0012] The present invention provides a secondary AC circuit insulation monitoring method and system based on multimodal machine learning, which has the following beneficial effects: 1. Rapidly locate the secondary circuit with insulation fault: Taking advantage of the fact that high-frequency pulse signals will be generated when insulation defects occur in the secondary circuit, high-frequency current transformers are installed on each secondary circuit to be monitored. When a fault occurs, the corresponding fault line can be quickly located, making it easier for inspection personnel to troubleshoot and solve the fault; 2. Deep mining of fault time and frequency domain information: The high-frequency current signal of insulation defect fault is decomposed into multiple modes from the time domain and frequency domain using the MVMD algorithm and FFT. Through the different characteristics of each mode, the characteristics of the fault signal can be analyzed in depth and from all aspects, making fault identification more realistic and reliable; 3. Comprehensive and reliable monitoring of the insulation status of the secondary circuit: After LLE dimensionality reduction processing of various time-frequency domain feature parameters and SVM feature classification, insulation defect fault types such as broken wire, grounding short circuit, aging and damage of the insulation layer, and loose and falling off of the terminal block can be obtained. When a certain insulation defect occurs, the fault type can be accurately identified through monitoring and the failure rate of the fault type can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 is a flow chart of the method of the present invention; Figure 2 A diagram of a secondary AC circuit insulation monitoring system of the present invention; Figure 3 It is a technical flow chart of the present invention. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0015] Embodiment 1: The present invention provides a secondary AC circuit insulation monitoring method based on multimodal machine learning, such as Figure 1 and Figure 3 As shown, including: S1. First, the time-frequency domain of the high-frequency current signal of the secondary circuit is processed by the multivariate variational mode decomposition algorithm (MVMD) to obtain the time-frequency domain characteristics of the fault signal under different modes.

[0016] The high-frequency pulse current signal generated by the secondary circuit insulation defect fault is decomposed by time-frequency MVMD. If the signal is directly subjected to fast Fourier transform (FFT), it can only reflect the entire frequency domain characteristics of the signal at all times. However, the time domain signal of the high-frequency current is first decomposed by MVMD, and then the different modal time domain signals generated after the decomposition are processed by FFT to obtain the corresponding frequency domain mode, so that the fault signal can be analyzed at multiple scales.

[0017] The basic principle of FFT is to use the discrete Fourier transform formula to convert the time domain signal into the frequency domain signal. The formula is: ; in, is the frequency domain signal, is the time domain signal corresponding to the frequency domain signal, represents the amplitude of the frequency domain signal, Indicates different frequency components.

[0018] The specific process is: 1) Take the power plant secondary circuit fault time domain signal X' as the input signal and set the initial conditions , ,in is the modal parameter, is the penalty factor, when When , MVMD has the strongest adaptability and the best denoising effect; 2) Perform MVMD on the fault time domain signal X' and decompose it into K The intrinsic mode functions (IMFs) are superimposed to form the recombined signal Y'; 3) Calculate the Pearson correlation coefficient between the fault time domain signal X' and the reconstructed signal Y' to determine with a magnitude of 0.95, where is the Pearson correlation coefficient, is the standard deviation, is the covariance; 4) When When the fault time domain reconstruction signal Y1 is output, if , reprocess the fault time domain signal X' through MVMD until ; At the same time, each IMF of the reconstructed signal Y1 is subjected to FFT to obtain the corresponding fault frequency domain reconstructed signal Y2.

[0019] The recombined signal Y' is selected by calculating the Pearson correlation coefficient (correlation coefficient is an indicator of the linear correlation between two variables) with the fault time domain signal X'. When the conditions are met, these recombination signals Y' are used as Y1.

[0020] S2. Select the skewness and kurtosis of the time domain signal and the amplitude and frequency distribution of the frequency domain signal as the characteristic parameters of fault diagnosis, and use the local linear embedding algorithm (LLE) to reduce the dimension of the characteristic parameters of the fault time domain reconstructed signal Y1 and the fault frequency domain reconstructed signal Y2 in S1.

[0021] S201. Select the time-frequency characteristic parameters of the secondary AC circuit insulation defect. When an insulation defect occurs in a secondary AC circuit, partial discharge will occur. The high-frequency pulse current generated by partial discharge is a non-stationary transient random signal. The skewness is often used in mathematics. Kurtosis To describe the characteristics of the distribution of random variables, the high-frequency pulse current signal of partial discharge is considered to be a discrete signal of time-amplitude, so skewness and kurtosis are selected as time domain characteristic parameters of fault diagnosis. Since the fault spectrum diagrams corresponding to different fault types and fault severities are different, they all have corresponding spectrum characteristics, so frequency content and frequency distribution are selected as frequency domain characteristic parameters of fault diagnosis.

[0022] S202, using the LLE algorithm to reduce the dimension of fault characteristic parameters. The initial characteristic parameters of the partial discharge high-frequency pulse signal are not globally linear. Since the LLE algorithm has the advantages of global optimization without iteration, high calculation efficiency and few variable parameters, the LLE algorithm is used to reduce the dimension of the fault characteristic parameters.

[0023] Will N The initial eigenvector matrix composed of the initial time-frequency characteristic parameters of each mode The fault time-frequency feature vector composed of D feature parameters is obtained by dimensionality reduction processing using the LLE algorithm. , in, express N The time-frequency characteristic parameters of each initial mode, Indicates that the dimension reduction process is obtained D time-frequency characteristic parameters.

[0024] The calculation process is: 1) Calculate neighboring points: According to the metric space distance formula , calculate the Euclidean distance between any two sample points, and select the calculated closest distance to the sample point K points as neighboring points.

[0025] in, is the Euclidean distance between two sample points, , is any sample point selected, D for D A collection of sample points in dimensional Euclidean space, p= 2 is a constant.

[0026] 2) Sample points are mapped to low-dimensional space According to the minimum reconstruction error formula Calculate, where According to the sample of K neighboring points, for and is the weight between, For the i The local covariance matrix of sample points is For the i The local reconstruction weight vector of samples.

[0027] in, It is a sample of K neighboring points, for and The weight between Z i For the i The local covariance matrix of sample points, w i For the i The local reconstruction weight vector of samples.

[0028] 3) Calculate the reconstruction weight matrix The following conditions must be met when mapping all sample points to a low-dimensional space: ; in, is the output function, For sample The output vector of for of k neighboring points, Satisfy the conditions , , for N×N A sparse matrix used to store , for No. i List, for N×N The identity matrix i List, , represents the output feature vector, It is the final fault time-frequency feature parameter extracted after dimensionality reduction. S3. The massive historical fault time-frequency feature vectors after dimensionality reduction are used as the training set, and the real-time fault time-frequency feature vectors after dimensionality reduction are used as the test set. The support vector machine algorithm (SVM) is used for training, and the fault type and the failure rate of each fault type are obtained after classification.

[0029] S301. The process of fault classification using the SVM algorithm is as follows: 1) Feature mapping: To avoid overfitting and improve generalization performance, a kernel function is introduced to map linearly inseparable fault feature data into a high-dimensional space to achieve linear classification. The kernel function satisfies the condition ,in x , z is a point in the input space, k is the kernel function, is the mapping function; 2) Calculate the hyperplane: Find a hyperplane in the high-dimensional space so that the distance between the time-frequency characteristic data points of various faults and the hyperplane is the largest; 3) Feature classification: Obtain the classification results of fault feature data such as broken wire, grounding short circuit, aging and damage of insulation layer, loose and falling off of terminal, and obtain the fault rate of real-time faults by comparing historical fault feature data.

[0030] Embodiment 2: The present invention also provides a secondary AC circuit insulation monitoring system based on multimodal machine learning, such as Figure 2As shown, it includes: a high-frequency current transformer (HFCT), a collection unit, a convergence terminal and a secondary circuit insulation analysis information system. The high-frequency current transformer is installed on the A phase, B phase and C phase of the key secondary circuit, and is used to collect the high-frequency pulse current signal generated when the insulation defect occurs in the secondary AC circuit of the power plant. The frequency of the current signal is 3MHz~30MHz, and low-frequency signals such as the base frequency are filtered out at the same time, avoiding the interference of the low-frequency signal on the high-frequency pulse current signal. Each HFCT monitors the corresponding secondary circuit, so the fault line can be located, and the data is sent to the collection unit through the RS485 communication method. After receiving the pulse current high-frequency signal, the collection unit transmits the data signal to the convergence terminal in a wireless manner. The convergence terminal sends the collected collection unit data of multiple unit cabinets to the secondary circuit insulation analysis information system for data processing and fault diagnosis.

[0031] The present invention can solve the single problem of scarcity of current monitoring technology for insulation defect faults in secondary circuits, solve the problem of conventional power frequency signal analysis (insulation defect fault pulse current signals are mainly concentrated in the high frequency range of 3-30 MHz), solve the problem that single-angle signal analysis cannot deeply mine fault signal feature data from multiple angles, and solve the problem that conventional technology cannot effectively identify the type of secondary circuit insulation defect faults (broken wire grounding short circuit, aging and damage of insulation layer, loose and falling off of terminal head) and the failure rate.

[0032] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A secondary AC circuit insulation monitoring method based on multimodal machine learning, characterized in that: The following steps are involved: S1. The high-frequency pulse current signal generated by the secondary circuit insulation defect fault is processed by multivariate variational mode decomposition to obtain the time-frequency domain characteristics of the fault signal under different modes; S2, selecting the skewness and kurtosis of the time domain signal of the high-frequency pulse current and the amplitude and frequency distribution of the frequency domain signal as characteristic parameters of fault diagnosis, and using the local linear embedding algorithm to reduce the dimension of the characteristic parameters of the fault time domain reconstructed signal and the fault frequency domain reconstructed signal in step S1; S3. Use the reduced-dimensional historical fault time-frequency feature vector as a training set, and use the reduced-dimensional real-time fault time-frequency feature vector as a test set. Classify them using the support vector machine algorithm, and obtain the fault type and the failure rate of each fault type after classification.

2. A secondary AC circuit insulation monitoring method based on multimodal machine learning according to claim 1, characterized in that: In the step S1, the time domain signal of the high-frequency pulse current signal is first subjected to multivariate variational modal decomposition, and then the different modal time domain signals generated after the decomposition are subjected to fast Fourier transform processing to obtain the corresponding frequency domain mode.

3. A secondary AC circuit insulation monitoring method based on multimodal machine learning according to claim 2, characterized in that: The step S1 includes the following steps: S101, taking the secondary circuit fault time domain signal X' as the input signal, setting the initial conditions , ,in is the modal parameter, is the penalty factor; S102, perform multivariate variational mode decomposition on the fault time domain signal X' to obtain K The eigenmode functions are superimposed to form the recombined signal Y'; S103, calculate the Pearson correlation coefficient between the fault time domain signal X' and the reconstructed signal Y', and determine the size of the Pearson correlation coefficient and 0.

95. The expression of the Pearson correlation coefficient is: ； in, represents the Pearson correlation coefficient, and Respectively represent the standard deviation of the fault time domain signal X' and the reconstructed signal Y', is the covariance of the fault time domain signal X' and the reconstructed signal Y'; S104, when When the fault time domain reconstruction signal Y1 is output, each intrinsic mode function of the reconstructed signal Y1 is transformed by fast Fourier transform to obtain the corresponding fault frequency domain reconstruction signal Y2; like , re-process the fault time domain signal X' by multivariate variational modal decomposition until .

4. The secondary AC circuit insulation monitoring method based on multimodal machine learning according to claim 1 is characterized in that: The step S2 includes the following steps: S201. Select the time-frequency characteristic parameters of the secondary AC circuit insulation defect: the skewness of the time domain signal Sk The expression is: ； Kurtosis of time domain signal Ku The expression is: ； in, is the time domain signal sample, is the mean value of the time domain signal samples, n is the number of time domain signal samples; S202, using a local linear embedding algorithm to reduce the dimension of the fault characteristic parameters: N The initial eigenvector matrix composed of the initial time-frequency characteristic parameters of each mode , the fault time-frequency feature vector composed of D feature parameters is obtained by dimensionality reduction processing through local linear embedding algorithm ,in, express N The time-frequency characteristic parameters of each initial mode, Indicates that the dimension reduction process is obtained D time-frequency characteristic parameters; The calculation process is: 1) Calculate neighboring points: Calculate the Euclidean distance between any two sample points according to the metric space distance formula, and select the calculated closest distance to the sample point. K points as neighboring points, the expression of the Euclidean distance between any two sample points is: ； in, is the Euclidean distance between two sample points, , is any sample point selected, D for D A collection of sample points in dimensional Euclidean space, p= 2 is a constant; 2) Sample points are mapped to low-dimensional space: The calculation is performed according to the minimum reconstruction error formula, and the expression of the minimum reconstruction error formula is: ； in, It is a sample of K neighboring points, for and The weight between Z i For the i The local covariance matrix of sample points is w i For the i The local reconstruction weight vector of samples; Z i The expression is: ； w i The expression is: ； 3) Calculate the reconstruction weight matrix: The following conditions must be met when mapping all sample points to a low-dimensional space: ； in, is the output function, For sample The output vector of for of k neighboring points, Satisfy the conditions , , for N×N A sparse matrix used to store , for No. i List, for N×N The identity matrix i List, , represents the output feature vector, It is the final fault time-frequency feature parameter extracted after dimensionality reduction.

5. The secondary AC circuit insulation monitoring method based on multimodal machine learning according to claim 1 is characterized in that: In the step 3, classification is performed using a support vector machine algorithm, including the following steps: S301, feature mapping: introduce kernel function to map linearly inseparable fault feature data into high-dimensional space to achieve linear classification. The kernel function satisfies the following conditions: ； in, x , z is a point in the input space, k is the kernel function, is the mapping function; S302, calculating a hyperplane: finding a hyperplane in a high-dimensional space so that the distance between each type of fault time-frequency characteristic data point and the hyperplane is maximized; S303, feature classification: obtain classification results of various fault feature data, and obtain the fault rate of real-time faults by comparing historical fault feature data.

6. A secondary AC circuit insulation monitoring system based on multimodal machine learning, characterized in that: It includes a high-frequency current transformer, a collection unit, a convergence terminal and a secondary circuit insulation analysis information system. The high-frequency current transformer is installed on phases A, B and C of the secondary circuit and is used to collect high-frequency pulse current signals generated when insulation defects occur in the secondary AC circuit of the power plant; after receiving the pulse current high-frequency signal, the collection unit transmits the data signal to the convergence terminal in a wireless manner; the convergence terminal transmits the collected collection unit data of multiple unit cabinets to the secondary circuit insulation analysis information system; the secondary circuit insulation analysis information system executes the secondary AC circuit insulation monitoring method based on multimodal machine learning as described in claims 1 to 5 to perform data processing and fault diagnosis.

Images