A method and system for detecting substation faults

Abstract

This invention relates to the field of fault detection technology and discloses a substation fault detection method and system. The method includes: windowing the collected substation time-series data to extract a first feature vector and calculate a second feature vector; constructing a hybrid kernel function composed of a weighted combination of radial basis function (RBF) kernel functions and polynomial kernel functions, and calculating the weight coefficients of the RBF kernel function and polynomial kernel function; training a single-class support vector machine (SVM) using training samples under normal operating conditions based on preset model parameters and the hybrid kernel function to obtain the fault decision boundary; calculating an early warning threshold and setting an early warning boundary within the fault decision boundary; and performing the same feature extraction on the test data as on the training samples and inputting it into the trained SVM to determine the state of the test data. This invention improves the accuracy of fault detection, can identify early signs of equipment deterioration, and provides early warning information for operation and maintenance personnel.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, specifically to a method and system for detecting faults in substations. Background Technology

[0002] During operation, substation equipment may experience various faults due to internal and external factors. If these faults are not detected and addressed in a timely manner, they could lead to widespread power outages, causing severe economic losses and social impact. Therefore, researching efficient and accurate substation fault detection methods is of significant practical importance.

[0003] Traditional substation fault detection methods largely rely on manual inspections and fixed threshold alarms. However, these traditional methods generally suffer from high subjectivity and poor flexibility, being insensitive to subtle early-stage faults and ill-suited to complex and changing operating conditions. With the development of artificial intelligence, data-driven machine learning offers new approaches to fault detection. Among them, the One-Class Support Vector Machine (OCSVM) algorithm exhibits unique advantages: it only requires training the model using normal operating data, making it highly suitable for real-world scenarios where fault samples are sparse or difficult to obtain. However, applying OCSVM to substation fault detection still faces the following challenges: the time-series signals collected by substations typically possess complex, non-stationary, and nonlinear characteristics, containing rich time and frequency domain information. Extracting features solely from a single dimension (such as purely statistical features or frequency domain analysis) will fail to comprehensively capture subtle changes in equipment status, resulting in insufficient feature information and ultimately affecting the accuracy of the detection model. Furthermore, the performance of single-class support vector machines (SVMs) is highly dependent on the choice of kernel function. Both commonly used radial basis function (RBF) kernels and multinomial kernels have inherent advantages and disadvantages, making it difficult to simultaneously and accurately characterize both local details and global distributions of the data. This often results in suboptimal decision boundaries for the model, affecting the accuracy of the detection model. Moreover, traditional single-class SVMs are essentially binary classifiers, typically only able to distinguish between "normal" and "faulty" states. They cannot identify the intermediate transition process from a healthy state to a faulty state, thus lacking effective early warning capabilities. Summary of the Invention

[0004] This invention provides a substation fault detection method and system to solve the problems of low accuracy and lack of early warning capability in the existing technology.

[0005] In a first aspect, the substation fault detection method of the present invention includes the following steps:

[0006] The collected substation time-series data is processed by windowing. Multi-level wavelet packet decomposition is performed on each data window to extract the energy of each frequency band as the first feature vector. The statistical characteristics of the data within each data window are calculated as the second feature vector. The first and second feature vectors constitute the complete feature vector of the data window. A hybrid kernel function is constructed, consisting of a weighted combination of radial basis function (RBF) kernel function and polynomial kernel function. The weight coefficients of the RBF kernel function and polynomial kernel function are calculated based on the intra-class scatter of the first and second feature vectors under their respective kernel function mappings. Based on the preset model parameters and the hybrid kernel function, a single-class support vector machine is trained using training samples under normal operating conditions to obtain the fault decision boundary that distinguishes between normal and fault states. The distance from all training samples under normal operating conditions to the fault decision boundary is calculated. The warning threshold is determined based on the distance distribution, and a warning boundary defined by the warning threshold is set inside the fault decision boundary.

[0007] The test data is subjected to the same feature extraction as the training samples and input into the trained single-class support vector machine. When the output value of the single-class support vector machine falls outside the fault decision boundary, it is determined to be a fault state; when the output value of the single-class support vector machine falls between the fault decision boundary and the warning boundary, it is determined to be a warning state.

[0008] Preferably, the step of performing windowing processing on the collected substation time-series data includes:

[0009] The collected substation time-series data was divided into windows with a window length of 256 sampling points and a sliding step size of 64 sampling points.

[0010] Preferably, the step of performing multi-level wavelet packet decomposition on each data window and extracting the energy of each frequency band as the first feature vector includes:

[0011] The db4 wavelet basis is used to perform 3-level wavelet packet decomposition on each data window to obtain 8 frequency band coefficients; the energy of each frequency band coefficient is calculated to form an 8-dimensional first eigenvector.

[0012] Preferably, the calculation of statistical characteristics of the data within each data window as the second feature vector includes:

[0013] The mean, standard deviation, kurtosis, and skewness of the 256 sample points within the data window are calculated to form a 4-dimensional second feature vector.

[0014] Preferably, the step of calculating the weight coefficients of the radial basis function kernel function and the polynomial kernel function based on the intra-class scatter of the first and second eigenvectors under their respective corresponding kernel function mappings includes:

[0015] Calculate the trace of the within-class discreteness matrix of the first feature vector of all training samples under normal operating conditions under the radial basis kernel function mapping. ;

[0016] Calculate the trace of the within-class discreteness matrix of the second feature vector of all training samples under normal operating conditions under the polynomial kernel function mapping. ;

[0017] Weighting coefficients of radial basis kernel function Calculated using the following formula:

[0018] ;

[0019] Weighting coefficients of polynomial kernel function Calculated using the following formula:

[0020] .

[0021] Preferred, hybrid kernel function for:

[0022] ;

[0023] in, For data window The complete feature vector, For data window The complete feature vector, For data window The first eigenvector, For data window The first eigenvector, For data window The second eigenvector, For data window The second eigenvector, These are the weighting coefficients of the radial basis kernel function. The weighting coefficients are those of the polynomial kernel function. For radial basis kernel functions, It is a polynomial kernel function;

[0024] ;in, For kernel parameters;

[0025] Where d is the order of the polynomial kernel function and c is a constant.

[0026] Preferably, determining the warning threshold based on the distance distribution includes:

[0027] The distances from all training samples under normal operating conditions to the fault decision boundary constitute the decision value sample set;

[0028] Calculate the mean of the decision value sample set. with standard deviation ;

[0029] Use 3 The criteria determine the warning threshold T, which is calculated using the following formula:

[0030] .

[0031] Preferably, the preset model parameters include penalty coefficients, kernel parameters of the radial basis function, and the order of the polynomial kernel function.

[0032] Preferably, the preset model parameters are determined by grid search and cross-validation.

[0033] Secondly, the substation fault detection system of the present invention includes a memory and a processor. The memory stores computer instructions, and when the processor executes the computer instructions, it implements the above-mentioned substation fault detection method.

[0034] The beneficial effects of this invention are as follows: By combining the multi-layer wavelet packet energy features and statistical features of time-series data, this invention can obtain the operating status of substation equipment from both the time and frequency domains, enhancing the expressive power of feature vectors. Based on this, the constructed hybrid kernel function utilizes the complementary advantages of radial basis function and polynomial kernel function, assigning reasonable weights to different features based on intra-class discreteness, enabling a single-class support vector machine to generate a more accurate fault decision boundary, improving the accuracy of fault detection and reducing the false alarm rate. This invention, through the concept of an early warning boundary and the analysis of the distance distribution from normal samples to the fault boundary, achieves a three-level classification of equipment status: normal, early warning, and fault. This can identify early signs of equipment status deterioration, providing early warning information for operation and maintenance personnel, and improving the safety and reliability of power grid operation. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the substation fault detection method provided in an embodiment of the present invention. Detailed Implementation

[0036] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0037] like Figure 1 As shown, an embodiment of the substation fault detection method provided by the present invention includes the following steps:

[0038] S1. The collected substation time-series data is processed by windowing. Multi-level wavelet packet decomposition is performed on each data window. The energy of each frequency band is extracted as the first feature vector. The statistical characteristics of the data in each data window are calculated as the second feature vector. The first feature vector and the second feature vector constitute the complete feature vector of the data window.

[0039] In an optional embodiment, the step of windowing the collected substation time-series data includes: dividing the collected substation time-series data into windows with a window length of 256 sampling points and a sliding step of 64 sampling points.

[0040] The step of performing multi-level wavelet packet decomposition on each data window and extracting the energy of each frequency band as the first feature vector includes: performing three-level wavelet packet decomposition on each data window using the db4 wavelet basis to obtain eight frequency band coefficients; calculating the energy of each frequency band coefficient to form an eight-dimensional first feature vector.

[0041] The calculation of the statistical characteristics of the data within each data window is used as the second feature vector, including: calculating the mean, standard deviation, kurtosis, and skewness of the 256 sampling points within the data window, forming a 4-dimensional second feature vector.

[0042] For example, suppose a current time series data segment containing 1024 sampling points is collected from a circuit breaker in a substation. The current time series data is windowed with a window length of 256 points and a sliding step size of 64 points. The first data window contains sampling points 1 to 256, the second data window contains sampling points 65 to 320, and so on. For each data window of 256 sampling points, a 3-level wavelet packet decomposition is performed using the db4 wavelet basis, decomposing the signal of the data window into 8 different frequency bands from low to high frequency, obtaining a set of wavelet packet coefficients for each frequency band. The energy of the wavelet packet coefficients in each of the 8 frequency bands is calculated by summing the squares of all coefficients within each frequency band. For example, the energy of the first frequency band is E1, the second frequency band is E2, and so on up to the energy of the eighth frequency band, E8, thus forming an 8-dimensional first feature vector, such as [E1, E2, E3, E4, E5, E6, E7, E8]. Simultaneously, for the original 256 sampling points within the data window, statistical characteristics are calculated. For example, the mean is 150, the standard deviation is 5.2, the kurtosis is 3.1, and the skewness is 0.15. These four values ​​constitute a 4-dimensional second feature vector, such as [150, 5.2, 3.1, 0.15]. The first and second feature vectors together represent the state of the data window.

[0043] S2, construct a hybrid kernel function consisting of a weighted combination of radial basis function kernel function and polynomial kernel function, and calculate the weight coefficients of radial basis function kernel function and polynomial kernel function based on the intra-class scatter of the first eigenvector and the second eigenvector under their respective kernel function mappings.

[0044] In an optional embodiment, the step of calculating the weight coefficients of the radial basis function kernel function and the polynomial kernel function based on the intra-class scatter of the first feature vector and the second feature vector under their respective corresponding kernel function mappings includes:

[0045] Calculate the trace of the within-class discreteness matrix of the first feature vector of all training samples under normal operating conditions under the radial basis kernel function mapping. ;

[0046] Calculate the trace of the within-class discreteness matrix of the second feature vector of all training samples under normal operating conditions under the polynomial kernel function mapping. ;

[0047] Weighting coefficients of radial basis kernel function Calculated using the following formula:

[0048] ;

[0049] Weighting coefficients of polynomial kernel function Calculated using the following formula:

[0050] .

[0051] For example, suppose 1000 data windows under normal operating conditions are collected, and 1000 sets of first feature vectors and 1000 sets of second feature vectors are extracted. The 1000 first feature vectors are mapped to a high-dimensional space using a radial basis function kernel, and the trace of their within-class scatter matrix in the high-dimensional space is calculated. This reflects the degree of dispersion of energy characteristics under normal operating conditions, assuming that the calculated values ​​are... The value is 20. Then, a similar operation is performed on the 1000 second eigenvectors, mapping them using a polynomial kernel function and calculating the trace of the within-class scatter matrix, resulting in... . This reflects the degree of dispersion of statistical characteristics under normal operating conditions, assuming that the calculated values ​​are... The value is 80. The weighting coefficients of the radial basis function are calculated according to the formula. The weighting coefficient of the polynomial kernel function is 0.8. The value is 0.2. Because the first eigenvector has lower intra-class dispersion and more concentrated data, a higher weight is assigned to the corresponding radial basis kernel function.

[0052] In an optional embodiment, the hybrid kernel function for:

[0053] ;

[0054] in, For data window The complete feature vector, For data window The complete feature vector, For data window The first eigenvector, For data window The first eigenvector, For data window The second eigenvector, For data window The second eigenvector, These are the weighting coefficients of the radial basis kernel function. The weighting coefficients are those of the polynomial kernel function. For radial basis kernel functions, It is a polynomial kernel function;

[0055] ;in, For kernel parameters, It equals 0.5;

[0056] Where d is the order of the polynomial kernel function, d equals 2, and c is a constant, c equals 1.

[0057] For example, a hybrid kernel function is used to measure the similarity between any two data samples, fusing information from two different features. Suppose we need to calculate the similarity between sample i and sample j. The first feature vector of sample i is... The second feature vector is The first feature vector of sample j is The second feature vector is Calculate the radial basis function kernel function. and The square of the Euclidean distance between them, assuming a result of 0.5, then the value of the radial basis function kernel is approximately 0.779. Calculate... and The inner product of the polynomials, assuming the result is 30, then the value of the polynomial kernel function is 961. Using the weighting coefficients calculated in the previous step... equal to 0.8 and The value is equal to 0.2, and the calculated value of the mixed kernel function is 192.8232.

[0058] S3, based on preset model parameters and hybrid kernel functions, trains a single-class support vector machine using training samples under normal operating conditions to obtain the fault decision boundary that distinguishes between normal and fault states; calculates the distance from all training samples under normal operating conditions to the fault decision boundary, determines the warning threshold based on the distribution of distances, and sets the warning boundary defined by the warning threshold inside the fault decision boundary.

[0059] Specifically, the optimal model parameters of the single-class support vector machine are determined through grid search and cross-validation. These parameters include the penalty coefficient ν, the kernel parameter γ of the radial basis function, and the order d of the polynomial kernel function. All feature vectors extracted under normal operating conditions are used as the training set and input into the single-class support vector machine with configured parameters and a hybrid kernel function for training. An optimal hyperplane is then found, which serves as the fault decision boundary separating normal and fault states. The decision function f(x) of the fault decision boundary is equal to 0.

[0060] Each normal training sample is input into a pre-trained single-class support vector machine model to obtain the decision function output value f(x). The decision function output value represents the distance from the sample point to the fault decision boundary. The distances from all training samples under normal operating conditions to the fault decision boundary constitute the decision value sample set; the mean of the decision value sample set is calculated. with standard deviation ; Adopt 3 The criteria determine the warning threshold T, which is calculated using the following formula: Furthermore, within the fault decision boundary f(x) equal to 0, a warning boundary is determined by f(x) equal to T.

[0061] For example, after training a single-class support vector machine using a hybrid kernel function and normal operating condition training samples, a clear warning threshold needs to be obtained to determine whether new data is abnormal. All normal operating condition samples used for training, for example, 1000 samples, are input one by one into the trained model to obtain 1000 decision function output values.

[0062] Statistical analysis was performed on a sample set of decision function outputs consisting of 1000 values. It was assumed that the mean μ of the sample set was 1.5 and the standard deviation σ was 0.2. This indicates that the decision function outputs of the normal operating condition samples are concentrated around 1.5, with the fluctuation range determined by the standard deviation of 0.2. To obtain a warning boundary with high confidence, a 3... The criteria determine the warning threshold T, which is calculated to be 0.9. In subsequent online monitoring, when a new data window is input into a single-class support vector machine, if the output value of the decision function is less than 0.9, it is determined that the data point has significantly deviated from the normal state, triggering an early warning.

[0063] S4. The same features as those in the training samples are extracted from the test data and input into the trained single-class support vector machine. When the output value of the single-class support vector machine falls outside the fault decision boundary, it is determined to be a fault state; when the output value of the single-class support vector machine falls between the fault decision boundary and the warning boundary, it is determined to be a warning state.

[0064] Specifically, for the newly collected test data, the same windowing method, wavelet packet decomposition method, and statistical feature calculation method as in the training phase are used to obtain the first and second feature vectors of the test data. These two feature vectors are then input into a pre-trained single-class support vector machine to calculate the output value f(x) of the decision function. If f(x) is less than 0, the device state corresponding to the test data is determined to be a fault state; if f(x) is greater than or equal to 0 and less than the warning threshold T, it is determined to be a warning state; if f(x) is greater than or equal to the warning threshold T, it is determined to be a normal state.

[0065] The implementation principle of the substation fault detection method in this invention is as follows: This invention integrates the multi-layer wavelet packet energy features and statistical features of time-series data to comprehensively characterize the equipment operating status from both the time and frequency domains, enhancing the discriminative power of the feature vectors. Secondly, a hybrid kernel function based on intra-class discreteness weighting is constructed. This function combines the advantages of radial basis function and polynomial kernel function, establishing a better fault decision boundary for single-class support vector machines, thereby improving fault detection accuracy and reducing false alarms. Finally, by introducing the concept of "early warning boundary" and analyzing the distance from normal data points to the decision boundary, this invention establishes a three-level state assessment system of "normal-early warning-fault". This system can keenly capture the degradation trend of equipment performance, providing crucial early warnings for preventative maintenance.

[0066] An embodiment of the substation fault detection system provided by the present invention includes a memory and a processor. The memory stores computer instructions, and when the processor executes the computer instructions, it implements the substation fault detection method in the above embodiment.

[0067] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A substation fault detection method characterized by, Includes the following steps: The collected substation time-series data is windowed, and each data window undergoes multi-level wavelet packet decomposition. The energy of each frequency band is extracted as the first feature vector, and the statistical characteristics of the data within each data window are calculated as the second feature vector. The first and second feature vectors constitute the complete feature vector of the data window. A hybrid kernel function is constructed, consisting of a weighted combination of radial basis function (RBF) kernel function and polynomial kernel function. Based on the intra-class scatter of the first and second feature vectors under their respective kernel function mappings, the weight coefficients of the RBF kernel function and polynomial kernel function are calculated, including: Calculate the trace of the within-class scatter matrix of the first eigenvector under the radial basis kernel function mapping of all training samples under normal working conditions respectively ; respectively calculate the trace of the within-class scatter matrix of the second feature vector of the training sample under all normal working conditions in the polynomial kernel function mapping ; Weighting coefficients of radial basis kernel function for: ； Weighting coefficients of polynomial kernel function for: ； Based on preset model parameters and a hybrid kernel function, a single-class support vector machine is trained using training samples under normal operating conditions to obtain the fault decision boundary that distinguishes between normal and fault states. The distance from all training samples under normal operating conditions to the fault decision boundary is calculated, and a warning threshold is determined based on the distance distribution. A warning boundary defined by the warning threshold is then set within the fault decision boundary. The hybrid kernel function... for: ； For data window The complete feature vector, For data window The complete feature vector, For data window The first eigenvector, For data window The first eigenvector, For data window The second eigenvector, For data window The second eigenvector, These are the weighting coefficients of the radial basis kernel function. The weighting coefficients are those of the polynomial kernel function. For radial basis kernel functions, It is a polynomial kernel function; ;in, For kernel parameters; Where d is the order of the polynomial kernel function, and c is a constant; The test data is subjected to the same feature extraction as the training samples and input into the trained single-class support vector machine. When the output value of the single-class support vector machine falls outside the fault decision boundary, it is determined to be a fault state; when the output value of the single-class support vector machine falls between the fault decision boundary and the warning boundary, it is determined to be a warning state.

2. The substation fault detection method according to claim 1, characterized in that, The step of windowing the collected substation time-series data includes: The collected substation time-series data was divided into windows with a window length of 256 sampling points and a sliding step size of 64 sampling points.

3. The substation fault detection method according to claim 2, characterized in that, The step of performing multi-level wavelet packet decomposition on each data window and extracting the energy of each frequency band as the first feature vector includes: The db4 wavelet basis is used to perform 3-level wavelet packet decomposition on each data window to obtain 8 frequency band coefficients; the energy of each frequency band coefficient is calculated to form an 8-dimensional first eigenvector.

4. The substation fault detection method according to claim 2, characterized in that, The calculation of statistical characteristics of the data within each data window as the second feature vector includes: The mean, standard deviation, kurtosis, and skewness of the 256 sample points within the data window are calculated to form a 4-dimensional second feature vector.

5. The substation fault detection method according to claim 1, characterized in that, The method of determining the warning threshold based on the distribution of distances includes: The distances from all training samples under normal operating conditions to the fault decision boundary constitute the decision value sample set; Calculate the mean of the decision value sample set. with standard deviation ; Use 3 The criteria determine the warning threshold T, which is calculated using the following formula: 。 6. The substation fault detection method according to claim 1, characterized in that, The preset model parameters include the penalty coefficient, the kernel parameters of the radial basis kernel function, and the order of the polynomial kernel function.

7. The substation fault detection method according to claim 6, characterized in that, The preset model parameters are determined through grid search and cross-validation.

8. A substation fault detection system, characterized in that, It includes a memory and a processor. The memory stores computer instructions, and when the processor executes the computer instructions, it implements the substation fault detection method as described in any one of claims 1-7.

Images