A method for discriminating intermittent faults of an inverter system by distribution characteristics

Abstract

The application discloses a method for discriminating intermittent faults of an inverter system through distribution characteristics, and the method comprises the following steps: collecting three-phase voltage signals of the inverter system under normal operation at equal intervals and processing the three-phase voltage signals to obtain characteristic data at each sampling moment; then, a plurality of basic detectors trained offline are used to complete feature extraction of the characteristic data; subsequently, KL divergence of sequence distribution is calculated as a difference factor to complete long-term sequence distribution feature extraction, and a sliding window singular value is used as a single-point feature of a monitoring object; finally, the two types of characteristic factors are fused as a monitoring index to complete real-time online detection of intermittent faults of the inverter system.

Description

A method for identifying intermittent faults in inverter systems based on distribution characteristics Technical Field

[0001] This invention belongs to the field of power system technology, and more specifically, relates to a method for identifying intermittent faults in inverter systems based on distribution characteristics. Background Technology

[0002] With the development of science and technology, the importance of power systems in modern society has become increasingly prominent. Inverter systems, as one of the key components in the DC-AC conversion process of the power grid, are widely used in photovoltaic energy, wind power generation, aerospace, and other fields. Once an inverter system fails, it may cause the entire power system to shut down, severely impacting social production and causing substantial property damage. Furthermore, as a core component of the power system, inverter system faults can spread and propagate along the grid topology, affecting the stability and lifespan of other components in the system. Therefore, timely and effective fault detection technology can effectively protect the safety and stability of the power system and minimize downtime and maintenance costs.

[0003] Intermittent faults are a common type of fault in power electronic systems. Unlike persistent faults, intermittent faults often disappear spontaneously without human intervention, but reappear randomly during system operation. Intermittent faults typically have small amplitudes and short durations. Therefore, detection methods for intermittent faults need to possess both good real-time performance and high sensitivity. Existing intermittent fault detection methods can be divided into analytical model-based methods and data-driven methods. Analytical model-based methods require mathematical modeling of the actual process, achieving fault detection by comparing actual data with model outputs. Existing analytical models often calculate residuals by constructing observers. However, analytical model-based methods heavily rely on highly accurate modeling and parameter estimation of the actual system, resulting in significant costs and long development cycles. Furthermore, with technological iterations, existing models require structural modifications and parameter adjustments, leading to even greater overhead in practical applications. Currently, data-driven methods have received widespread attention, replacing the complex and cumbersome model construction through training with rich historical data. Principal component analysis (PCA) is a classic data-driven method, but its direct application in the field of intermittent fault detection has not achieved satisfactory detection results. To extract the dynamic characteristics of a process, Dynamic Principal Component Analysis (DPCA) was proposed. Furthermore, the sliding window concept was introduced to extract sequence information, leading to Moving Average Principal Component Analysis (MAPCA). It is noted that different methods are often more sensitive to different data features, and integrating different detectors into a new detection model often yields better fault detection performance. However, most existing ensemble frameworks only utilize single-point features of the data or simply perform linear transformations using sliding window techniques, which often fails to effectively extract the sequence distribution characteristics of the process. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for identifying intermittent faults in inverter systems through distribution characteristics. This method measures the differences in the sequence distribution information of multiple variables within a sliding window and designs intermittent fault detection indicators based on singular value decomposition, thereby completing the intermittent fault detection of inverter systems.

[0005] To achieve the above-mentioned objective, the present invention provides a method for identifying intermittent faults in an inverter system based on distribution characteristics, characterized by comprising the following steps:

[0006] (1) Collect and process the three-phase voltage signals of the inverter system under normal operation at equal intervals;

[0007] (2) Construct a weak detector and train it offline;

[0008] (3) Set control limits for detecting intermittent faults in the inverter system through sliding window analysis and processing;

[0009] (4) Collect characteristic data of the inverter system in real time, calculate the fault monitoring index at the current moment, and then perform real-time online detection of intermittent faults of the inverter system according to the set control limits.

[0010] The objective of this invention is achieved as follows:

[0011] This invention discloses a method for identifying intermittent faults in an inverter system based on distribution characteristics. The method involves collecting and processing three-phase voltage signals of the inverter system under normal operation at equal intervals to obtain characteristic data at each sampling time. Then, several basic detectors trained offline are used to extract features from the characteristic data. Subsequently, the KL divergence of the sequence distribution is calculated as a difference factor to extract long-term sequence distribution features, and sliding window singular values ​​are used as single-point features of the monitoring object. Finally, the two types of feature factors are fused as a monitoring index to achieve real-time online detection of intermittent faults in the inverter system.

[0012] Meanwhile, the method for identifying intermittent faults in an inverter system based on distribution characteristics, as described in this invention, also has the following advantages:

[0013] Beneficial effects:

[0014] (1) This invention utilizes integrated sliding window technology to fuse statistics obtained from different basic detectors through the sliding window singular value fusion, and calculates the second norm of the singular value vector to characterize the single-point features of the sample for intermittent fault detection.

[0015] (2) The present invention calculates the probability density distribution in the sliding window and the overall probability density distribution, and then measures the difference between the two density distributions by KL divergence. Finally, it combines the sliding window singular value with the state monitoring index of the three-phase inverter system, realizing effective process feature extraction and intermittent fault detection. Attached Figure Description

[0016] Figure 1 is a flowchart of a method for identifying intermittent faults in an inverter system based on distribution characteristics according to the present invention.

[0017] Figure 2 shows the detection results of the present invention for intermittent faults caused by inductor degradation in inverter circuits. Detailed Implementation

[0018] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0019] Example

[0020] Figure 1 is a flowchart of a method for identifying intermittent faults in an inverter system based on distribution characteristics according to the present invention.

[0021] In this embodiment, as shown in Figure 1, the method for identifying intermittent faults in an inverter system based on distribution characteristics according to the present invention includes the following steps:

[0022] (1) Data acquisition and processing;

[0023] (1.1) The three-phase voltage signals of the inverter system under normal operation are collected at equal intervals of 1 second. Let U be the single-phase voltage signal collected at the i-th moment. i,Θ Θ represents any one of the three phases of the inverter system, i = 1, 2, ..., N, and N represents the number of data collections. In this embodiment, N = 4000.

[0024] (1.2) Extracting the single-phase voltage signal U i,Θ Feature data e i,Θ This includes frequency, mean, peak-to-peak value, RMS value, margin, and kurtosis;

[0025] (1.3) The feature data extracted from the three single-phase voltage signals are used to form a feature dataset E = {e1,…,e...} i ,…e N}, e i This represents the feature data extracted at the i-th time. In this embodiment, the three-phase voltage signal includes three single-phase voltage signals. Six types of feature data are extracted from each single-phase voltage signal. Therefore, the feature data at one time contains a total of 18 feature quantities.

[0026] (2) Construct a weak detector and train it offline;

[0027] Constructing principal component analysis (PCA), dynamic principal component analysis (DPCA), kernel principal component analysis (KPCA), sliding window average principal component analysis (MAPCA), exponentially weighted sliding window average principal component analysis (EWMAPCA), and most weighted sliding window average T... 2 The control graph method OWMA-TCC uses six basic detectors, and then each of the six basic detectors is trained offline until convergence.

[0028] (3) Sliding window analysis;

[0029] (3.1) Extract the feature data e from the feature dataset E at each time step. i The inputs are fed into six basic detectors trained offline, and T is output through PCA, DPCA, MAPCA, KPCA, and EWMAPCA respectively. 2 The two statistics, Q and T, are used to output a T value through OWMA-TCC. 2The statistics are used to form a feature vector u from 11 extracted statistics at each time step. i ={f1(e i ),f2(e i ),...,f 11 (e i )};

[0030] (3.2) The feature vectors extracted at N sampling times are concatenated in the column dimension to form a feature integration matrix U;

[0031]

[0032] (3.3) Normalize the feature matrix U by first calculating the mean of each column of U, where the mean of the k-th column of U is denoted as μ. k k = 1, 2, ..., 11;

[0033] The mean of each column u k Form a row matrix

[0034] Then, U is normalized to a matrix with a mean of 0 and a variance of 1 using the following formula.

[0035]

[0036] Where the superscript T indicates transpose, 1 ω Let Σ be an identity matrix of length ω with elements of 1, and let Σ be the covariance matrix of U. * This indicates the inverse transformation of Σ;

[0037] remember The normalized elements in are Then matrix Represented as:

[0038]

[0039] (3.4) Set a sliding window with a size of ω×11 and a sliding step of 1, where the width ω ~ [12, N);

[0040] In this embodiment, assuming the sliding window is 45×11, the sliding window stops when it reaches the (N-44th)th row during the top-to-bottom sliding process.

[0041] (3.5) Starting with the first element f1(e1) of the unnormalized feature matrix U, slide the matrix U from top to bottom using a sliding window with a sliding step, and record the matrix after each slide. Let U be the sliding window matrix after the j-th slide. j ;

[0042] (3.6) Normalize each sliding window matrix;

[0043] For any sliding window matrix U j First calculate U j The mean of each column of elements in the array, where U j The mean of the elements in the k-th column is denoted as μ. j,k k = 1, 2, ..., 11;

[0044] The mean of each column u j,k Form a row matrix

[0045] Then, use the following formula to apply U j Normalized to a matrix with a mean of 0 and a variance of 1

[0046]

[0047] Where, Σ j It's U j The covariance matrix, (Σ j ) * Expressing the search for Σ j inverse transform;

[0048] remember The normalized element in is Then matrix Represented as:

[0049]

[0050] (3.7) For each normalized matrix Perform differential measurement factor calculation;

[0051] (3.7.1) First, examine the matrix. For each column of statistics, perform kernel density distribution estimation, where the estimate for the k-th column is:

[0052]

[0053] Among them, g k(x) represents the estimated probability density distribution of the k-th column, x is the variable corresponding to the required estimated probability density, h is a constant, and K(·) represents the Gaussian kernel function. In this embodiment, h is set to 1.

[0054] (3.7.2) Then, for the matrix For each column of statistics, perform kernel density distribution estimation, where the estimate for the k-th column is:

[0055]

[0056] Among them, g k (x) represents the estimated probability density distribution of the k-th column, x is the variable corresponding to the required estimated probability density, h is a constant, and K(·) represents the Gaussian kernel function. In this embodiment, h is set to 1.

[0057] (3.7.3) Calculate the matrix The probability density distribution and matrix of each column The KL divergence between the probability density distributions of each column, where the KL divergence between the probability density distributions of the k-th column is calculated as follows:

[0058]

[0059] Where X represents the range of values ​​for variable x;

[0060] (3.7.4) Sum the KL divergence of each column and use it as a measure of dissimilarity θ. j :

[0061]

[0062] (3.8) For the matrix Perform singular value decomposition to obtain the singular value vector [σ]. j,1 ,...,σ j,k ,...,σ j,11 ], where σ j,k It is the kth singular value obtained from the decomposition;

[0063] (3.9) Calculate the singular value factor d j :

[0064]

[0065] in, It is the mean of the k-th element of the singular value vector, std. k It is the standard deviation of the k-th element of the singular value vector;

[0066] (3.10) Based on the singular value factor d j and the difference measure θ j Calculate fault monitoring index D j :

[0067]

[0068] (3.11) Set control limits;

[0069] (3.11.1) Regarding fault monitoring index D j Perform kernel density estimation:

[0070]

[0071] Where g(x) is the estimated probability density, x is the variable corresponding to the required estimated probability density, h is a constant, and K(·) represents the Gaussian kernel function. In this embodiment, h is set to 1.

[0072] (3.11.2) Set the control limit Δ, which satisfies:

[0073]

[0074] Wherein, α is the set confidence level, and in this embodiment, α is 99%;

[0075] (4) Online status monitoring of the inverter system;

[0076] (4.1) According to step (1), collect the characteristic data e of the inverter system at time N+1. N+1 , will e N+1 The inputs are fed into six basic detectors to extract the feature vector u at time N+1. N+1 ={f1(e N+1 ),f2(e N+1 ),…,f 11 (e N+1 )};

[0077] Then add it to the sliding window matrix U. j The last line and the sliding window matrix U are deleted. j The first row yields the sliding window matrix U at time N+1. N+1 ;

[0078] (4.2) Repeat steps (3.5) to (3.10) to calculate the fault monitoring index D at time N+1. N+1 ;

[0079] (4.3) Determine the state of the inverter system at time N+1 based on the set control limit Δ. If D N+1 If the value is greater than Δ, it indicates that the inverter system is experiencing an intermittent fault at this moment; otherwise, it is considered that the inverter system is operating normally.

[0080] Experimental simulation

[0081] The platform used for verification in this invention consists of four parts: a DC power supply, an inverter circuit system, electronic loads, and a data acquisition module. The DC power supply is controlled at 80V. The IGBTs in the inverter circuit system are Semikron SKM100GAL12T4 modules with a 2SC0435T2F1-17 IGBT driver core and a 2BB043 adapter board. The three electronic loads are set to AC constant resistance mode with a resistance of 50 ohms. The implanted fault is located in the filter inductor at the output of the inverter circuit system, with a resistor connected in parallel to simulate intermittent inductor degradation. In this experiment, 4000 normal data points were collected to train the model. The test data consists of 4000 samples, with parallel resistances of 500Ω for samples 1000-2000 and 3300-3600, 1000Ω for samples 2400-3000, and the remainder being normal data. The fault detection rate is calculated based on the number of samples exceeding the control limit during the fault duration. Then, the false alarm rate is obtained by using the number of falsely identified normal samples outside the fault duration.

[0082] In this experiment, Principal Component Analysis (PCA), Dynamic Principal Component Analysis (DPCA), Kernel Principal Component Analysis (KPCA), Moving Average Principal Component Analysis (MAPCA), Exponentially Weighted Moving Average Principal Component Analysis (EWMAPCA), and Optimal Weighted Moving Average Principal Component Analysis (TPA) were applied. 2 Control chart (Optimal weighted moving average T) 2 The performance of the control chart (OWMA-TCC) and the FEDD of this invention was verified in an inverter system. The CPV of the PCA-based method was 80%, and the delay of DPCA was 1. The kernel function of KPCA was a radial basis function with a parameter of 0.000016. The moving window length of MAPCA was 7, and the forgetting factor was 0.7. The window width of OWMA-TCC was set to 8, and the window width of FEDD was set to 15. The significance level of each detector was set to 1%, and the control limits were calculated by KDE.

[0083] Table 1 lists the fault detection rates of the verified methods for intermittent inductor degradation faults in inverter systems. Among them, PCA, DPCA, and KPCA showed unsatisfactory detection results. For PCA, the false alarm rate obtained from the T2 statistic monitoring was 0.29%, and the fault detection rate was only 1.32%, failing to achieve effective fault detection. Furthermore, the Q statistic of PCA, with a false alarm rate of 0.38%, only achieved a final fault detection rate of 0.84%, which is also not an effective detection indicator. Similarly, the T2 and Q statistics of DPCA and KPCA also failed to achieve effective fault detection, with the fault detection rates of the above methods not exceeding 2%. Due to the introduction of the sliding window technique, the fault detection rates of MAPCA, EWMAPCA, and OWMA-TCC methods were significantly improved. OWMA-TCC, as the best performing intermittent fault detection model, could detect 84.37% of intermittent faults. The integrated detection method proposed in this invention achieved a fault detection rate of 93.68% with a false alarm rate of 2.71%, which is 9.31% higher than the best existing method. The above results demonstrate the superiority and effectiveness of the proposed method.

[0084] Table 1 shows the fault detection rate (%) of the verification method for intermittent inductor degradation faults in inverter systems;

[0085]

[0086] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for identifying intermittent faults in an inverter system based on distribution characteristics, characterized in that, The steps include: (1) collecting and processing the three-phase voltage signals of the inverter system under normal operation at equal intervals; (2) constructing a weak detector and training it offline; (3) setting the control limit for detecting intermittent faults in the inverter system through sliding window analysis and processing; (4) collecting the characteristic data of the inverter system in real time, calculating the fault monitoring index at the current moment, and then performing real-time online detection of intermittent faults in the inverter system according to the set control limit; wherein, the method for collecting and processing the three-phase voltage signals in step (1) is as follows: (1.1) collecting the three-phase voltage signals of the inverter system under normal operation at equal intervals, wherein, let the first phase be denoted as the third phase. The single-phase voltage signal collected at each moment is ， This represents any one of the three phases in the inverter system. ， Indicates the number of data acquisitions; (1.2) Extract single-phase voltage signals. Feature data in Including frequency, mean, peak-to-peak value, RMS value, margin, and kurtosis; (1.3) The feature data extracted from the three single-phase voltage signals are combined into a feature dataset. ， Indicates the first Feature data extracted at each time point; wherein, the process of setting the control limit in step (3) is as follows: (3.1) The feature dataset is... Feature data at various times The inputs are fed into six basic detectors trained offline, and T is output through PCA, DPCA, MAPCA, KPCA, and EWMAPCA respectively. 2 The two statistics, Q and T, are used to output a T value through OWMA-TCC. 2 The statistics are used to construct a feature vector from the 11 statistics extracted at each time point. (3.2) will The feature vectors extracted at each sampling time are concatenated along the column dimension to form a feature integration matrix. ； (3.3) Regarding the characteristic matrix Perform normalization processing, first calculate The mean of each column of elements, where, The The mean of the column elements is denoted as ， ; calculate the mean of each column Form a row matrix ， Then, follow the formula below to... Normalized to a matrix with a mean of 0 and a variance of 1 ； ; where superscript Indicates transpose. This indicates that the element is 1 and the length is 1. The identity matrix, yes The covariance matrix, Expressing the request Inverse transform; denoted The normalized elements in are Then the matrix Represented as: (3.4) Set a sliding window with a window size of [size missing]. The sliding step is 1, where the width is... (3.5) Using the unnormalized characteristic matrix The first element Starting from the matrix, the sliding window is used to adjust the matrix according to the sliding step. First, slide from top to bottom, recording the matrix corresponding to each slide, where is denoted as the first . The sliding window matrix after the next slide is: (3.6) Normalize each sliding window matrix; for any sliding window matrix First calculate The mean of each column of elements, where, The The mean of the column elements is denoted as ， ; calculate the mean of each column Form a row matrix ， Then, follow the formula below to... Normalized to a matrix with a mean of 0 and a variance of 1 ； ;in, yes The covariance matrix, Expressing the request Inverse transform; denoted The normalized element in is Then the matrix Represented as: (3.7) For each normalized matrix Perform differential measurement factor calculation; (3.7.1) First, perform matrix... For each column of statistics, kernel density distribution estimation is performed, where the th column is... The column estimate is: ;in, To estimate the first Column probability density distribution, For the variable corresponding to the required estimated probability density, To set a constant, Represents the Gaussian kernel function; (3.7.2), then for the matrix For each column of statistics, kernel density distribution estimation is performed, where the th column is... The column estimate is: ;in, To estimate the first Column probability density distribution, For the variable corresponding to the required estimated probability density, To set a constant, Represents the Gaussian kernel function; (3.7.3), calculate the matrix. The probability density distribution and matrix of each column The KL divergence between the probability density distributions of each column, where the th column... The KL divergence between probability density distributions is calculated as follows: ;in, Representative variable The range of values; (3.7.4), sum the KL divergence of each column and use it as a measure of dissimilarity. ： (3.8) For the matrix Perform singular value decomposition to obtain singular value vectors. ,in, The decomposition yielded the first (3.9) Calculate the singular value factor. ： ;in, It is the singular value vector of the first The mean of the elements. It is the singular value vector of the first The standard deviation of each element; (3.10), based on the singular value factor and difference measure factor Calculate fault monitoring indicators ： (3.11) Set control limits; (3.11.1) Monitor fault indicators Perform kernel density estimation: ;in, To estimate the resulting probability density, For the variable corresponding to the required estimated probability density, To set a constant, Represents the Gaussian kernel function; (3.11.2), set control limits. Control Limit satisfy: ;in, The confidence level to be set.

2. The method for identifying intermittent faults in an inverter system based on distribution characteristics according to claim 1, characterized in that, The weak detectors include: Principal Component Analysis (PCA), Dynamic Principal Component Analysis (DPCA), Kernel Principal Component Analysis (KPCA), Moving Window Average Principal Component Analysis (MAPCA), Exponentially Weighted Moving Window Average Principal Component Analysis (EWMAPCA), and Most Weighted Moving Window Average Principal Component Analysis (TPA). 2 The OWMA-TCC control chart method uses six basic detectors.

3. The method for identifying intermittent faults in an inverter system based on distribution characteristics according to claim 1, characterized in that, The method for real-time online detection of intermittent faults in the inverter system is as follows: (S1) According to step (1), the inverter system data is collected in the first... Feature data of time ,Will The data is input into six basic detectors respectively, and the first one is extracted. Feature vector at time step Then add it to the sliding window matrix. The last line and the sliding window matrix are deleted. The first line, get Sliding window matrix of time (S2) By adjusting the sliding window matrix Perform sliding window analysis to calculate the first... Fault monitoring indicators at all times (S3) According to the set control limits Determine the inverter system in the first... The state at any given moment, if If the inverter system experiences an intermittent fault at this moment, it is determined that the inverter system is operating normally; otherwise, it is considered that the inverter system is operating normally.

Images