Photovoltaic DC system fault detection method based on time-frequency graph texture feature extraction

Abstract

The invention discloses a photovoltaic direct current system fault detection method based on time-frequency graph texture feature extraction. According to the method, time-frequency point space distribution characteristics of arc fault signals are analyzed through a time-frequency diagram, and a detection method based on texture analysis is provided. The method comprises the following steps: firstly, inhibiting spectrum outlier noise by adopting a sliding quartile distance frequency domain filtering algorithm, and determining a fault characteristic frequency band by combining a spectrum minimum point and frequency band energy; secondly, extracting time-frequency graph texture features of a feature frequency band by using two-dimensional constant false alarm rate detection and a gray-level co-occurrence matrix, and constructing a time-frequency feature quantization system; and finally, inputting the extracted time-frequency characteristics into an extreme learning machine for decision judgment. The method can be widely applied to arc fault detection in different voltage scenes, solves the problems of noise interference, misjudgment, missing detection and the like, and improves the reliability of fault detection.

Description

Technical Field

[0001] The present application relates to the technical field of power system fault detection, and particularly to a photovoltaic system DC fault arc detection method based on time-frequency graph texture feature extraction. Background Art

[0002] In recent years, with the rapid development of DC power systems such as photovoltaic power generation, electric vehicles, and DC microgrids, DC fault arc detection technology has become one of the key challenges to ensure the safe operation of the system. DC fault arcs are usually caused by factors such as line aging, insulation deterioration, or poor contact. They have characteristics such as strong nonlinearity, instantaneousness, and randomness, and the arc signal is easily submerged in complex environmental noise, resulting in problems of insufficient sensitivity and high misjudgment rate in traditional detection methods.

[0003] The technical means for DC fault arc detection are mainly divided into methods based on physical characteristics, methods based on time-frequency analysis, and detection methods based on deep learning. Methods based on physical characteristics, such as acoustic signal and electromagnetic radiation detection, detect through the acoustic fingerprint characteristics or electromagnetic field disturbances of the arc, but are greatly affected by environmental interference and have high requirements for the installation position of sensors, with limited applicability. Methods based on time-frequency analysis, such as short-time Fourier transform and wavelet transform, extract time-frequency domain features and combine machine learning for classification. However, the dimension of time-frequency feature extraction is high, and directly inputting into a neural network will lead to a sharp increase in computational complexity, making it difficult to run in real time in scenarios with limited hardware resources. Under high-voltage conditions, the signal-to-noise ratio of the arc signal is low, and traditional time-frequency analysis methods are difficult to effectively distinguish noise from weak fault features, and high-frequency harmonic interference introduced by transient processes such as inverter startup and shutdown is prone to cause misjudgment. End-to-end detection methods based on deep learning, such as convolutional neural networks, directly process time-frequency graphs. Although some studies have achieved high accuracy, the model has a large number of parameters, relies on a large amount of labeled data, and has insufficient generalization ability for weak arcs. At present, there is an urgent need for a detection method that can accurately extract fault features and suppress noise interference to improve the reliability, real-time performance, and adaptability of fault diagnosis in high-noise and multi-scenario environments. Summary of the Invention

[0004] The purpose of the present application is to provide a fault detection method for a photovoltaic DC system based on time-frequency graph texture feature extraction. By sliding filtering to suppress spectral outliers, determine the fault feature frequency band, and extract the time-frequency features of the feature frequency band according to the non-stationarity of the fault, the fault detection of the photovoltaic DC system under different voltage levels is realized, and the accuracy and robustness of fault detection are improved.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions: A photovoltaic DC system fault detection method based on time-frequency graph texture feature extraction, characterized by including the following steps:

[0006] S1: Build an experimental platform for collecting fault signals. Simulate the power generation and power consumption processes with two collection scenarios: connecting a photovoltaic panel to an inverter and connecting a DC power supply to a resistor. The collection frequency is f0;

[0007] S2: Perform a fast Fourier transform (FFT) on the collected signals to generate the single-sided spectrum |X(f)| of the signals. Perform a sliding interquartile range filter on the spectrum to obtain the filtered spectrum X′(f);

[0008] S3: Perform Gaussian filtering and smoothing on X'(f) to obtain the spectrum envelope. Use the minimum value points of the spectrum envelope as the boundaries for spectrum division, and take the envelope interval with the maximum energy as the characteristic frequency band f arc ;

[0009] S4: Perform a short-time Fourier transform (STFT) on the collected signals to generate the time-frequency spectrum Y of the signals. Determine the characteristic time-frequency spectrum Y arc according to the characteristic frequency band f arc , and then perform two-dimensional constant false alarm detection to obtain the binary characteristic time-frequency diagram Y arcb ;

[0010] S5: Extract the time-frequency features of Y arcb , determine the direction θ and distance d of the gray-level co-occurrence matrix, and extract the energy, entropy value, contrast, and correlation features in different directions of Y arcb ;

[0011] S6: Input the extracted features into an extreme learning machine (ELM) neural network, and perform fault detection after training into a model.

[0012] Further, the specific content of step S1 is: Build a DC series fault arc experimental platform including a photovoltaic panel, a DC power supply, an inverter, an arc generator, a load resistor, and a current transformer. The current transformer is connected to the resistor to obtain voltage signals in real time from the sampling point of the photovoltaic DC system. Simulate different load conditions by adjusting the load resistor and the inverter. The collection frequency is f0.

[0013] Further, the specific content of step S2 is: In the optional embodiment S2 of this application, perform an FFT transform on the collected signal x(t) to generate the single-sided spectrum X(f).

[0014] Determine the window function for sliding filtering. The window function W k at the frequency point f k The interception range at:

[0015]

[0016] where L is the window length, f k is the current center frequency point, and calculate the lower quartile Q1 and upper quartile Q3 of the spectrum data within the window:

[0017]

[0018] Construct a frequency-varying determination boundary:

[0019] B(f) = [Q1 - kI QR , Q3 + kI QR

[0020] where k is a threshold coefficient, and I QR is the interquartile range, and the calculation formula is:

[0021] I QR = Q3 - Q1

[0022] Replace the outliers with the window median to better retain the characteristics of the non-outliers within the window, and obtain the filtered spectrum X'(f).

[0023] Furthermore, the specific steps of step S3 are as follows: For the filtered spectrum X′(f) generated in S2, perform Gaussian filtering and smoothing:

[0024]

[0025]

[0026] where S(f i - f) is the Gaussian kernel function, and σ is the standard deviation of the Gaussian function, which controls the smoothing degree of the Gaussian filter. Take the minimum point of the envelope after Gaussian filtering as the frequency band division boundary. Calculate the energy within each frequency band:

[0027]

[0028] where E j is the energy of the jth frequency band, Δf is the frequency resolution, f k is the discrete frequency point, and take the envelope interval with the maximum energy as the characteristic frequency band f arc ;

[0029] Furthermore, the specific steps of step S4 are as follows: Perform STFT transformation on the collected signal, use a Hanning window to weight the signal segments, set the window length to N win , set the length of the overlapping part of the window to N overlap , the number of points for the fast Fourier transform is N FFT , generate the time-frequency spectrum Y of the signal, determine the characteristic time-frequency spectrum Y arc according to the characteristic frequency band f arc , calculate the mean μ of Y arc , and obtain the standard deviation σ0 of the noise envelope:

[0030] ​

[0031] Define the standardized variable:

[0032]

[0033] where |z m,n | represents the envelope amplitude at the m-th moment and n-th frequency point in the time-frequency plane. The standardized noise envelope probability density function can be expressed as:

[0034]

[0035] When U0 is used as the detection threshold, the false alarm probability of a single detection is:

[0036]

[0037] Find U0 from the given false alarm probability to obtain the absolute detection threshold η:

[0038] η = U0σ0

[0039] Set the time-frequency points with amplitudes greater than the detection threshold to 1 and those with amplitudes less than the detection threshold to 0 to obtain the binary feature time-frequency diagram Y arcb .

[0040] Furthermore, the specific step S5 is as follows: Y arcb The image size is M×N, the gray level is 2, and the GLCM is a matrix. The matrix element C(i,j) represents the frequency of simultaneous occurrence of pixels with gray value i and pixels with gray value j in a specific direction θ and distance d. The calculation formula is as follows:

[0041] C (θ,d) (i,j) = |{(x,y),(x + Δx,y + Δy)|f(x,y) = i, f(x + Δx,y + Δy) = j}|

[0042] where (x,y) is the coordinate of a certain pixel point in the image, f(x,y) is the gray value of the pixel point (x,y), and the value range is 1, 2. Δx and Δy are the spacings of two pixel points on the X-axis and Y-axis.

[0043] The energy is as follows:

[0044]

[0045] The entropy is as follows:

[0046]

[0047] The contrast is as follows:

[0048]

[0049] The correlation is as follows:

[0050]

[0051] Further, step S6 is specifically as follows: randomly initialize the connection weight matrix W from the input layer to the hidden layer and the bias vector b of the hidden layer neurons, and calculate the hidden layer output matrix H:

[0052] H = g(W·X + b)

[0053] where X is the input feature matrix and g(·) is the activation function. Solve the output weight β through the Moore - Penrose generalized inverse:

[0054] β = H + ·T

[0055] where T is the target output matrix and H + is the generalized inverse of the hidden layer output matrix. Calculate the network output Z:

[0056] Z = H·B

[0057] The beneficial effects of the present invention are as follows:

[0058] (1) This application suppresses spectral outlier noise through sliding interquartile range (IQR) frequency - domain filtering and optimizes frequency band division by combining Gaussian smoothing, effectively enhancing the recognizability of weak fault features and being applicable to the determination of fault feature frequency bands in multiple scenarios.

[0059] (2) This application extracts time - frequency map texture features, compresses high - dimensional time - frequency data into low - dimensional feature vectors. Compared with the traditional method of using neural networks to train time - frequency maps, it can save computing resources, improve detection efficiency, and is easy to deploy at the edge. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, so they should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0061] Figure 1 A photovoltaic DC system fault detection method based on time - frequency map texture feature extraction provided by an embodiment of the present application;

[0062] Figure 2 A flowchart for suppressing signal line - spectrum frequencies provided by an embodiment of the present application;

[0063] Figure 3Schematic diagram of the extreme learning machine neural network structure provided by the embodiments of the present application. Detailed implementation manners

[0064] Next, the technical solutions in the embodiments of the present application will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present application. Apparently, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0065] Figure 1 Flowchart of a photovoltaic DC system fault detection method based on time-frequency diagram texture feature extraction provided by the present application, specifically including the following steps:

[0066] Step S101: Input the fault arc signal collected by the mutual inductor.

[0067] As an embodiment, the mutual inductor is connected to both ends of the resistor, and the voltage signal is collected in real time from the sampling point of the photovoltaic system, and the sampling frequency is f0.

[0068] Step S102: Perform FFT transformation and frequency-domain sliding filtering on the collected signal.

[0069] As an embodiment, determine the filter window length, based on the center frequency point of the window, extract the spectral data within the corresponding frequency band, calculate its upper and lower quartiles Q1 and Q3, construct a frequency-varying decision boundary, and replace the outliers outside the boundary with the window median to obtain the filtered spectrum.

[0070] Step S103: Perform Gaussian smoothing and determine the characteristic frequency band.

[0071] As an embodiment, perform Gaussian smoothing on the filtered spectrum to obtain the spectral envelope. Take the minimum value point of the envelope as the frequency band division boundary, and take the frequency band with the largest energy as the characteristic frequency band.

[0072] Step S104: Extract the characteristic time-frequency spectrum and perform binarization processing.

[0073] As an embodiment, extract the characteristic time-frequency spectrum according to the characteristic frequency band, calculate the standard deviation σ0 and the detection threshold η according to the expected false alarm probability P f , set the points with amplitudes greater than the threshold to 1 and those less than the threshold to 0, and generate the binary image Y arcb .

[0074] Step S105: Extract the texture features of the time-frequency diagram.

[0075] As an embodiment, construct a gray-level co-occurrence matrix for the binary time-frequency diagram, and calculate four texture features of its energy, entropy, contrast, and correlation.

[0076] Step S106: Use an extreme learning machine for fault identification.

[0077] As an embodiment, the extracted four-dimensional texture feature vector is input into the extreme learning machine neural network for training and classification recognition.

[0078] Step S107, output the recognition result for DC system arc fault detection and warning.

[0079] Furthermore, the experimental platform for signal acquisition in step S101 consists of a photovoltaic panel, a DC power supply, an inverter, an arc generator, a load resistor, and a current transformer, and the voltage signal of the acquisition system is collected with a sampling frequency of 100 kHz.

[0080] Furthermore, the number of spectrum points in step S102 is 50k, the window length L of the moving interquartile range filter is 2500, the threshold coefficient k is 3, and a frequency-varying decision boundary is constructed:

[0081] B(f) = [Q1 - 3I QR , Q3 + 3I QR

[0082] Replace the outliers outside the boundary with the window median to obtain the filtered spectrum X′(f).

[0083] Furthermore, in step S103, Gaussian filtering is performed on X'(f), the window length of the Gaussian filtering is 250, and the standard deviation is 1500, to obtain the filtered spectrum:

[0084]

[0085] Determine the boundary of spectrum division through the minimum points, and calculate the energy values of each frequency band:

[0086]

[0087] Take the envelope interval with the maximum energy as the characteristic frequency band f arc .

[0088] Furthermore, in step S104, short-time Fourier transform is performed on the acquired signal, weighted with a Hanning window, the window length is set to 512, the length of the overlapping part of the window is 256, the number of fast Fourier transform points is 1024, and according to the characteristic frequency band f arc Intercept the characteristic time-frequency spectrum Y arc , the expected false alarm probability P f is 0.1, obtain the standard deviation σ0 and the detection threshold η based on the expected false alarm probability, and obtain the binary image Y arcb .

[0089] Further, the relationship between the direction θ and the distance d in the step S105 is as Figure 2 shown. When the vertical and horizontal stripe directions of the characteristic time-frequency spectrum have certain rules, d = 1, θ = 0°, and d = 1, θ = 90° are selected, and the gray level is 2. The gray-level co-occurrence matrix is used to analyze the texture information of the binary characteristic time-frequency diagram, and four texture features of energy, entropy, contrast, and correlation are extracted.

[0090] Further, the structure of the extreme learning machine in the step S106 is as Figure 3 . The connection weights between the input layer and the hidden layer and the thresholds of the hidden layer neurons are randomly initialized, and the weight matrix from the hidden layer to the output layer is calculated by analytical method. The eight texture features of energy, entropy, contrast, and correlation in two directions are used as the inputs of the extreme learning machine neural network. The input of the normal signal is mapped to 1, and the input of the fault signal is mapped to -1. The signal state is judged according to the value of the output matrix Z. If it is greater than 0, it is normal; if it is less than 0, it is a fault.

[0091] Finally, it should be noted that the above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A DC fault arc detection method for a photovoltaic system based on the extraction of time-frequency diagram texture features, characterized in that, The method includes: Constructing a multi-voltage fault signal acquisition experimental platform to collect fault arc signal data under different voltages; Performing fast Fourier transform (FFT) and short-time Fourier transform (STFT) on the fault arc signal to generate the single-sided spectrum and time-frequency diagram of the signal; Performing frequency-domain filtering on the single-sided spectrum using the sliding interquartile range (IQR) to suppress spectrum outliers and determine the characteristic frequency band; Performing two-dimensional constant false alarm rate (CFAR) detection on the time-frequency diagram to remove background noise and generate a binary time-frequency diagram; Performing gray-level co-occurrence matrix (GLCM) analysis on the binary time-frequency diagram to extract texture features; Inputting the extracted texture features into an extreme learning machine (ELM) for fault arc detection decision-making.

2. The method according to claim 1, wherein In the sliding interquartile range filtering, the frequency-varying decision boundary is constructed using the interquartile range of the data within the frequency band window. The spectrum points exceeding this boundary are regarded as outliers and replaced with the median within the window.

3. The method according to claim 1, wherein The specific determination of the characteristic frequency band includes: Performing Gaussian smoothing on the filtered spectrum, taking the minimum value point of the smoothed spectrum as the frequency band boundary, calculating the total energy of each frequency band, and selecting the frequency band with the largest energy as the characteristic frequency band.

4. The method according to claim 3, wherein The steps of the two-dimensional constant false alarm rate (CFAR) detection include: Calculating the standard deviation of the noise envelope and defining the standardized variable; Determining the detection threshold according to the given false alarm probability, setting the time-frequency points with amplitudes greater than the threshold to 1 and those less than the threshold to 0 to generate a binary time-frequency diagram.

5. The method according to claim 1, wherein The implementation steps of the gray-level co-occurrence matrix (GLCM) analysis are as follows: Setting the direction and distance parameters and calculating the gray-level co-occurrence matrix of the binary time-frequency diagram; Extracting texture features, including contrast, energy, entropy, and correlation.

6. The method according to claim 1, wherein The extreme learning machine (ELM) is used for classification decision-making and outputs the result of whether there is a fault arc after training.

7. The method according to any one of claims 1-6, characterized in that, It is applicable to fault arc detection in the scenarios of a photovoltaic system connected to an inverter load and a DC power supply with different voltage levels connected to a resistive load.

Images