Multi-feature fusion DC fault arc detection method based on compressed sensing

Abstract

The invention belongs to the technical field of direct-current arc fault detection, and particularly discloses a multi-feature fusion direct-current arc fault detection method based on compressed sensing. The problem of mass data processing and storage caused by direct current series fault arc real-time monitoring and the problem of high misjudgment rate under the condition of single monitoring arc signal are solved. According to the method, compressed sensing domain data is obtained from mass data by using a compressed sensing technology, the data volume of storage and processing of arc signals is greatly reduced, arc sound signals are extracted when the arc occurs and are combined with current signals, and fault arc monitoring based on multi-feature fusion is realized. The BP neural network is combined for decision making, the accuracy rate reaches up to 99%, and normal signals and fault arc signals can be effectively distinguished. On the basis of the detection method, an arc monitoring interface is designed, and arc signals and sound signals can be obtained in real time through the monitoring interface, so that arc monitoring is more real-time and convenient.

Description

Technical Field

[0001] The invention relates to a novel DC fault arc detection method, belonging to the technical field of DC arc fault detection. Background Art

[0002] As the global demand for renewable energy increases, new energy sources such as solar energy have become an important part of the building energy structure. The output of the photovoltaic power generation system is direct current. This feature has promoted the transformation of building electricity consumption from traditional alternating current to direct current. At present, there are three main detection methods for DC fault arcs by scholars at home and abroad: a method that uses physical characteristics such as sound, light, heat and electromagnetic radiation when the arc is generated; a method that uses time-frequency domain feature quantities to select threshold judgment; and a method that uses artificial intelligence technology to detect fault arcs. However, in the process of arc signal detection, there are problems of high sampling rate and large amount of measurement point data, which increase the storage and calculation burden of the computer, and single condition judgment may result in missed judgments and misjudgments, resulting in a high error rate in arc fault detection. Summary of the invention

[0003] To address the above problems, a DC fault arc detection method based on compressed sensing and multi-feature fusion is proposed, which can effectively avoid the above problems and make the results more accurate.

[0004] Through compressed sensing technology, high-dimensional signals are converted into low-dimensional signals, which achieves data simplification. Combined with the sound signal when the arc occurs, the accuracy of arc detection is improved.

[0005] The novel fault arc detection method comprises the following steps:

[0006] Step 1: Use an oscilloscope to collect the DC series current signal in the circuit;

[0007] The present invention builds a DC arc fault test platform with reference to UL1699B standard, which mainly includes a DC power supply, an arc generator, a data acquisition device and a DC load. The current signal in the circuit is collected by an oscilloscope, and the subsequent signal analysis and processing has been carried out.

[0008] Step 2: Using the compressed sensing method on the collected current signal, using the sensing matrix to obtain data in the compressed sensing domain from the massive data;

[0009] Y=ΦΨx (1)

[0010] Among them, Y is the perception signal, Φ is the observation matrix, Ψ is the sparse matrix, and x is the original signal;

[0011] After the oscilloscope collects the current signal y, it performs compressed sensing sampling on it. Reconstruction error and Pearson correlation coefficient are two commonly used evaluation indicators of CS, which are used to analyze the degree of effective information retention of the original signal after compressed sensing.

[0012]

[0013]

[0014] Where x is the original signal, y is the reconstructed signal, and N is the dimension of the original signal;

[0015] According to two evaluation indicators, it is determined that the measurement matrix of compressed sensing is a Gaussian measurement matrix and the compression rate is 85%.

[0016] Step 3: The signal after compressed sensing is random and disordered. It is impossible to directly distinguish whether an arc fault occurs in the system by relying solely on the current signal after compressed sensing. The misjudgment rate in the DC system is relatively high. Therefore, feature extraction is performed on the sensed arc signal.

[0017] The perceived signal is analyzed from the perspectives of time domain and energy. The time domain includes four aspects: standard deviation, peak value, root mean square, and root square amplitude. The energy value of the signal in the time window is extracted. Five characteristic values ​​are extracted from the perceived signal, and a characteristic data sample is formed with five characteristic values ​​as a group.

[0018] Standard deviation s1:

[0019]

[0020] Peak s2:

[0021] s2=max|Y(n)| (5)

[0022] RMS value s3:

[0023]

[0024] Square root amplitude s4:

[0025]

[0026] Energy s5:

[0027]

[0028] Among them, Y(n) is the perception signal, and M is the dimension of the perception signal.

[0029] Step 4: Analyze the collected arc sound signals, and extract the sound signals when there is no arc and the sound signals when an arc occurs;

[0030] By performing frequency, spectrogram, and short-time energy analysis on the collected signals and using wavelet packet decomposition to check the energy differences of each sub-band, the interference of fans, talking and other noises can be removed, and the sound signals when there is no arc and the sound signals when an arc occurs can be obtained.

[0031] Step 5: Use the compressed sensing method on the collected sound signal and use the perception matrix to obtain data in the compressed sensing domain from the massive data;

[0032] The compressed sensing algorithm is also used, the measurement matrix of the compressed sensing is a Gaussian measurement matrix and the compression rate is 85%.

[0033] Step 6: Perform feature analysis on the collected sound perception signals;

[0034] volume:

[0035] energy:

[0036] Amplitude: a3 = max|Y(i)| (11)

[0037] RMS value:

[0038] Y(i) is the perception signal, and M is the dimension of the perception signal.

[0039] Step 7: Perform feature fusion on the collected feature quantities, that is, form a feature data sample with 9 feature values ​​as a group.

[0040] Step 8: Use BP neural network to make decisions on detection features, and obtain a detection accuracy of up to 99%;

[0041] The experiment selected two voltage levels of 60V and 90V DC power supplies, and the loads were resistors, household LED lamps, and 12V DC fans. 1500 sets of data under normal system conditions and 3000 sets of data under arc fault conditions were selected to form a feature data training sample set. Then 253 sets of feature data under normal conditions and 253 sets of feature data under fault conditions under different loads were selected as test set data. Using BP neural network as the decision-making for detection features, a detection accuracy of up to 99% was obtained.

[0042] Step 9: Based on this detection method, an arc software monitoring system is designed. Through this monitoring system, arc signals and sound signals can be obtained, making arc monitoring more real-time and convenient.

[0043] The present invention has the following advantages:

[0044] First, applying the compressed sensing algorithm to arc signal extraction can significantly reduce the amount of data stored and processed by the arc signal while retaining the main features of the original signal.

[0045] Second, the multi-feature fusion fault arc detection method reduces the misjudgment rate caused by a single variable or other environmental noise interference, and more accurately detects the generation of DC fault arcs.

[0046] Third, improve the detection accuracy of arc faults and provide important reference ideas for arc fault detection.

[0047] The above description is only an overview of the present invention. In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the specific implementation method of the present invention is exemplified. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to enable the present invention to be better understood by users, the following drawings required for describing the embodiments are listed:

[0049] Figure 1 It is an overall flow chart of the technical solution of the present invention;

[0050] Figure 2 It is a DC fault arc data acquisition experimental platform;

[0051] Figure 3 is the original data waveform;

[0052] Figure 4 is the arc signal waveform under FFT basis;

[0053] Figure 5 It is the arc signal waveform of compressed sensing;

[0054] Figure 6 is the reconstructed arc signal waveform;

[0055] Figure 7 The complex sound signal collected;

[0056] Figure 8 is the time-frequency diagram of the sound signal;

[0057] Fig. 9 It is the short-time energy analysis of sound signal;

[0058] Fig.10 is the energy value of the eight sub-bands obtained after the three-layer wavelet packet decomposition of the sound signal;

[0059] Fig.11 Constructing feature datasets for arc perception signals and sound perception signals

[0060] Fig.12 , Fig.13 , Fig.14 It is the arc monitoring interface; DETAILED DESCRIPTION

[0061] The present invention will be further described in detail below in conjunction with the overall flow chart and accompanying drawings of the present invention. Based on the embodiments of the present invention, all other similar cases obtained by ordinary technicians in the field without making substantial innovations are within the scope of protection of the present invention.

[0062] Figure 1 The figure is the overall flow chart of the present invention. In the experimental scene, the current signal is collected by an oscilloscope, and then the original signal is compressed to obtain the perception signal; the sound signal is processed to extract the arc sound signal and the normal signal, and the useful sound signal is compressed to obtain the perception signal, and the processed signal is subjected to equal period extraction of features, and finally the BP neural network is used to make a decision to determine whether a fault arc is generated.

[0063] Figure 2 The DC fault arc data acquisition experimental platform built by the present invention according to the UL1699B standard uses a DC power supply as the DC source input of the experimental platform, connects the arc generator in series in the trunk line, simulates the normal operation and arcing in the line through the connection and disconnection state of the arc generator, and connects the load module in series or in parallel to the circuit to simulate the actual power consumption. The current signal in the loop is collected by an oscilloscope to obtain the normal signal and fault signal required for signal processing. The data of a certain time window is selected, the time length is 7ms, the data of each time window is 14000, and the data length after compressed sampling is 2100.

[0064] Figure 3 This is the current signal waveform obtained by the present invention under the experimental platform. Single load operation: resistance load, DC fan load, LED lamp load; multiple loads in parallel operation, resistance load and DC fan load, resistance load and LED lamp load, resistance load and DC fan load and LED lamp load. Each signal contains 14,000 points, which shows that the current fluctuates violently when the fault arc occurs.

[0065] Figure 4 The spectrum diagram of the arc signal collected by the present invention based on FFT is simulated by MATLAB software and the signal is transformed by FFT. It can be seen that the energy of the signal is mainly concentrated in the low-frequency part, the frequency of the arc signal is mainly concentrated in 0-100KHz, and the high-frequency part is small or zero, that is, the signal has frequency domain sparsity, which meets the theoretical basis of compressed sensing.

[0066] Figure 5In order to adopt the Gaussian observation matrix, the arc signal is compressed and sampled at a compression rate of 85%, that is, 2100 data points are sampled from the original 14000 data points. The sampled data is analyzed and processed.

[0067] Figure 6 In order to reconstruct the arc perception, the compressed sampling matching pursuit (CoSaMP) algorithm is used in the FFT sparse basis. The reconstructed signal is simulated by MATLAB software. The number of iterations of the CoSaMP reconstruction algorithm is set to 500. It can be seen from the reconstructed signal waveform that the reconstructed signal is similar to the original signal ( Figure 3 ) have great similarities.

[0068] Figure 7 The sound signals when an arc occurs and when there is no arc are obtained by the voice recorder in the present invention. Through the spectrum analysis of the arc sound signal, it can be concluded that the frequency band of the arc sound signal is mainly concentrated in 0-16KHz. First, a low-pass filter is used on the signal to eliminate the interference of high-frequency noise, and the wavelet noise reduction method is used to process the original signal to make the signal characteristics more prominent.

[0069] Figure 8 For the time-frequency analysis of the sound signal, the sound signal can be divided into three parts. The frequency range of area 1 and area 2 is mainly concentrated in 0-16KHz, the relative intensity variation range of area 1 is mainly concentrated in -20-20, and the relative intensity variation range of area 2 is -10-10; the frequency range of area 3 is mainly concentrated in 0-2KHz, and the relative intensity variation range is 0-10. It can be seen that area 1 and area 2 have the frequency domain characteristics of arc signals.

[0070] Fig. 9 In order to perform short-time energy analysis on the sound, the collected digital signal is framed so that each frame has N sampling points, the total number of frames is L1 = L / N, and the duration of each frame is t = T / L1. After energy calculation, the envelope waveform of the signal shows 6 peak waveforms, indicating that the energy of the sound signal has changed to a certain extent.

[0071] Fig.10 Sym8 wavelet is used to perform 3-layer wavelet packet decomposition on the 6 signals to obtain eight sub-bands, and the energy normalization value of each sub-band is calculated. It can be seen that signal 1 and signal 2 belong to the same type of signal, signal 3 and signal 4 belong to the same type of signal, and signal 5 and signal 6 belong to the same type of signal. In summary, signals 3 and 4 are arc sound signals, that is, the 10-18s time period.

[0072] Fig.11A feature data set for arc perception signals and sound perception signals was constructed, with 9 feature quantities as a group as the input of the BP neural network, 4500 groups of data as the training set, and 506 groups of data as the test set. The accuracy of the model was tested after training the training set data, and its accuracy rate was as high as 99%.

[0073] Fig.12 , Fig.13 , Fig.14 The figure shows the created arc fault detection system platform, which mainly includes arc monitoring interface, sound monitoring interface and arc fault identification interface. Through this system, the fault current can be detected more intuitively.

[0074] Compared with the traditional arc fault detection method, the novel arc data collection and diagnosis method implemented by the present invention greatly reduces the amount of data collected and stored, improves the working efficiency of the computer, and combines the arc current signal with the sound signal to improve the accuracy of arc system detection.

[0075] The above description is only an embodiment of the present invention and is not intended to limit the application of the present invention. Any modifications, equivalent substitutions, improvements, etc. made without departing from the basic features and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A DC fault arc detection method based on multi-feature fusion of compressed sensing, characterized in that: Step 1: Use an oscilloscope to collect the DC series current signal in the circuit; Step 2: Use the compressed sensing method on the collected current signal and use the sensing matrix to obtain data in the compressed sensing domain from the massive data. Y=ΦΨx (1) Among them, Y is the perception signal, Φ is the observation matrix, Ψ is the sparse matrix, and x is the original signal; Step 3: The signal after compressed sensing is random and disordered. It is impossible to directly distinguish whether an arc fault occurs in the system by relying solely on the current signal after compressed sensing. The misjudgment rate in the DC system is relatively high. Therefore, it is necessary to extract the features of the sensed arc signal: Standard Deviation: Peak value: s2=max|Y(n)| (3) RMS value: Square root amplitude: energy: Among them, Y(n) is the perception signal, M is the dimension of the perception signal; Step 4: Analyze the collected arc sound signals, and extract the sound signals when there is no arc and the sound signals when an arc occurs; Step 5: Use the compressed sensing method on the collected sound signal and use the perception matrix to obtain data in the compressed sensing domain from the massive data; Step 6: Perform feature analysis on the collected sound perception signals: volume: energy: Amplitude: a3 = max|Y(i)| (9) RMS value: Among them, Y(i) is the perception signal, M is the dimension of the perception signal; Step 7: Perform feature fusion on the collected feature quantities, that is, form a feature data sample with 9 feature values ​​as a group; Step 8: Use BP neural network to make decisions on detection features, and obtain a detection accuracy of up to 99%; Step 9: Based on the detection method, an arc monitoring interface is designed, through which arc signals and sound signals can be obtained, making arc monitoring more real-time and convenient.

Images