A research method for enhancing the detection features of direct-current arc in photovoltaic systems based on adaptive morphology

Abstract

The application belongs to the technical field of photovoltaic electrical fault detection, and particularly discloses a research method for detecting features of a direct-current fault arc of a photovoltaic system based on adaptive morphological enhancement. In order to ensure the reliability of the system, the fault arc is accurately detected and the detection features are enhanced. The enhancement effect of morphological on the detection features of the direct-current fault arc is studied, and an adaptive morphological algorithm is proposed. The particle swarm algorithm is combined with signal features to dynamically optimize the structure element, improve the signal signal-to-noise ratio, enhance the detection features, and compared with the stochastic resonance and wavelet analysis algorithm, so that the good feature enhancement effect of the application is verified, and the normal and fault arcs can be effectively distinguished. Finally, the features are decided by combining the support vector machine, and a detection accuracy of 96.23% is obtained. The adaptive morphological algorithm can effectively enhance the detection features of the fault arc, is suitable for adaptive optimization in different experimental scenes, and is helpful for accurately and effectively detecting the direct-current fault arc.

Description

Technical Field

[0001] This application belongs to the field of photovoltaic electrical fault detection technology, and is a method for detecting DC fault arcs in photovoltaic systems. It can effectively extract and enhance the detection features of fault arcs, which is beneficial for the accurate and effective detection of DC fault arcs. Background Technology

[0002] With the proposal of my country's carbon peak and carbon neutrality goals, photovoltaic power generation has gradually become an important way to promote low-carbon transformation. In recent years, my country's total photovoltaic installed capacity has increased from 6.24 million kW to 253.17 million kW. The existing photovoltaic installed capacity has an average annual emission reduction benefit of about 200 million tons, which can reach 1.92 billion tons by 2030 ([1] Han Mengyao, Xiong Jiao, Liu Weidong. Spatiotemporal distribution, competitive pattern and emission reduction benefits of photovoltaic power generation in China [J]. Journal of Natural Resources, 2022, 37(05): 1338-1351.). However, the safety problems that follow are also increasing. According to statistics, fires caused by fault arcs account for about 70% of photovoltaic power plant fires, and the hazards caused by connectors cannot be ignored. In fact, a 1MW photovoltaic power plant will use about 4,000 sets of connectors ([2] Peng Qijun, PV-HCB30 photovoltaic connector. Zhejiang Province, Zhejiang Xinhui Photovoltaic Technology Co., Ltd., 2016-11-26.). With the increase in the number of photovoltaic installed capacity, the number of connectors is also gradually increasing. Improper connector installation, long-term weathering, and animal biting can all cause fault arcing. Therefore, developing an effective DC fault arcing detection scheme is of great significance for early warning of fault arcing.

[0003] Domestic and foreign scholars have conducted a large number of studies on the detection of DC fault arc. The main ones are: ([3] Xiong Qing, Ji Shengchang, Zhu Lingyu, et al. Review of research progress on fault arc characteristic detection and location methods in low voltage DC system[J]. Proceedings of the CSEE, 2020, 40(18):6015-6026. [4] Guo Fengyi, Ruan Junyi, Liu Dawei, Wang Zhiyong. Research status and development trend of DC fault arc in photovoltaic system[J]. Electrical Appliances and Energy Efficiency Management Technology, 2018(10):1-7+19. [5] Gao Shaobin, Zhu Hongwei. Review of DC fault arc detection technology[J]. Electrical Appliances and Energy Efficiency Management Technology, 2018(10):20-24. [6] X. Yao, J. Wang, DLSchweickart. Review and recent developments in DC arc fault detection, 2016 IEEE International Power Modulator and High Voltage Conference (IPMHVC), 2016, pp.467-472.: This method employs physical characteristics such as sound, light, and electromagnetic radiation; it uses a time-frequency domain threshold determination method; and it employs artificial intelligence technology. This method mainly extracts features through time-frequency domain methods and finally uses artificial intelligence technology for fault arc detection.

[0004] Fault arc detection method based on physical characteristics such as sound, light and electromagnetic radiation. S Zhao et al. calculated the structural similarity index and 6-dB bandwidth box for electromagnetic radiation emitted by arc to extract the similarity of steady-state arc spectrum ([7] S. Zhao, Y. Wang, F. Niu, C. Zhu, Y. Xu, K. Li. A Series DC Arc Fault Detection Method Based on Steady Pattern of High-Frequency Electromagnetic Radiation[J]. in IEEE Transactions on Plasma Science, 2019, 47(9): 4370-4377.). Xiong Qing et al. designed a 4th-order Hilbert antenna based on electromagnetic radiation characteristics, which can effectively distinguish between normal and fault-induced system transients ([8] Xiong Qing, Xiao Rong, Ji Shengchang, Zhu Lingyu, Liu Yuan, Zhong Lipeng. DC arc detection method based on electromagnetic radiation characteristics[J]. High Voltage Engineering, 2017, 43(09): 2967-2975.). Wang Yao et al. designed a third-order Hilbert antenna to distinguish fault arcs and analyzed the influence of measurement distance on the intensity of electromagnetic radiation signal, providing ideas for fault arc location technology ([9] Wang Yao, Zhang Yanfeng, Niu Feng, Zhao Shuangle, Li Kui. Analysis and measurement method of electromagnetic radiation characteristics of photovoltaic DC arc [J]. Journal of Electrical Engineering, 2019, 34(14):2913-2921.).

[0005] The detection method based on time-frequency domain threshold determination is currently the main method for detecting fault arcs. Tang Shengxue et al. proposed a method for detecting weak DC series fault arcs by combining current wavelet energy entropy with extreme learning machine based on the UI characteristics of photovoltaic cells (

[10] Tang Shengxue, Diao Xudong, Chen Li, Zhang Jixin, Yao Fang. Research on the detection method of weak DC series fault arcs in photovoltaic power generation system [J]. Journal of Instrumentation, 2021, 42(03): 150-160.). Q.Lu et al. proposed a method for detecting DC series arc faults by comprehensively utilizing circuit current and power supply voltage based on the UI characteristics of DC fault arcs (

[11] Q.Lu, Z.Ye, M.Su, Y.Li, Y.Sun, H.Huang. A DC Series Arc Fault Detection Method Using Line Current and Supply Voltage [J]. IEEE Access, 2020, 8: 10134-10146.). H.-P. Park et al. proposed a fault arc detection method that compares the frequency domain and time domain current abrupt changes (

[12] H.-P. PARK, S. CHAE. DC Series Arc Fault Detection Algorithm for Distributed Energy Resources Using Arc Fault Impedance Modeling[J].in IEEE Access,2020,8:179039-179046.). H.-L. Dang et al. studied and analyzed several typical loads, selected 5 time domain features for fault arc detection, and compared the detection accuracy of various machine learning algorithms (

[13] H.-L. Dang, J. Kim, S. Kwak, S. Choi. Series DC Arc Fault Detection Using Machine Learning Algorithms[J].in IEEE Access,2021,9:133346-133364.).WMiao et al. proposed an improved method for detecting series fault arcs by combining empirical mode decomposition to extract time-frequency features with support vector machines (SVM) (

[14] W.Miao, Q.Xu, KHLam, PWTPong, HVPoor. DC Arc-Fault Detection Based on Empirical Mode Decomposition of Arc Signatures and Support Vector Machine[J].in IEEE Sensors Journal,2021,21(5):7024-7033.). Di Zhenguo et al. proposed a new method for detecting DC fault arcs based on weighted differential current, which weights the characteristic frequency band and characteristic parameters by differential (

[15] Di Zhenguo, Xiong Qing, Zhang Chen, Li Rui, Ji Shengchang. DC fault arc detection method based on weighted differential current[J / OL]. Journal of Xi'an Jiaotong University,2022(05):1-9[2022-05-19].). Meng Yu et al. proposed an improved Catboost algorithm for DC fault arc detection, which can quickly detect fault arcs under six different electrode materials (

[16] Meng Yu, Chen Silei, Wu Zihao, Wang Chenxi, Li Xingwen. Improved Catboost algorithm for DC fault arc detection under different electrode materials [J]. Journal of Xi'an Jiaotong University, 2022, 56(03): 124-134.). Meanwhile, the study proved that random resonance can effectively enhance the arc detection characteristics under different DC system topologies (

[17] Meng Yu, Chen Silei, Wu Zihao, Wang Chenxi, Li Xingwen. Study on enhancing photovoltaic DC fault arc detection characteristics based on random resonance method [J]. Proceedings of the CSEE, 2022, 42(06): 2396-2407.). Zhao Tiejun et al. obtained the series fault arc signal of the filter capacitor branch at the input end of the string based on the high frequency characteristics of the fault arc, and established a series fault arc detection algorithm based on sample entropy and standard deviation (

[18] Zhao Tiejun, Meng Jing, Song Yueqi, Xie Xiaoying, Zhang Mengchen. DC series arc fault detection and protection strategy for string photovoltaic system [J]. Power System Protection and Control, 2020, 48(20):74-82.). Huang Xiaoxiao et al. collected arc noise signals under various conditions and analyzed and compared the signals from the time domain, frequency domain and time-frequency domain respectively, proving that the time-frequency domain detection method is more suitable for the comprehensive analysis and detection of fault arc (

[19] Huang Xiaoxiao, Wu Chunhua, Li Zhihua, Wang Fei. Comparative study on DC arc fault detection methods of photovoltaic system [J]. Acta Energiae Solaris Sinica, 2020, 41(08):204-214.).M.Kavi et al. captured the continuous burning of electric arcs by generating continuous random spikes at the output of the decomposed open-closed alternating sequence algorithm, and further incorporated fault location technology based on increasing effective resistance (

[20] M.Kavi, Y.Mishra, M.Vilathgamuwa. DC Arc Fault Detection For Grid-Connected Large-Scale Photovoltaic Systems[J].in IEEEJournal ofPhotovoltaics,2020,10(5):1489-1502.). Xiong Qing et al. used the spectrum integral difference of the parallel capacitor current and the polarity of the first pulse of the capacitor current to realize the detection and location of fault arcs (

[21] Xiong Qing, Liu Xiaojun, Guo Ziqing, Feng Xianyong, Ji Shengchang, Zhu Lingyu. Detection and location of electric arc faults in photovoltaic systems based on the spectrum integral difference of the current[J]. High Voltage Engineering,2021,47(05):1625-1633.).

[0006] In summary, most current DC fault arc detection methods have not studied or considered the importance of enhancing fault detection features under the noise interference of complex systems. Reference (

[10] Tang Shengxue, Diao Xudong, Chen Li, Zhang Jixin, Yao Fang. Research on DC series weak fault arc detection method of photovoltaic power generation system[J]. Journal of Instrumentation, 2021, 42(03):150-160.) uses wavelet energy entropy to enhance the detection of weak arcs. Reference (

[17] Meng Yu, Chen Silei, Wu Zihao, Wang Chenxi, Li Xingwen. Research on enhancing the detection features of photovoltaic DC fault arc based on random resonance method[J]. Proceedings of the CSEE, 2022, 42(06):2396-2407.) uses random resonance to enhance the detection features obtained by wavelet transform under the consideration of noise, which has a certain effect, but the selection of its parameters is not adaptive. Therefore, in order to effectively extract feature information, it is necessary to adaptively enhance the detection features of fault arcs.This invention proposes an adaptive morphological algorithm for feature enhancement of signals. There are existing literatures that use mathematical morphology (MM) to study fault arcs (

[20] M.Kavi, Y.Mishra, M.Vilathgamuwa. DC Arc Fault Detection For Grid-Connected Large-Scale Photovoltaic Systems[J].in IEEE Journal). of Photovoltaics, 2020, 10(5): 1489-1502.

[22] Gao Shaobin. Research on DC fault arc detection of photovoltaic system [D]. Zhejiang University, 2019.

[23] Cui Ruihua, Li Sisi, Jia Xiaoxiang, Hu Wenda. Research on the application of mathematical morphology in the detection of series fault arc in aviation [J]. Electrical Appliances and Energy Efficiency Management Technology, 2017, (19): 13-17.

[24] Liu Zhenguo, Sun Peng, You Guangxiang. Application of mathematical morphology in the diagnosis of series fault arc [J]. High Voltage Apparatus, 2016, 52(09): 190-195.

[25] M. Weerasekara, M. Vilathgamuwa, Y. MISHRa. Detection of high impedance faults in PV systems using mathematical morphology [C]. 2018 IEEE International Conference on Industrial Electronics for Sustainable Energy Systems (IESES), 2018, pp. 357-361.), its structure and general functions can be seamlessly applied in DC fault arc detection (

[26] RFAmmerman, T. Gammon, PKSen, JPNelso. DC-arc models and incident-energy computing[J]. IEEE Trans. Ind. Appl., 2010, 46(5): 1810-1819.). However, the role of the MM algorithm in enhancing detection features has not been studied. This invention is the first to use the MM algorithm to study the feature enhancement of signals, and proposes an adaptive MM algorithm. The particle swarm algorithm is used to dynamically and adaptively select the length of the structural element and its internal parameters in combination with the signal features, and the feature enhancement effect is compared with that of the stochastic resonance and wavelet analysis algorithms. Finally, the energy features of the signal processed by the adaptive MM algorithm are extracted by using the method of moving step size equal to the sliding window size, and the DC fault arc detection algorithm is constructed using SVM. Summary of the Invention

[0007] Complex system noise interference in photovoltaic DC systems makes it difficult to effectively extract the characteristics of fault arcs, leading to misjudgments or missed detections. Therefore, to ensure system reliability, accurate detection of fault arcs and enhancement of their detection characteristics are crucial.

[0008] To solve the above problems and achieve the above objectives, the present invention provides the following technical solution:

[0009] A research method for detecting DC fault arcs in photovoltaic systems based on adaptive morphology enhancement is characterized by the following steps:

[0010] S1: Use a current transformer to collect the series current signal in the line;

[0011] S2: Adaptive morphological processing is performed on the acquired signal. Its adaptive mechanism uses the particle swarm algorithm to combine signal features to adaptively and dynamically optimize the structural elements, thereby improving the signal-to-noise ratio and enhancing the detection features.

[0012] S3: The morphological filtering algorithm used is a closed-closed-open hybrid operation constructed based on the basic operations of erosion and dilation combined with signal features;

[0013] S4: The objective function of the particle swarm optimization algorithm is the energy ratio of fault and normal signals;

[0014] S5: Extract detection features from the signal after it has been enhanced by an adaptive morphological algorithm using a feature extraction method with a step size equal to the size of the sliding window.

[0015] S6: Finally, after extracting features from the data collected in four different scenarios, a support vector machine was used to make decisions on the detection features, resulting in a detection accuracy of up to 96.23%.

[0016] Furthermore, step S1 specifically involves: in the experimental scenario, the present invention constructs four scenarios for the detection and research of DC series fault arcs: photovoltaic connected to the grid, photovoltaic connected to a resistive load, DC power supply connected to the inverter, and DC power supply connected to a resistive load. Current signals in the line are collected using current transformers for subsequent analysis and processing.

[0017] Furthermore, step S2 specifically involves: after acquiring the current signal, performing adaptive morphological algorithm enhancement processing on it, and using particle swarm optimization algorithm combined with signal features to dynamically optimize the length and internal parameters of the morphological structural elements throughout the entire dynamic adaptive processing process.

[0018] Furthermore, step S3 specifically involves: the proposed morphological algorithm is a closed-closed-open hybrid operation constructed based on the basic operations of erosion and dilation combined with signal features.

[0019] Let the one-dimensional voltage signal be x(n), and the structuring element be g(m). Determine the corrosion (Θ) and expansion of x(n) with respect to g(m). The operation is as follows:

[0020] (xΘg)(n)=min{x(n+m)-g(m)} (1)

[0021]

[0022] Where the domain of x(n) is D x ={x0,x1,…,x N-1 The domain of g(m) is D. g ={g0,g1,…,g M-1}, and N>M, (n+m), (nm)∈D x ,m∈D g The closing operation (·) and opening operation (o) formed by combining equations (1) and (2) are:

[0023]

[0024]

[0025] Therefore, this invention, through a series of experimental tests, performs a closed-closed-open-open mixed operation on the signal with the origin of the structural element at the starting point, as shown in the following formula:

[0026] a[x(n)]=(x·g·g)(n) (5)

[0027]

[0028] y(n)=a[x(n)]+b[x(n)] (7)

[0029] Where g is the structuring element and n (n = 1, 2, 3, ...) is the number of points in the signal.

[0030] Furthermore, step S4 specifically involves using a particle swarm optimization algorithm during the adaptive morphological processing, wherein the objective function of the particle swarm optimization algorithm is the energy ratio of the fault signal to the normal signal.

[0031] Furthermore, step S5 specifically involves extracting detection features from the signal after adaptive morphological filtering using a feature extraction method with a step size equal to the sliding window size.

[0032] Furthermore, step S6 specifically involves: extracting features from 40 sets of data randomly collected under four scenarios—photovoltaic connected to an inverter, photovoltaic connected to a resistor, DC power supply connected to an inverter, and DC power supply connected to a resistor—and then using a support vector machine to make decisions on the detection features, achieving a detection accuracy of up to 96.23%.

[0033] The beneficial effects of this invention are as follows:

[0034] (1) The adaptive MM algorithm proposed in this invention is used to perform feature enhancement processing on DC fault arc signal, which can effectively improve the signal-to-noise ratio of the original signal and enhance the distinction between fault signal and normal signal. In the comparative experiment with random resonance and wavelet analysis algorithms, the feature enhancement effect of the adaptive MM algorithm proposed in this invention is also highlighted. It can effectively improve the reliability and stability of DC fault arc detection system based on transformer acquisition and reduce fault misjudgment.

[0035] (2) The adaptive MM algorithm proposed in this invention can dynamically and adaptively optimize parameters by combining signal features, thereby optimizing the feature enhancement effect of the MM algorithm. It avoids manual optimization, reduces human error and workload, and makes the algorithm more intelligent and systematic. Moreover, it can perform adaptive optimization for different experimental scenarios and has a good feature enhancement effect.

[0036] (3) The DC series fault arc detection method for photovoltaic systems based on the SVM algorithm achieved a high detection accuracy and can effectively detect DC fault arcs in different experimental scenarios. It has reference value for other DC fault arc detection systems based on current transformer data acquisition. Attached Figure Description

[0037] Figure 1 This is a flowchart of the overall technical solution of the present invention.

[0038] Figure 2 Experimental platform for acquiring DC fault arc data

[0039] Figure 3 Original data waveform

[0040] Figure 4 Flowchart of the adaptive morphology algorithm

[0041] Figure 5 Diagram of the length optimization process for structural elements

[0042] Figure 6 Signal-to-noise ratio comparison chart

[0043] Figure 7 Features extracted from four experimental scenarios

[0044] Figure 8Features extracted by a single-step sliding window

[0045] Figure 9 Comparison of different feature enhancement algorithms Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] Figure 1 This is a flowchart of the overall technical solution of the present invention. In the constructed experimental scenario, current signals are collected through a current transformer. Based on the particle swarm optimization algorithm combined with signal features, the length and parameters of the structural elements are adaptively optimized. Then, the original signal undergoes adaptive morphological enhancement processing. Features are extracted from the processed signal using a feature extraction method where the step size equals the sliding window size. Finally, SVM is used for decision-making, outputting whether a faulty arc exists.

[0048] Figure 2 This invention is a DC fault arc data acquisition experimental platform built according to the UL1699B standard. Compared with parallel fault arcs, the current is smaller when a series fault arc occurs, making it difficult for protection devices to detect (

[27] Jiao Zhijie, Li Teng, Wang Lina, Mou Longhua, Alexandra Khalyasmaa. DC series arc fault detection of photovoltaic system based on convolutional neural network [J]. New Technology of Electrical Engineering and Energy, 2019, 38(07). 29-34.). In order to obtain DC series fault arc information, this invention built a DC series fault arc experimental platform in the laboratory scenario of State Grid Chongqing Electric Power Company. Four series photovoltaic panels and DC power supply can be powered separately. The load is a high-power resistor and a grid-connected inverter. The component parameters of the experimental platform are shown in Table 1.

[0049] The arc generator was constructed according to the UL1699B standard. The arc generating device consists of a fixed electrode, a movable electrode, a side adjuster, a slider, a screw rod, and a fixed base. The fixed electrode is a flat-headed copper rod, and the movable electrode is a pointed copper rod. Arc initiation is achieved by adjusting the side adjuster using the drawing arc method. Before each experiment, the surface of the copper rods is sanded smooth to remove any traces left from the previous experiment's arc ignition, ensuring successful arc initiation.

[0050] In this invention, the sampling frequency is 100kHz. The current signal in the circuit is measured by a current transformer. The current signal is converted into a voltage signal by an external 200-ohm sampling resistor. The signal is then acquired and transmitted to a PC via a PicoScope 5000Series. Finally, the acquired data is analyzed and processed using MATLAB software.

[0051] Table 1 Component parameters of the experimental platform

[0052]

[0053] Figure 3 The original data collected by this invention under four experimental scenarios—photovoltaic connected to inverter, photovoltaic connected to resistor, DC power supply connected to inverter, and DC power supply connected to resistor—shows that the signal fluctuation is small before the fault arc occurs and large after the fault arc occurs. In particular, strong system noise interference occurs in the scenario of connecting to the grid with the inverter. The main reason why the waveform does not decrease but increases when the arc occurs is that the operating point of the photovoltaic power supply (or DC power supply) in the circuit is in the constant current source region or the non-constant current source region before and after the fault arc occurs (

[28] Wu Jingjing. Research on series DC fault arc detection method [D]. Hefei University of Technology, 2020.). At the same time, the voltage signal collected by this invention has the same shape characteristics as the voltage signal collected by the current transformer in the literature (

[29] Chen Yonghui, Xiong Lan, Fan Yuyi, Liu Xuan, Guo Ke. Photovoltaic arc fault detection method based on current transformer voltage signal [J]. Acta Energiae Solaris Sinica, 2021, 42(10):68-75.).

[0054] In practical acquisition systems, there is often a large amount of system noise, and useful signals are submerged in noise signals, making it difficult to extract effective features of the signal. Existing literature generally uses empirical mode decomposition (

[14] W.Miao, Q.Xu, KHLam, PWTPong, HVPoor. DC Arc-Fault Detection Based on Empirical Mode Decomposition of Arc Signatures and Support Vector Machine[J].in IEEE Sensors Journal, 2021, 21(5): 7024-7033.), wavelet transform (

[30] Hu Jixin, Xu Yongxin, Geng Yicheng et al. A method for identifying series DC arc faults in photovoltaic systems based on multi-feature fusion [J]. Modern Power, 2022, 39(05): 529-536.) and singular value decomposition (

[31] Liu Xuan, Xiong Lan, Wang Yun et al. Detection algorithm and protection experiment of series DC arc in photovoltaic power station [J]. Acta Energiae Solaris Sinica, 2022, 43(01): 348-355.) and other methods to reduce noise in signals, but empirical mode decomposition and wavelet transform, as time-frequency domain methods, are more computationally complex than time-domain methods in processing signals in both the time and frequency domains. Singular value decomposition, as a two-dimensional matrix signal, has a higher computational complexity than directly processing one-dimensional signals. Moreover, the local features of DC fault arc signals are relatively complex, and effectively extracting the local features of the signal is helpful for arc detection. To reduce computational complexity while considering the importance of local signal features, an effective algorithm is needed.Therefore, this invention proposes an adaptive MM algorithm to process one-dimensional signals. Since MM is a nonlinear analysis method, it can effectively extract local feature information of small changes in DC fault arc signals (

[32] M.Kavi, Y.Mishra, M.Vilathgamuwa. DC Arc-Fault Detection in PV Systems Using Multistage Morphological Fault Detection Algorithm[C].IECON2018-44th Annual Conference of the IEEE Industrial Electronics Society, 2018, pp. 1746-1751.), and the algorithm has low complexity and small delay (

[23] Cui Ruihua, Li Sisi, Jia Xiaoxiang, Hu Wenda. Research on the application of mathematical morphology in the detection of arc faults in aviation series [J]. Electrical and Energy Efficiency Management Technology, 2017, (19): 13-17.), and its good noise reduction performance has also been gradually applied to the power system (

[33] Zhao Zhao, Liu Lilin, Zhang Chengxue, Li Haitao, Wang Jianmin. Research and analysis on the selection principle of structural elements of morphological filter [J]. Power System Protection and Control, 2009, 37 (14): 21-25+35.).

[0055] MM is an algebraic operation that mainly uses structuring elements to perform neighborhood operations with signals. The structuring element acts as a sliding window and interacts with the signal to detect features near each point of the signal (

[25] M. Weerasekara, M. Vilathgamuwa, Y. MISHRa. Detection of high impedance faults in PV systems using mathematical morphology[C]. 2018 IEEE International Conference on Industrial Electronics for Sustainable Energy Systems (IESES), 2018, pp. 357-361.). Its basic operations are erosion and dilation operations, as well as the closing and opening operations formed by the combination of the two.

[0056] Let the one-dimensional voltage signal be x(n), and the structuring element be g(m). Determine the corrosion (Θ) and expansion of x(n) with respect to g(m). The operation is as follows:

[0057] (xΘg)(n)=min{x(n+m)-g(m)} (1)

[0058]

[0059] Where the domain of x(n) is D x ={x0,x1,…,x N-1 The domain of g(m) is D. g ={g0,g1,…,g M-1}, and N>M, (n+m), (nm)∈D x ,m∈D g The closing operation (·) and opening operation formed by combining equations (1) and (2) for:

[0060]

[0061]

[0062] In the closed operation, the signal is first expanded and then corroded, while the open operation is the opposite. Since the arc signal collected by the transformer actually contains impacts in both positive and negative directions, the traditional corrosion and expansion operations can only process information in one direction. In order to construct an effective morphological algorithm, it is necessary to combine its basic operations (

[34] Jiang Wanlu, Zheng Zhi, Zhu Yong, Liu Siyuan. Research on hydraulic pump fault diagnosis based on the optimal flat structural element length [J]. Vibration and Shock, 2014, 33(15): 35-41.). Therefore, this invention performs a series of experimental tests, with the origin of the structural element at the starting point, to perform a closed-closed-open-open mixed operation on the signal, as shown in the following formula:

[0063] a[x(n)]=(x·g·g)(n) (5)

[0064]

[0065] y(n)=a[x(n)]+b[x(n)] (7)

[0066] Where g is the structuring element and n (n = 1, 2, 3, ...) is the number of points in the signal.

[0067] Because different structural element lengths and internal parameters have different effects on signal processing, they should be matched as closely as possible to the characteristics of the signal to be analyzed (

[22] Gao Shaobin. Research on DC fault arc detection of photovoltaic system [D]. Zhejiang University, 2019.). For one-dimensional signals, flat structural elements have advantages such as small computational load, zero height and few parameters to be optimized (

[35] Wang Tao. Research on axle box bearing diagnosis method based on optimized morphological filtering [D]. Southwest Jiaotong University, 2020.). Therefore, this invention selects flat structural elements to process the signal. The length and internal parameters of flat structural elements have a large selection range. In order to better adapt to the local characteristics of the signal and make the feature enhancement effect of the MM algorithm proposed in this invention the best, this invention proposes an adaptive MM algorithm to dynamically and adaptively analyze and process the signal.

[0068] Figure 4 This is a flowchart of the adaptive MM algorithm proposed in this invention. First, within the custom length range of the candidate structuring elements, a one-dimensional array g[i] = [m(1), m(2), ..., m(i)] containing unknown variables is constructed based on the length i of the current structuring element. This array represents the structuring element to be optimized. Then, the current structuring element is applied to the original signal for MM algorithm processing. After processing, the energy ratio of the current fault and normal signals is calculated to obtain E_new. This is compared with the custom E value. If E_new is greater than the E value, the current E is replaced; otherwise, the E value remains unchanged. Finally, the length and its internal parameters corresponding to the maximum E value within the custom structuring element length range are output. This invention uses the energy ratio (E) of the fault and normal signals as the objective to adaptively select the structuring element length and parameters. Throughout the dynamic parameter selection process, the particle swarm optimization algorithm, combined with the MM algorithm, uses the E value as the objective function to adaptively select the structuring element parameters based on the current structuring element length, optimizing the output structuring element with the maximum E value. Each time the length of the structuring element increases, a comparison and replacement of the E value is performed. Finally, the length and parameters of the structuring element corresponding to the maximum E value are output and imported into the MM algorithm (formula (7)) designed in this invention for signal enhancement processing.

[0069] Because the particle swarm optimization algorithm has a fast convergence speed, few parameters, and is simple and easy to implement (

[36] Kennedy J. Particleswarm optimization[J]. Proc. of 1995 IEEE Int. Conf. Neural Networks, (Perth, Australia), Nov. 27-Dec. 2011, 4(8): 1942-1948.), this invention uses the particle swarm optimization algorithm to adaptively optimize the selection of structural elements. The parameters of the particle swarm optimization algorithm are set as follows: particle position range [-11], particle motion speed range [-2525], learning factor [1.31.7], inertia weight [0.20.9], initial swarm size is 100, and maximum number of iterations is set to 10. The selection of the above parameters refers to the literature (

[37] Wang Dongfeng, Meng Li. Performance analysis and parameter selection of particle swarm optimization algorithm[J]. Acta Automatica Sinica, 2016, 42(10): 1552-1561.), and is also selected in combination with actual test results.

[0070] To verify the enhancement effect of the adaptive MM algorithm proposed in this invention, this invention processed a random set of data collected under four experimental scenarios: photovoltaic connected to an inverter, photovoltaic connected to a resistor, DC power supply connected to a resistor, and DC power supply connected to an inverter, and output the final adaptively obtained structural element. The parameter optimization results of the final adaptive structural element obtained under the four experimental scenarios are shown in Table 2.

[0071] Table 2. Optimal parameters of structural elements

[0072]

[0073] Figure 5 The diagram illustrates the length optimization process of the structural element proposed in this invention. In the photovoltaic-inverter scenario, a length of 17 is optimal; in the photovoltaic-resistor scenario, a length of 7 is optimal; in the DC power supply-inverter scenario, a length of 9 is optimal; and in the DC power supply-resistor scenario, a length of 3 is optimal. This invention only provides adaptive parameter-finding examples under four experimental scenarios. For different voltages and currents, especially under different solar power generation conditions in photovoltaic scenarios, the length of the structural element and its internal parameters are different. The innovation of this invention lies in how to dynamically and adaptively find the optimal parameters so that the feature enhancement effect of the original signal after processing by the MM algorithm proposed in this invention is optimal.

[0074] Figure 6This invention presents a comparison of the signal-to-noise ratio (SNR) of the original signal and the signal processed by the adaptive MM algorithm. To better demonstrate the effectiveness of the proposed adaptive MM algorithm, additive white Gaussian noise (SNR set to 20dB) is simultaneously superimposed on both signals, and the SNR of each signal with the noise is calculated separately. For noisy signals, a higher SNR is better. Compared to the original signal, the signal-to-noise ratio of the signal processed by the proposed adaptive MM algorithm is improved, indicating a higher quality of the useful signal. This demonstrates that the proposed adaptive MM algorithm can effectively improve the SNR of a signal, thereby effectively enhancing the detection characteristics of the signal.

[0075] After the signal is processed by the adaptive MM algorithm, the crucial feature extraction process begins. This invention employs a sliding window with a step size equal to the window size (10ms, 1ms = 100 points) to extract features from the signal within the window. The sliding window method reduces the sensitivity of features to noise. Using a step size equal to the sliding window size avoids interference from extracted signals to unextracted signals. When the step size is smaller than the sliding window size, duplicate calculations occur; when the step size is larger than the sliding window size, information is missed. Therefore, this invention uses a step size equal to the window size for DC fault arc feature extraction, and determines the DC fault arc based on the feature values ​​of the signal within the window. Here, the sum of squares of the signals within the window is extracted as the detection feature.

[0076] Let the signal sequence be a sequence of length n, T = [t0, t1, ..., tn]. n-1 Let the sliding window size be m (m < n), and the step size be b (b = m) for each forward movement. Starting from index 0, calculate the sum of squares of the signals within the windows T[0:m-1], T[m:2m-1], ..., T[n-m+1:n], which is the energy. The formula is:

[0077]

[0078] Figure 7 These are the feature maps extracted under four experimental scenarios presented in this invention. This invention conducted tests under four experimental scenarios, with the data collected in the photovoltaic experimental scenario covering representative operating states of the photovoltaic system. First, the signals collected under the four experimental scenarios were adaptively processed using the adaptive MM algorithm. Then, a sliding window with a step size equal to the window size was used to extract features from the processed signals. The extracted features are shown below. Figure 7 As shown (only one random set of data from four scenarios is used here). For 1 second of data, this invention chooses a movement step size and window size of 10ms for ease of calculation and analysis. Each feature has 100 points, with the first 50 points representing normal data and the last 50 representing faulty data. Figure 7It can be seen that, in the four experimental scenarios, the features of the fault area after optimization by the adaptive MM algorithm are significantly enhanced compared with the original signal, and the distinction between normal and faulty signals is greater. This means that the adaptive MM algorithm proposed in this invention can effectively enhance the detection features of fault arcs, increase the distinction between normal signals and faulty arcs, and reduce fault misjudgment.

[0079] Figure 8 This invention provides a feature map extracted using a single-step sliding window. Considering that the signal within the window may simultaneously contain both normal and fault information in actual testing, to demonstrate the effectiveness of the feature extraction method used in this invention, under the same conditions, a single-step sliding window (window size 10ms) is used to extract the signal energy within 1s of data. This approach considers all possible signal scenarios within the window. Figure 8 As shown, the first 49001 points are normal, and from point 49002 onwards, the window contains fault information until the window is entirely filled with fault information. As can be seen from the marked points, the energy gradually increases from these points. Therefore, the feature extraction method used in this invention can effectively distinguish between normal and faulty conditions, reducing false fault identification.

[0080] Figure 9 This paper presents a comparison of different enhancement algorithms presented in this invention. Enhancing the detection features of fault arcs is of great significance in practical applications. Existing literature has used algorithms such as stochastic resonance and wavelet analysis to enhance the detection features of fault arcs, achieving some enhancement effects, but the results are generally limited and the parameter selection is not adaptive. This invention compares the enhancement effects of stochastic resonance, wavelet analysis, and adaptive MM algorithms on the original signal detection features. The parameters of stochastic resonance are also optimized using the particle swarm optimization algorithm, with the parameter settings unchanged and the objective function remaining the energy ratio of the fault and normal signals. The main factors affecting the stochastic resonance effect are system parameters a and b. Additive white Gaussian noise is added to the original signal, and the final optimized parameters are shown in Table 3. Wavelet analysis uses Rbio3.1 as the wavelet basis and employs a 6-level decomposition. Because the Rbio3.1 wavelet basis has better feature applicability than Db9 (

[17] Meng Yu, Chen Silei, Wu Zihao, Wang Chenxi, Li Xingwen. Research on enhancing the characteristics of photovoltaic DC fault arc detection based on stochastic resonance method [J]. Proceedings of the CSEE, 2022, 42(06):2396-2407.). Figure 9As shown in the figure, (Category 1 represents the original signal, Category 2 represents wavelet analysis, Category 3 represents random resonance, Category 4 represents random resonance + wavelet analysis, and Category 5 represents the adaptive MM algorithm proposed in this invention) In the four experimental scenarios, from the perspective of fault characteristics, random resonance and wavelet analysis only have a weak feature enhancement effect on the original signal. Random resonance + wavelet analysis is the method of first performing random resonance processing on the signal and then performing wavelet analysis. Compared with random resonance and wavelet analysis alone, it has a significant feature enhancement effect, but it is much worse than the adaptive MM algorithm proposed in this invention. Under the same experimental scenario conditions, by comparing the feature enhancement effect with related enhancement algorithms, it is proved that the adaptive MM algorithm proposed in this invention has a very good feature enhancement effect.

[0081] Table 3. Optimal parameters of stochastic resonance

[0082]

[0083]

[0084] In recent years, with the rapid development of artificial intelligence technology, machine learning and deep learning algorithms have been widely used in the field of arc fault detection (

[39] S. Lu, T. Sirojan, BTPhung, D. Zhang, E. AMBIKAIRAJAH. DA-DCGAN: An Effective Methodology for DC Series ArcFault Diagnosis in Photovoltaic Systems[J]. in IEEE Access, 2019, 7: 45831-45840.

[40] L. Xing, Y. Wen, S. Xiao, D. Zhang, J. Zhang. A Deep Learning Approach for Series DC ArcFault Diagnosing and Real-Time Circuit Behavior Predicting[J]. in IEEE Transactions on Electromagnetic Compatibility, 2022, 64(2): 569-579.). Traditional arc fault detection methods have problems such as high misjudgment rate and low detection rate, while machine learning and deep learning algorithms have the characteristics of automation and intelligence, and have outstanding advantages in arc fault detection. Because this invention constructs a small sample dataset, SVM is a classic algorithm for solving small sample data, with advantages such as solving nonlinear problems and good generalization ability. Therefore, this invention uses SVM to construct a DC fault arc detection algorithm and analyzes and compares its detection performance.

[0085] This invention randomly collected 40 sets of data in four experimental scenarios. After processing these 160 sets of data using the adaptive MM algorithm, features were extracted. Each set of data was divided into 100 periods by a sliding window, resulting in a total of 16,000 samples. The data was randomly shuffled in a 5:5 ratio to divide it into training and test sets, and then normalized to the range [-11] to eliminate the adverse effects of outlier samples. This is a crucial and essential step in machine learning.

[0086] The features constructed in this invention achieved a detection accuracy of 96.23% after SVM decision-making. To verify the effectiveness of the DC fault arc detection algorithm constructed in this invention, this invention was compared with some existing literature that uses SVM algorithm for decision-making. The comparison results are shown in Table 4. Due to the non-sharing of experimental scenarios and datasets, it is not possible to compare under the same conditions. This invention only compares the final fault arc detection accuracy with that of the literature (

[14] W.Miao, Q.Xu, KHLam, PWTPong, HVPoor. DC Arc-Fault Detection Based on Empirical Mode DecompositionofArc Signatures and Support Vector Machine[J].inIEEE Sensors). Journal, 2021, 21(5):7024-7033.), and literature (

[17] Meng Yu, Chen Silei, Wu Zihao, Wang Chenxi, Li Xingwen. Research on enhancing the characteristics of photovoltaic DC fault arc detection based on stochastic resonance method [J]. Proceedings of the CSEE, 2022, 42(06):2396-2407.) and literature (

[38] Ding Xin, Zhu Hongwei, Yin Haonan, Wang Yiwen. DC arc fault detection based on machine learning method [J]. Sensors & Microsystems, 2017, 36(11):123-127.) were compared, and the detection accuracy was improved by 0.23%, 4.53% and 1.75%, respectively. Reference (

[14] W.Miao, Q.Xu, KHLam, PWTPong, HVPoor. DC Arc-Fault Detection Based on EmpiricalMode Decomposition of ArcSignatures and Support Vector Machine[J].in IEEESensors Journal,2021,21(5):7024-7033.) uses empirical mode decomposition to suppress system noise, but the empirical mode decomposition parameters are not adaptively selected, and their effect is greatly affected by the parameters, and there is mode aliasing. Reference (

[17] Meng Yu, Chen Silei, Wu Zihao, Wang Chenxi, Li Xingwen. Research on enhancing the characteristics of photovoltaic DC fault arc detection based on stochastic resonance method[J]. Proceedings of the CSEE,2022,42(06):2396-2407.) uses stochastic resonance to enhance the detection characteristics after wavelet transform, and also uses the sum of squares of the signal as the detection feature, but the final detection effect is worse than that of this invention, and it is not an adaptive parameter selection.Reference (

[38] Ding Xin, Zhu Hongwei, Yin Haonan, Wang Yiwen. DC arc fault detection based on machine learning method [J]. Sensors & Microsystems, 2017, 36(11):123-127.) only extracted multiple features from the time domain and frequency domain, without considering the influence of system noise and the experimental scenario was relatively simple. This invention not only considers the influence of system noise and conducts tests in multiple experimental scenarios, but also enhances the detection features. Therefore, it can be seen that the overall arc fault detection algorithm proposed in this invention is superior to these references and can provide reliable and stable DC arc fault detection guarantee.

[0087] Table 4 Comparison of fault arc detection accuracy

[0088]

Claims

1. A research method for detecting DC fault arc characteristics in photovoltaic systems based on adaptive morphology enhancement, characterized in that, Includes the following steps: S1: Use a current transformer to collect the series current signal in the line; S2: Adaptive morphological processing is performed on the acquired signal. In this process, the length and internal parameters of the morphological structural elements are dynamically optimized by combining the particle swarm optimization algorithm with the signal features. The optimization objective function is the energy ratio of the fault signal and the normal signal, so as to improve the signal-to-noise ratio and enhance the detection features. S3: The morphological filtering algorithm used is a closed-closed-open-open hybrid operation constructed based on the basic operations of erosion and dilation combined with signal features; S4: Extract detection features from the signal after it has been enhanced by an adaptive morphological algorithm using a feature extraction method where the step size is equal to the size of the sliding window. S5: Finally, feature extraction is performed on the data collected under four different scenarios: photovoltaic connected to inverter, photovoltaic connected to resistor, DC power supply connected to inverter, and DC power supply connected to resistor. Support vector machine is then used to make decisions on the detection features to identify whether there is a fault arc.

2. The method as described in claim 1, characterized in that, The morphological filtering algorithm is a hybrid operation of closing-closing-opening and opening-closed. Its characteristics are: Let the one-dimensional voltage signal be x(n), and the structuring element be g(m). The erosion and dilation operations of x(n) with respect to g(m) are as follows: ； ； Where the domain of x(n) is D x ={x0,x1,…,x N-1 The domain of g(m) is D. g ={g0,g1,…,g M-1 }, and N>M, (n+m) and (nm)∈D x ,m∈D g， The closing and opening operations formed by combining the above two equations are: ； ； The signal is subjected to a closed-closed-open-open mixed operation with the origin of the structuring element at the starting point, as shown in the following formula: ； ； ； Where g is the structuring element and n (n=1,2,3,…) is the number of points in the signal.

Images