Arc fault diagnosis system and method based on iao-vmd center frequency

Abstract

The arc fault diagnosis system and method based on IAO-VMD center frequency comprises a data acquisition module, an optimal parameter acquisition module, a decomposition component acquisition module, a frequency band division module, a signal reconstruction frequency band, a fault feature extraction module and an integrated learning processing module; each module is connected in turn. The improved eagle algorithm is used to optimize the optimal parameter combination of VMD decomposition, and the final optimization result is output. The IMF components in the same frequency band are superimposed and combined, the multi-domain fault features are extracted, and the low-dimensional fault feature set is reconstructed. Different base learners are selected for integrated learning of the low-dimensional feature set in different frequency bands to output the final fault diagnosis result. The improved eagle algorithm is used to optimize the VMD parameters, the data dimension reduction method is used to reduce the calculation complexity, and the current signal parameters of multiple angles are selected to extract the arc fault features, so that the fault information extracted by a single feature is not comprehensive, the fault information can be accurately reflected, and the diagnosis efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of arc fault diagnosis, and particularly to an arc fault diagnosis system and method based on IAO-VMD center frequency. BACKGROUND

[0002] According to the fire statistics report in recent years, the proportion of electrical fire is at the top, and arc fault is an important reason for electrical fire. In the alternating current system, low-voltage power supply line insulation aging, ground fault and loose connection of wiring terminal can cause alternating current arc fault. Arc fault is mainly divided into three categories, among which parallel and ground arc faults are easy to be detected by traditional protection devices. When series arc fault occurs, the line current decreases, and the protection devices such as circuit breaker and fuse cannot trip to protect the line. Therefore, how to effectively solve the low-voltage alternating current arc fault has become a hot issue and has been widely concerned and researched by domestic and foreign scholars.

[0003] The detection methods of alternating current arc fault mainly include methods based on arc mathematical model, physical characteristics and pattern recognition. The existing arc mathematical models include Cassie model, Mayr model, Stokes model, Ayrton model for describing the macroscopic external characteristics of arc, and magnetohydrodynamic model for describing the internal physical characteristics of arc. The radiation characteristic method is affected by environmental factors and has limited positioning range, so it is rarely used. The most commonly used method is the pattern recognition method, which often processes the fault signal in time and frequency domain, and then uses machine learning algorithm for fault diagnosis.

[0004] VMD (Variational Mode Decomposition) is a time-frequency domain separation algorithm based on iterative operation, which can effectively overcome mode mixing. However, if the parameter selection is not proper, it will take too much execution time and the decomposition effect will not be good. The low-voltage alternating current system has complex and variable load, weak arc fault characteristics and many influencing factors. Inaccurate signal decomposition will greatly affect the diagnosis result of fault arc. In addition, many methods using ML to process alternating current series arc fault have been studied, but these ML algorithms are mostly single prediction models, and the generalization ability is generally limited to improve the diagnosis accuracy. At the same time, due to the dynamic and uncertain nature of low-voltage distribution system, it will bring great challenges to arc fault detection and diagnosis. SUMMARY

[0005] The purpose of the present application is to overcome the above-mentioned deficiencies of the prior art and provide an arc fault diagnosis system and method based on IAO-VMD center frequency.

[0006] The technical scheme of the application is: an arc fault diagnosis system based on IAO-VMD center frequency, comprising a data acquisition module, an optimal parameter acquisition module, a decomposition component acquisition module, a frequency band division module, a signal reconstruction frequency band, a fault feature extraction module and an integrated learning processing module.

[0007] The data acquisition module is used to select current signal data in existing circuit normal state and arc fault state, or to obtain current signal data in simulated circuit normal state and arc fault state through an alternating current arc fault simulation and acquisition experiment platform.

[0008] The optimal parameter acquisition module is used to process feature data of a group of randomly extracted fault arcs by using a VMD method, to optimize the best parameter combination of VMD decomposition by using an improved eagle algorithm, and to output the final optimization result.

[0009] The decomposition component acquisition module is used to input the finally optimized VMD decomposition best parameter combination into VMD, to process feature data sets of circuit normal state and fault arc state, and to obtain K IMF components of the feature data sets of circuit normal state and fault arc state and the center frequencies corresponding to each component.

[0010] The frequency band division module is used to divide the feature data sets of normal state and fault arc state into four equal parts after FFT frequency domain conversion, and to determine the frequency bands to which the IMF components belong according to the center frequencies of the IMF components.

[0011] The signal reconstruction module is used to superimpose and merge the IMF components in the same frequency band to obtain new low-frequency, medium-low-frequency, medium-high-frequency and high-frequency sub-sequences, and to output the reconstructed signal.

[0012] The fault feature extraction module is used to extract multi-domain fault features from each reconstructed sub-sequence, to perform data dimension reduction by using a kernel principal component dimension reduction method, and to construct a low-dimensional fault feature set for each frequency band reconstructed sub-sequence.

[0013] The integrated learning processing module is a Stacking integrated learning model, wherein the base learners include an SVM processing part, a KNN processing part, a GBDT processing part and an RF processing part; different base learners are selected for low-dimensional feature sets of different frequency bands, output results of each base learner are obtained by training, and the output structure is used as an input quantity, a GBDT model is used as a meta-learner for classification, and a final fault diagnosis result is output.

[0014] The data acquisition module, the optimal parameter acquisition module, the decomposition component acquisition module, the frequency band division module, the signal reconstruction frequency band, the fault feature extraction module and the integrated learning processing module are connected in sequence.

[0015] The further technical scheme of the present application is: the alternating current arc fault simulation and acquisition experiment platform comprises a power supply, a current sensor, a data acquisition device, an intelligent gateway, an upper computer, a load, two switches and an arc generator; the power supply, the current sensor, the first switch, the load and the second switch are sequentially connected to form a closed loop, the current sensor is electrically connected with the data acquisition device, the data acquisition device is electrically connected with the intelligent gateway, the intelligent gateway is electrically connected with the upper computer, and the arc generator is arranged between the current sensor and the load and is connected in parallel with the first switch; the first switch and the second switch are closed when normal state data of the load is collected, the first switch is opened and the second switch is closed when arc fault state data of the load is collected, a plurality of sets of normal state data and arc fault state data are obtained for each load, and a feature data set is constructed.

[0016] The further technical scheme of the present application is: the arc generator is composed of a flat-end fixed carbon rod electrode and a sharp-end movable copper rod electrode, and an electrode gap is simulated by slowly rotating a copper electrode adjuster to control the electrode gap to simulate arc drawing.

[0017] The further technical scheme of the present application is: the multi-domain fault features include 11 time domain features and 5 frequency domain features, respectively, namely average value, variance, effective value, square root amplitude value, peak-peak value, standard deviation, kurtosis, skewness, peak factor, pulse factor, waveform index, barycentric frequency, average frequency, root mean square frequency, frequency variance, frequency variance and frequency standard deviation.

[0018] The further technical scheme of the present application is: different base learners are selected for different frequency band low-dimensional feature sets, specifically: a support vector machine (SVM) model is used for low-frequency band low-dimensional data, a K nearest neighbor (KNN) algorithm model is used for low-medium frequency band low-dimensional data, a gradient boosting decision tree (GBDT) model is used for medium-high frequency band low-dimensional data, and a random forest (RF) model is used for high frequency band low-dimensional data.

[0019] Another technical scheme of the present application is: the method of the arc fault diagnosis system based on the IAO-VMD center frequency, comprising the following steps,

[0020] Step one, the data acquisition module acquires current signal data under the normal state and the arc fault state of the circuit.

[0021] Step two, the optimal parameter acquisition module adopts the VMD method to process the feature data of a group of randomly extracted arc fault states, optimizes the best parameter combination of VMD decomposition using the improved eagle algorithm, and outputs the final optimization result.

[0022] Step 3: The decomposition component acquisition module substitutes the optimized VMD decomposition optimal parameter combination into VMD, processes the characteristic data sets of the normal circuit state and the fault arc state, and obtains K IMF components of the characteristic data sets of the normal circuit state and the fault arc state and the center frequency corresponding to each component.

[0023] Step 4: The frequency band division module performs FFT frequency domain transformation on the frequency of the characteristic data set of the normal state and the fault arc state and divides it into four equal parts, namely low frequency band, medium-low frequency band, medium-high frequency band, and high frequency band. The frequency band to which each IMF component obtained in step 3 belongs is determined based on its center frequency, and the division result is output; in the signal reconstruction module, the IMF components in the same frequency band are superimposed and merged to obtain new low-frequency, medium-low frequency, medium-high frequency, and high frequency subsequences, and the reconstructed signal is output.

[0024] Step 5: In the fault feature extraction module, multi-domain fault features are extracted for each reconstructed subsequence, and the kernel principal component dimensionality reduction method is used to reduce the data dimension, and a low-dimensional fault feature set is constructed for each frequency band reconstructed subsequence.

[0025] Step 6. Repeat steps 4 and 5 until all feature data sets of the normal circuit state and the fault arc state are calculated; output low-dimensional feature data set groups of the normal circuit state and the fault arc state respectively, where each data set contains low-dimensional data of low frequency band, medium-low frequency band, medium-high frequency band, and high frequency band.

[0026] Step 7: In the integrated learning processing module, different base learners are selected for the low-dimensional feature sets of different frequency bands, and each base learner is trained to obtain the output result. The output structure is then used as the input, and the GBDT model is used as the meta-learner for classification to output the final diagnosis result.

[0027] A further technical solution of the present invention is: the optimal parameter combination of the improved Sky Eagle algorithm for optimizing VMD decomposition is specifically to introduce an Archimedean screw mechanism to strengthen the Sky Eagle algorithm when performing shrinking search and shrinking development.

[0028] A further technical solution of the present invention is: the data sampling frequency in the step 4 is 8000 Hz, the frequency after FFT frequency domain transformation is 4000 Hz, which is divided into four equal parts to obtain a low frequency band of [0, 1000 Hz], a mid-low frequency band of [1000 Hz, 2000 Hz], a mid-high frequency band of [2000 Hz, 3000 Hz], and a high frequency band of [3000 Hz, 4000 Hz].

[0029] Compared with the prior art, the present invention has the following characteristics:

[0030] (1) The application superimposes and combines according to the center frequency of the IMF component decomposed by VMD, reconstructs into new low-frequency, medium-low-frequency, medium-high-frequency and high-frequency sub-sequences, and the reconstructed signal can better reflect the effective fault characteristics of each frequency band, thereby effectively improving the precision of fault diagnosis.

[0031] (2) The application avoids the contingency and inaccuracy of single features by extracting multi-domain fault features, and can effectively and accurately extract arc fault features.

[0032] (3) The data dimension reduction method of the application avoids the problems of too large data set calculation scale and rising redundancy caused by correlation between features and too many features, and reduces the complexity of calculation.

[0033] (4) The application adopts an integrated learning model for fault diagnosis, has good generalization performance, and has remarkable fault recognition effect, and has higher recognition precision than a single model, thereby providing a new idea for improvement and research of the alternating series arc fault diagnosis algorithm.

[0034] The detailed structure of the application will be further described in combination with the drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0035] FIG. 1 is a schematic diagram of the system structure of the DC arc fault diagnosis system of the application; Figure 1 FIG. 2 is a schematic diagram of the alternating arc fault simulation and acquisition experimental platform structure of the application;

[0036] Figure 2 FIG. 3 is an optimization flowchart of the improved eagle algorithm of the application;

[0037] FIG. 4 is a VMD decomposition diagram and a frequency band division diagram of the optimal parameters of the application; Figure 3 FIG. 5 is a signal reconstruction diagram of the application;

[0038] Figure 4 FIG. 6 is a comparison diagram of the predicted value and the actual value of the arc fault of the method of the application.

[0039] FIG. 7 is a comparison diagram of the predicted value and the actual value of the arc fault of the method of the application. Figure 5 FIG. 8 is a comparison diagram of the predicted value and the actual value of the arc fault of the method of the application.

[0040] Figure 6 FIG. 9 is a comparison diagram of the predicted value and the actual value of the arc fault of the method of the application. DETAILED DESCRIPTION

[0041] Example 1, as shown in FIG. 1, the arc fault diagnosis system based on IAO-VMD center frequency includes a data acquisition module, an optimal parameter acquisition module, a decomposition component acquisition module, a frequency band division module, a signal reconstruction frequency band, a fault feature extraction module, and an integrated learning processing module. Figures 1-6

[0042] ​​​​The data acquisition module is used to select the current signal data in the existing circuit normal state and arc fault state, or to obtain the current signal data in the simulated circuit normal state and arc fault state through the alternating current arc fault simulation and collection experiment platform.

[0043] Due to various interference information existing in the actual power distribution line, it is difficult to accurately obtain the current waveform of the series arc fault, therefore, the data acquisition module is the alternating current arc fault simulation and collection experiment platform. The alternating current arc fault simulation and collection experiment platform comprises a power supply 1, a current sensor 2, a data collection device 3, an intelligent gateway 4, an upper computer 5, a load 6, two switches 7 and an arc generator 8. The power supply 1 is a 220V / 50Hz alternating current power supply, and the sampling frequency is set to 8000Hz. The arc generator 8 is composed of a flat-end fixed carbon rod electrode and a sharp-end moving copper rod electrode, and the electrode gap is simulated to draw arc by slowly rotating the copper electrode regulator. The load 6 is selected as four kinds of typical household loads, which are LED, electric hair dryer, electric kettle and notebook computer respectively, and the parameter information of each load is shown in Table 1.

[0044] Table 1 Parameter information of each load

[0045]

[0046] The power supply 1, the current sensor 2, the first switch 7, the load 6 and the second switch 7 are connected in sequence to form a closed loop, the current sensor 2 is electrically connected with the data collection device 3, the data collection device 3 is electrically connected with the intelligent gateway 4, the intelligent gateway 4 is electrically connected with the upper computer 5, and the arc generator 8 is arranged between the current sensor 2 and the load 6 and is connected in parallel with the first switch 7. The current waveform in the circuit loop is collected and transmitted by the data collection device 3 after passing through the current sensor 2 and is stored in the SD card of the intelligent gateway 4, and the intelligent gateway 4 transmits the data to the software platform of the upper computer 5 in real time for viewing and exporting.

[0047] The alternating current arc fault simulation and collection experiment platform acquires the load 6 data, closes the first switch 7 and the second switch 7 when collecting the normal state data of the load 6, opens the first switch 7 and closes the second switch 7 when collecting the arc fault state data of the load 6, acquires multiple sets of normal state data and arc fault state data for each load 6 respectively, and constructs a feature data set.

[0048] The optimal parameter acquisition module is used to process the feature data of a group of randomly extracted fault arc states by using the VMD method, and to optimize the best parameter combination [K, a] of VMD decomposition by using the improved eagle algorithm, and to output the final optimization result.

[0049] The decomposition component acquisition module is configured to input the final optimized VMD decomposition optimal parameter combination [K, a] into VMD, process the characteristic data sets of the normal state and the arc fault state of the circuit, and obtain K IMF components of the characteristic data sets of the normal state and the arc fault state of the circuit and the center frequencies ω corresponding to the components k .

[0050] The frequency band division module is configured to divide the characteristic data sets of the normal state and the arc fault state into four equal parts after performing FFT frequency domain conversion on the characteristic data sets, and determine the frequency bands to which the IMF components belong according to the center frequencies of the IMF components.

[0051] The signal reconstruction module is configured to superimpose and combine the IMF components in the same frequency band to obtain new low-frequency, medium-low-frequency, medium-high-frequency and high-frequency sub-sequences, and output the reconstructed signal.

[0052] The fault feature extraction module is configured to extract multi-domain fault features from each reconstructed sub-sequence, perform data dimension reduction on the multi-domain fault features by using a kernel principal component dimension reduction method, and construct a low-dimensional fault feature set for each frequency band reconstructed sub-sequence. The multi-domain fault features include 11 time domain features and 5 frequency domain features, which are average value, variance, effective value, square root amplitude value, peak-to-peak value, standard deviation, kurtosis, skewness, peak factor, pulse factor, waveform index, center frequency, average frequency, root mean square frequency, frequency variance, frequency variance and frequency standard deviation.

[0053] The integrated learning processing module is configured to select different base learners for the low-dimensional feature sets of different frequency bands, train each base learner to obtain an output result, and then use the GBDT model as a meta-learner to classify the output structure as an input quantity, and output the final fault diagnosis result. The different base learners selected for the low-dimensional feature sets of different frequency bands are as follows: the low-frequency band low-dimensional data uses a support vector machine (SVM) model, the medium-low-frequency band low-dimensional data uses a K-nearest neighbor (KNN) model, the medium-high-frequency band low-dimensional data uses a gradient boosting decision tree (GBDT) model, and the high-frequency band low-dimensional data uses a random forest (RF) model.

[0054] The data acquisition module, the optimal parameter acquisition module, the decomposition component acquisition module, the frequency band division module, the signal reconstruction module, the fault feature extraction module and the integrated learning processing module are sequentially connected, the integrated learning processing module is a Stacking integrated learning model, the base learners include an SVM processing part, a KNN processing part, a GBDT processing part and an RF processing part, and the meta-learner uses the GBDT processing part.

[0055] Embodiment two, a method applied to the arc fault diagnosis system based on the IAO-VMD center frequency, includes the following steps:

[0056] Step one, obtain the current signal data under the normal state and arc fault state of the circuit.

[0057] The data acquisition module selects the current signal data under the normal state and arc fault state of the existing circuit; or obtains the current signal data under the normal state and arc fault state of the simulated circuit through the alternating current arc fault simulation and acquisition experiment platform, and constructs the feature data set of the normal state and fault arc state of the circuit.

[0058] Step two, the VMD method is used to process a group of feature data of fault arc state randomly extracted in the optimal parameter acquisition module, the improved eagle algorithm is used to optimize the best parameter combination [K, a] of VMD decomposition, and the final optimization result is output, which specifically includes the following:

[0059] Step 2.1, VMD is used to decompose the signal, mainly including constructing a variational problem and solving the variational problem.

[0060] (1) Constructing a variational problem

[0061] The feature data of the extracted fault arc state is decomposed into K IMF components, and the formula of the kth IMF component u k (t) is:

[0062] u k (t) = A k (t) cos [φ k (t)], k ∈ {1, L, K} (1)

[0063] In the formula, the phase φ k (t) is a non-decreasing function, and φ′ k (t) 30; A k (t) is an envelope function, and A k (t) 30.

[0064] The solving formula of the constructed constrained variational problem is:

[0065]

[0066] In the formula, t is time; j is the imaginary unit; ω k is the center frequency of the kth IMF component, k = 1, 2, L K, and K is the maximum decomposition number; δ(t) is the impact function.

[0067] (2) Solving the variational problem

[0068] In formula (2), α and λ(t) are introduced, which converts the variational problem into an unconstrained variational problem, and the following formula is obtained:

[0069]

[0070] Where: λ(t) is the Lagrange multiplier; α is the penalty factor; f(t) is the signal to be decomposed.

[0071] Solving the saddle point of Equation (3) is equivalent to solving the optimal value of Equation (2). The main steps are:

[0072] ①Yes n is initialized;

[0073] ②Execute the loop, n=n+1;

[0074] ③For all ω>0, update {k=1,2,LK}, the update formula is:

[0075]

[0076] ④Updateω k , the update formula is:

[0077]

[0078] ⑤Update λ, the update formula is:

[0079]

[0080] Repeat steps ② to ⑤ above and stop the iteration when the following conditions are met:

[0081]

[0082] Where, are the decomposed IMFs components, e is the convergence accuracy, and e>0.

[0083] From the above decomposition calculation process, we can know that when VMD is used to decompose the signal, the VMD decomposition parameters [K, α] have a great influence on the decomposition result. If the parameters are not selected properly, it is easy to cause modal aliasing, which affects the extraction of fault information. Therefore, it is necessary to optimize the VMD decomposition parameters [K, α]. In order to ensure the best decomposition effect, the minimum envelope entropy value of each IMF component decomposed by VMD is used as the fitness value.

[0084] In step 2.2, the improved Sky Eagle algorithm is used to optimize the VMD decomposition parameters [K, α]. The specific process is as follows:

[0085] First, initialize the parameters of the Sky Eagle algorithm, including setting the current number of iterations t = 1, setting the maximum number of iterations T; and generating the fitness value of the Sky Eagle algorithm and the position X of the randomly initialized population within the search range. ij , the formula is as follows:

[0086] X ij= rand x (UB j - LB j + LB j i = 1, 2, L, N j = 1, 2, L, Dim (8)

[0087] where rand is a random vector with values in the range (0, 1); LB j is the jth lower bound of the solution, UB j is the jth upper bound of the solution; N is the population size; Dim is the dimension size. In this embodiment, the maximum number of iterations is set to 20, and the population size is set to 30.

[0088] Secondly, according to the relationship between the current iteration number t and the maximum iteration number T, and the size of the random vector rand, different eagle optimization algorithms are selected, including: (1) expanding search: selecting the search space by high-altitude flight of vertical dive, (2) narrowing search: exploring in the divergent search space by contour flight of short glide attack, (3) expanding development: developing in the convergent search space by low-altitude flight of slow descent attack, (4) narrowing development: through the behavior of rushing and grabbing prey by foot.

[0089] The specific settings are as follows:

[0090] A, when the expanding search is performed, the search space is selected by high-altitude flight of vertical dive to determine the search space area where the prey is located, and the mathematical model of this behavior is as shown in the following formula:

[0091]

[0092]

[0093] where t is the current iteration number, T is the maximum iteration number; X1(t+1) is the solution of the t+1th iteration; X best (t) is the best solution obtained before the tth iteration, which reflects the approximate position of the prey; is used to control the expansion of the search by the iteration number; X M (t) is the average value of the current solution at the tth iteration; rand is a random value between 0 and 1.

[0094] B, when and rand>0.5, the narrowing search is performed, after determining the search space area where the prey is located, the search space is explored in the divergent search space by contour flight of short glide attack to prepare for the attack, and the mathematical model of this behavior is as shown in the following formula:

[0095] X2(t+1) = X best(t) x Levy(D) + X R (t) + (y - x) * rand (11)

[0096]

[0097]

[0098]

[0099]

[0100] where X2(t+1) is the solution of the t+1th iteration; D is the dimension space size; Levy(D) is the Levy flight distribution function, X R (t) is the random solution obtained in the range of [1, N] at the tth iteration; u and v are Gaussian distribution random numbers subject to N(0, σ 2 ) and N(0, 1), respectively; r1 is a value between 1 and 20, used to fix the number of search periods; 1≤D1≤Dim, s=0.01, β=1.5, U=0.00565, ω=0.005.

[0101] The Levy flight distribution function is used by the algorithm in performing the reduced search, however, Levy flight only plays a role when the individual optimization stagnates, and it is difficult to meet the overall local search process of the algorithm, therefore, the improved algorithm introduces the Archimedes spiral mechanism to further improve the local search ability of the AO algorithm.

[0102] The Archimedes spiral is a motion trajectory generated by a point moving away from a fixed point at a constant speed and rotating around the fixed point at a constant angular velocity, and its polar coordinate expression is:

[0103] r=a+bθ (16)

[0104] where a is the distance from the initial point to the center of the polar coordinate, b is the distance between the spirals, and θ is the polar angle.

[0105] The Levy flight strategy is combined with the Archimedes spiral to improve the local search ability of the algorithm, and the mathematical model of its behavior is shown in the following formula:

[0106] X2(t+1)=X best (t) + |X best (t) - X Levy (t) | lcos(2πl) (17)

[0107]

[0108] where X Levy(t) represents the generated Levy flight solution; l is a random number between [-1, 1]. The Levy flight strategy is integrated into the Archimedean spiral mechanism to search for a local solution, which not only ensures the rigor and accuracy of the algorithm optimization process and enhances the local search capability, but also improves the population diversity of the algorithm in the later iteration to avoid the premature phenomenon, and optimizes the optimization accuracy and convergence speed of the algorithm.

[0109] C, when and rand < 0.5, the expansion development is performed, the selected area of the target prey is determined, and the prey is approached by slow descent in the selected area, which is mathematically shown as follows:

[0110] X3(t+1) = (X best (t) - X M (t)) x a1-rand + ((UB-LB) x rand+LB) x d (19)

[0111] In the formula, X3(t+1) is the solution of the t+1 iteration; a1 and d are adjustment parameters, and the values are fixed as 0.1; LB is the lower limit of the optimization parameter, and UB is the upper limit of the optimization parameter.

[0112] D, when and rand > 0.5, the reduction development is performed, when approaching the prey, the prey is grabbed by walking and attacking on land according to random motion, and the mathematical model of the behavior is shown as follows:

[0113] X4(t+1) = QF(t) x X best (t) - (G1 x X(t) x rand) - G2 x Levy(D) + rand x G1 (20)

[0114]

[0115] G1 = 2 x rand - 1 (22)

[0116]

[0117] In the formula, X4(t+1) is the solution of the t+1 iteration; QF(t) is a quality function used to balance the search strategy at the t iteration; G1 is various movements used to track the prey during the search for the prey; G2 is a flight speed, and the value is a decreasing value from 2 to 0; X(t) is the current solution of the t iteration.

[0118] The algorithm adopts Levy flight distribution function in the process of downsizing development, however, Levy flight only plays a role when the individual optimization stagnates, and it is difficult to meet the local search process of the overall algorithm, therefore, the improved algorithm introduces Archimedes spiral mechanism to further improve the local search ability of the algorithm, specifically:

[0119] Archimedes spiral is a motion trajectory generated by a point moving away from a fixed point at a constant speed and rotating around the fixed point at a constant angular velocity, and its polar coordinate expression is:

[0120] r=a+bθ (24)

[0121] In the formula, a is the distance from the initial point to the center of the polar coordinate, b is the distance between the spirals, and θ is the polar angle.

[0122] The Levy flight strategy is combined with Archimedes spiral to improve the local search ability of the algorithm, and the mathematical model of the behavior is as follows:

[0123] X4(t+1)=X best (t)+|X best (t)-X Levy (t)|lcos(2πl) (25)

[0124]

[0125] In the formula, X Levy (t) represents the generated Levy flight solution; l is a random number between [-1, 1]. The Levy flight strategy is integrated into the Archimedes spiral mechanism to search for local solutions, which ensures the rigor and accuracy of the optimization process, enhances the local search ability, and improves the population diversity of the algorithm in the later iteration to avoid premature phenomena, and optimizes the optimization accuracy and convergence speed of the algorithm.

[0126] Finally, it is judged whether the fitness value is improved, that is, the new fitness value calculated by the position of the new population is smaller than the last fitness value, when the fitness is not improved, the position of the hawk is updated according to the above iteration formula; when the fitness value is improved and the maximum iteration number is not reached, the fitness value of the algorithm is updated; when the fitness is improved and the maximum iteration number is reached, the global minimum envelope entropy is obtained, and the optimal parameters [K, α] obtained by the optimization of the algorithm are output.

[0127] Step three, the decomposition component acquisition module substitutes the optimal parameter combination [K, α] of the VMD optimization search into the VMD, processes the feature data sets of the normal state and fault arc state of the circuit, and obtains K IMF components of the feature data sets of the normal state and fault arc state of the circuit and the corresponding center frequency ω k .

[0128] Step four, the frequency band division module divides the characteristic data set frequency of the normal state and the arc fault state into four equal parts after FFT frequency domain transformation, which are low frequency band, medium-low frequency band, medium-high frequency band, and high frequency band. The center frequency of each IMF component obtained in step three is used to determine which frequency band it belongs to, and the division result is output. The signal reconstruction module superimposes and combines the IMF components in the same frequency band to obtain new low frequency, medium-low frequency, medium-high frequency, and high frequency subsequences, and outputs the reconstructed signal.

[0129] In this embodiment, the data sampling frequency is set to 8000 Hz, and the frequency after FFT frequency domain transformation is 4000 Hz. After four equal division, the low frequency band is [0, 1000 Hz], the medium-low frequency band is [1000 Hz, 2000 Hz], the medium-high frequency band is [2000 Hz, 3000 Hz], and the high frequency band is [3000 Hz, 4000 Hz], as shown in FIG. 2. The center frequency is used for division, and the division result is output. The IMF components in the same frequency band are superimposed and combined to obtain new low frequency S1, medium-low frequency S2, medium-high frequency S3, and high frequency S4 subsequences, as shown in Table 2 below. The reconstructed signal in the signal reconstruction module is shown in FIG. 3. Figure 4 Figure 5

[0130] Table 2 Center frequency of each IMF component and reconstructed sequence

[0131]

[0132] Step five, the multi-domain fault feature extraction module extracts multi-domain fault features from each reconstructed subsequence. The kernel principal component dimension reduction method is used for data dimension reduction, and a low-dimensional fault feature set is constructed for each frequency band reconstructed subsequence.

[0133] The multi-domain fault features include 11 time domain features and 5 frequency domain features, and a 16-dimensional fault feature matrix is constructed. By extracting multi-domain fault features, arc fault information can be more comprehensively reflected. Compared with algorithms established by single features, the method has stronger universality and is suitable for various load types. The definition and calculation formula of each fault feature are shown in Table 3 below:

[0134] Table 3 Definition formula of feature index

[0135]

[0136] The calculation process of the kernel principal component dimension reduction method is as follows:

[0137] Suppose D (D >> d) dimensional vector w i (i = 1,...,d) is the characteristic vector in the high-dimensional space, and λ i ​​(i=1,...,d) is the corresponding eigenvalue, and the principal component analysis in high-dimensional space is as follows:

[0138] Φ(X)Φ(X) T w i =λ i w i (27)

[0139] The eigenvector w i (i=1,...,d) is linearly represented by the sample set Φ(X), as follows:

[0140]

[0141] Will w i (i=1,...,d) Substitute into

[0142] Φ(X)Φ(X) T Φ(X)σ=λ i Φ(X)σ (29)

[0143] Multiply both sides of the equation by Φ(X) T , we get the following formula:

[0144] Φ(X)Φ(X) T Φ(X)σΦ(X) T =λ i Φ(X)σΦ(X) T (30)

[0145] Further replace it with the kernel matrix K1 and select the linear kernel function to obtain the following formula:

[0146] K1σ=λ i σ (31)

[0147] That is, find the eigenvectors corresponding to the largest eigenvalues ​​of K1. Since K1 is a symmetric matrix, the solution vectors are orthogonal to each other, thus obtaining the test sample X new The linear representation in this subspace, that is, the vector after dimensionality reduction, is:

[0148]

[0149] Step 6: Repeat steps 4 and 5 until all feature datasets for the normal circuit state and the arc fault state are calculated. Output low-dimensional feature datasets for the normal circuit state and the arc fault state, respectively. Each dataset contains low-dimensional data for the low-frequency band, the medium-low frequency band, the medium-high frequency band, and the high frequency band.

[0150] Step seven, different base learners are selected in the integrated learning processing module for different frequency bands of low-dimensional feature sets. The low-frequency band low-dimensional data uses a support vector machine (SVM) model, the low-medium frequency band low-dimensional data uses a K-nearest neighbor algorithm (KNN) model, the medium-high frequency band low-dimensional data uses a gradient boosting decision tree (GBDT) model, the high frequency band low-dimensional data uses a random forest (RF) model, and the output results of each base learner are obtained by training. The output structure is used as an input quantity, a GBDT model is used as a meta-learner for classification, and a final diagnosis result is output.

[0151] To make the diagnosis model have better performance, the base learner selects a model with excellent performance but different model principles. For the medium-high frequency and high frequency band low-dimensional data with complex fluctuation characteristics, the GBDT and RF models with relatively stronger learning ability are selected as the base learners. The meta-learner should select an algorithm with strong generalization ability, integrate the prediction advantages of the base learners, and therefore select the GBDT model with high prediction accuracy and good stability.

[0152] In this embodiment, each group of data in the normal state feature set is labeled "0", and each group of data in the arc fault state feature set is labeled "1". The parameters of each base learner are shown in Table 4. Different base learners are selected for each group of low-dimensional feature sets and the output results are compared with the labels before inputting. The accuracy of the same label is calculated as the arc fault diagnosis accuracy, as shown in the accompanying Figure 6

[0153] Table 4 Diagnosis model and hyperparameter setting

[0154]

[0155] As can be seen from the accompanying Figure 6 , only one group of data in the prediction set is predicted incorrectly, and all data in the test set is predicted correctly. Therefore, the accuracy and superiority of the arc fault diagnosis system and method based on the IAO-VMD center frequency are proved.​

Claims

1. The arc fault diagnosis system based on IAO-VMD center frequency is characterized by: It includes data acquisition module, optimal parameter acquisition module, decomposition component acquisition module, frequency band division module, signal reconstruction module, fault feature extraction module and integrated learning processing module; The data acquisition module is used to select the current signal data of the existing circuit in a normal state and in an arc fault state, or to obtain the current signal data of the simulated circuit in a normal state and in an arc fault state through an AC arc fault simulation and acquisition experimental platform; The optimal parameter acquisition module is used to process a set of randomly extracted characteristic data of the fault arc state using the VMD method, and use the improved Sky Eagle algorithm to optimize the optimal parameter combination of VMD decomposition, and output the final optimization result. The improved Sky Eagle algorithm optimizes the optimal parameter combination of VMD decomposition by introducing an Archimedean spiral mechanism to strengthen the Sky Eagle algorithm when performing reduced search and reduced development. The Sky Eagle algorithm uses the Levy flight distribution function when performing reduced development, and integrates the Levy flight strategy with the Archimedean spiral to improve the local search capability of the Sky Eagle algorithm. The mathematical model of its behavior is shown in the following formula: X(t+1)=X best (t)+|X best (t)-X Levy (t)|lcos(2πl) (1) Where, X best (t) is the best solution obtained before the tth iteration; X Levy (t) represents the generated Levy flight solution; l is a random number between [-1,1]; u and v are respectively the solutions that obey N(0, σ 2 ) and N(0,1) Gaussian distribution random numbers; β = 1.5; X R (t) is the random solution obtained in the range [1, N] at the tth iteration; The decomposition component acquisition module is used to substitute the final optimized VMD decomposition optimal parameter combination into the VMD, process the characteristic data sets of the normal state of the circuit and the fault arc state, and obtain K IMF components of the characteristic data sets of the normal state of the circuit and the fault arc state and the center frequency corresponding to each component; The frequency band division module is used to divide the frequency of the characteristic data set of the normal state and the fault arc state into four equal parts after performing FFT frequency domain transformation, and determine the frequency band to which each IMF component belongs according to its center frequency; The signal reconstruction module is used to superimpose and merge IMF components in the same frequency band to obtain new low-frequency, medium-low-frequency, medium-high-frequency, and high-frequency subsequences, and output a reconstructed signal; The fault feature extraction module is used to extract multi-domain fault features from each reconstructed subsequence, and use the kernel principal component dimensionality reduction method to perform data dimensionality reduction, and construct a low-dimensional fault feature set for each frequency band reconstructed subsequence; The ensemble learning processing module is a stacking ensemble learning model, in which the base learners include an SVM processing part, a KNN processing part, a GBDT processing part, and an RF processing part; it is used to select different base learners for low-dimensional feature sets of different frequency bands, train each base learner to obtain an output result, and then use the output structure as input, use the GBDT model as a meta-learner for classification, and output the final fault diagnosis result; The data acquisition module, the optimal parameter acquisition module, the decomposition component acquisition module, the frequency band division module, the signal reconstruction module, the fault feature extraction module and the integrated learning processing module are connected in sequence.

2. The arc fault diagnosis system based on IAO-VMD center frequency according to claim 1, characterized in that: The AC arc fault simulation and acquisition experimental platform includes a power supply, a current sensor, a data acquisition device, an intelligent gateway, a host computer, a load, two switches and an arc generator; the power supply, current sensor, first switch, load and second switch are connected in sequence to form a closed loop, the current sensor is electrically connected to the data acquisition device, the data acquisition device is electrically connected to the intelligent gateway, the intelligent gateway is electrically connected to the host computer, and the arc generator is arranged between the current sensor and the load and connected in parallel with the first switch; when collecting normal state data of the load, the first switch and the second switch are closed, and when collecting arc fault state data of the load, the first switch is opened and the second switch is closed, and multiple sets of normal state data and arc fault state data are obtained for each load to construct a feature data set.

3. The arc fault diagnosis system based on IAO-VMD center frequency according to claim 2, characterized in that: The arc generator is composed of a flat-end fixed carbon rod electrode and a tip movable copper rod electrode. The electrode gap is controlled by slowly rotating the copper pole regulator to simulate arcing. The load options are LED, hair dryer, electric kettle and notebook.

4. The arc fault diagnosis system based on IAO-VMD center frequency according to claim 1, characterized in that: The multi-domain fault features include 11 time domain features and 5 frequency domain features, namely, mean value, variance, effective value, root square amplitude, peak-to-peak value, standard deviation, kurtosis, skewness, peak factor, pulse factor, waveform index, center of gravity frequency, average frequency, root mean square frequency, frequency variance, frequency variance and frequency standard deviation.

5. The arc fault diagnosis system based on IAO-VMD center frequency according to claim 1, characterized in that: The specific selection of different base learners for low-dimensional feature sets in different frequency bands is: the support vector machine SVM model is used for low-dimensional data in the low-frequency band, the K-nearest neighbor algorithm KNN model is used for low-dimensional data in the medium-low frequency band, the gradient boosting decision tree GBDT model is used for low-dimensional data in the medium-high frequency band, and the random forest RF model is used for low-dimensional data in the high frequency band.

6. The method of the arc fault diagnosis system based on the IAO-VMD center frequency according to any one of claims 1 to 5, characterized in that: The following steps are included: Step 1: The data acquisition module acquires current signal data in a normal circuit state and an arc fault state; Step 2: In the optimal parameter acquisition module, the VMD method is used to process a set of randomly selected characteristic data of the fault arc state, and the improved Sky Eagle algorithm is used to optimize the optimal parameter combination of VMD decomposition, and the final optimization result is output; Step 3: The decomposition component acquisition module substitutes the optimized searched VMD decomposition optimal parameter combination into VMD, processes the characteristic data sets of the circuit normal state and the fault arc state, and obtains K IMF components of the characteristic data sets of the circuit normal state and the fault arc state and the corresponding center frequency of each component; Step 4: The frequency band division module performs FFT frequency domain transformation on the frequency of the characteristic data set of the normal state and the fault arc state and divides it into four equal parts, namely low frequency band, medium-low frequency band, medium-high frequency band, and high frequency band. The frequency band to which each IMF component obtained in step 3 belongs is determined based on its center frequency, and the division result is output; in the signal reconstruction module, the IMF components in the same frequency band are superimposed and merged to obtain new low-frequency, medium-low frequency, medium-high frequency, and high frequency subsequences, and the reconstructed signal is output; Step 5: In the fault feature extraction module, multi-domain fault features are extracted for each reconstructed subsequence, and the kernel principal component dimensionality reduction method is used to reduce the data dimension, and a low-dimensional fault feature set is constructed for each frequency band reconstructed subsequence. Step 6: Repeat steps 4 and 5 until all feature data sets of the normal circuit state and the fault arc state are calculated; output low-dimensional feature data sets of the normal circuit state and the fault arc state respectively, where each data set contains low-dimensional data of low frequency band, medium-low frequency band, medium-high frequency band, and high frequency band; Step 7: In the integrated learning processing module, different base learners are selected for the low-dimensional feature sets of different frequency bands, and each base learner is trained to obtain the output result. The output structure is then used as the input, and the GBDT model is used as the meta-learner for classification to output the final diagnosis result.

7. The method of the arc fault diagnosis system based on IAO-VMD center frequency according to claim 6, characterized in that: The data sampling frequency in step 4 is 8000 Hz, and the frequency after FFT frequency domain transformation is 4000 Hz. It is divided into four equal parts to obtain a low frequency band of [0, 1000 Hz], a mid-low frequency band of [1000 Hz, 2000 Hz], a mid-high frequency band of [2000 Hz, 3000 Hz], and a high frequency band of [3000 Hz, 4000 Hz].

Images