Direct current fault arc identification method and device based on wavelet transform algorithm

Abstract

The invention discloses a DC fault arc identification method and device based on a wavelet transform algorithm, and the method comprises the steps: collecting a DC line current signal through a high-bandwidth current sensor, carrying out the digitalization of a signal conditioning circuit and a high-speed ADC, and sequentially carrying out the preprocessing of the elimination of baseline drift through differential filtering and the suppression of specific frequency interference through adaptive notch; performing 3-8-layer multi-scale decomposition on the pure signal by adopting a Daubechies wavelet basis, preferably selecting high-frequency detail coefficients from the first layer to the third layer, and creatively fusing a multi-scale energy entropy representing energy distribution randomness and a time domain kurtosis representing transient impact to form a composite characteristic quantity; a threshold value (threshold value = mean value + adaptive coefficient * standard deviation) is dynamically generated based on sliding time window historical data, the adaptive coefficient is dynamically adjusted according to the load type, the resistive load is 1.8-2.2, and the inductive / capacitive load is 2.5-3.5; and when the composite characteristic quantity exceeds a threshold value and the sample proportion is greater than or equal to 85% within the duration of 0.5-5ms, a tripping execution mechanism is triggered.

Description

Technical Field

[0001] The present invention belongs to the field of DC fault arc identification, and in particular relates to a DC fault arc identification method and device based on a wavelet transform algorithm. Background Art

[0002] DC arc fault detection devices are key protection equipment for ensuring the safe operation of photovoltaic power generation systems, electric vehicle charging facilities, and DC microgrids. Their core function is to monitor DC line current in real time, identify arc faults with extremely high fire risks caused by line aging, insulation damage, or loose connections, and quickly disconnect the circuit before the arc continues to develop.

[0003] Typical DC arc detection devices typically consist of a current sensor, signal conditioning circuitry, a feature extraction module (often based on frequency-domain analysis using Fast Fourier Transforms (FFTs) or simple time-domain amplitude detection), and a fixed-threshold comparator. When the extracted feature (such as the energy in a specific frequency band or the current rate of change) exceeds a preset threshold, a trip signal is triggered. However, this traditional structure presents significant shortcomings in addressing complex and diverse real-world application scenarios.

[0004] The shortcomings of existing technologies are mainly reflected in three aspects:

[0005] (1) The fixed threshold strategy has poor adaptability: DC system load types are diverse (resistive, inductive, capacitive and their combinations), and the normal operating current fluctuation range is large. It is difficult for a fixed threshold to take into account different working conditions. If the threshold is set too low, it is easy to be affected by normal switching operations (such as DC-DC converter operation) or load switching interference, resulting in frequent false operations. If the threshold is set too high, the sensitivity to high resistive loads or small current arcs is insufficient, and there is a risk of missed judgment.

[0006] (2) Feature extraction is single and lacks robustness: The frequency domain analysis method relying on FFT has limited ability to capture the instantaneous high-frequency characteristics of the arc and is easily contaminated by background noise; while the simple time domain amplitude or rate of change characteristics are difficult to effectively distinguish between the arc and normal load fluctuations, resulting in a decrease in recognition accuracy.

[0007] (3) Lack of effective dynamic anti-noise mechanism: Existing devices have insufficient ability to suppress baseline drift (such as caused by MPPT tracking) and specific frequency interference (such as PWM harmonics). The lack or simplification of the pre-processing link reduces the signal-to-noise ratio of the feature quantity, further exacerbating the problems of misjudgment and missed judgment.

[0008] These inherent defects severely restrict the reliability and widespread application of existing DC arc detection devices. Summary of the Invention

[0009] Purpose of the invention: The present invention provides a DC fault arc identification method and device based on the wavelet transform algorithm, aiming to overcome the core defects of the traditional method, such as threshold rigidity, poor feature robustness and weak noise resistance, by constructing a composite feature quantity that integrates multi-scale energy entropy and time domain kurtosis, combined with a load-adaptive dynamic threshold model and a high-frequency detail coefficient optimization mechanism.

[0010] Technical solution: The present invention provides a DC fault arc identification method based on a wavelet transform algorithm, comprising:

[0011] Step 1: real-time acquisition of DC line current signals and preprocessing;

[0012] Step 2: Perform multi-scale wavelet decomposition on the pre-processed current signal to extract high-frequency detail coefficients at multiple scales;

[0013] Step 3: Based on the high-frequency detail coefficients of the selected scale layer, the multi-scale energy entropy representing the randomness of the signal energy distribution and the time domain kurtosis representing the signal impact characteristics are calculated, and the two are weighted and fused to form a composite feature quantity;

[0014] Step 4: Count the historical data of the composite feature quantity using a sliding time window, calculate the average value and dispersion of the composite feature quantity in the window in real time, and dynamically set the dynamic threshold of the arc based on the average value and dispersion;

[0015] Step 5: Within the preset time window, if the composite feature quantity continuously exceeds the dynamic threshold, and the proportion of the composite feature quantity exceeding the dynamic threshold to the total number of collected samples within the time window exceeds the proportion threshold, it is determined to be a fault arc.

[0016] Furthermore, in step 1, the preprocessing includes:

[0017] Eliminate baseline drift in current signals through differential filtering;

[0018] Adaptive notch filtering is used to suppress periodic interference of specific frequencies in the current signal.

[0019] Furthermore, in step 2, the multi-scale wavelet decomposition uses Daubechies wavelet basis function.

[0020] Furthermore, in step 3, the calculation of the multi-scale energy entropy includes:

[0021] Calculate the energy value of the high-frequency detail coefficient of each selected scale layer , the energy value is the sum of the squares of all detail coefficient amplitudes of this layer, and the formula is:

[0022] ;

[0023] in, Represents the i-th high-frequency detail coefficient of the k-th layer;

[0024] Calculate the sum of the energy values ​​of each selected scale layer to get the total energy , the formula is:

[0025] ;

[0026] Calculate the energy value of each selected scale layer Total energy proportion;

[0027] The information entropy is calculated based on the ratio to obtain the multi-scale energy entropy , the formula is:

[0028]

[0029] Where m is the number of selected scales;

[0030] The calculation of the time domain kurtosis includes:

[0031] Calculate the high-frequency detail coefficient sequence obtained in all selected scale layers, remove the DC component and obtain the target signal x[n]; set the sliding time window length to N, and calculate the average value of the target signal x[n] ,variance , the formula is:

[0032] ,

[0033] Calculate the fourth-order center distance , the formula is:

[0034]

[0035] Calculate time domain kurtosis , the formula is:

[0036]

[0037] The multi-scale energy entropy and time domain kurtosis weighted fusion form a composite feature quantity, including:

[0038] Multiscale energy entropy and temporal kurtosis Perform linear normalization and Z-score standardization respectively to obtain and ;

[0039] Based on the historical data in the sliding time window, the variance of multi-scale energy entropy and time domain kurtosis is dynamically calculated, and the weight coefficient is adaptively determined according to the "inverse variance" principle. , , the formula is:

[0040] ,

[0041] in, is the variance of the multi-scale energy entropy in the sliding time window, is the variance of the temporal kurtosis within the sliding time window;

[0042] Linear combination is used to form composite feature quantities, the formula is:

[0043] .

[0044] Furthermore, in step 4, the dynamic threshold is set as follows:

[0045] The dynamic threshold is equal to the average value of the composite feature quantity in the current sliding time window plus a floating amount. The floating amount is obtained by multiplying the discrete degree of the composite feature quantity in the current sliding time window by an adjustable adaptive coefficient. The dynamic threshold formula is:

[0046]

[0047] in, is the average value of the composite feature quantity in the current sliding time window, is the standard deviation of the composite feature quantity in the current sliding time window, is the adaptive coefficient.

[0048] The present invention also provides a DC fault arc identification device based on a wavelet transform algorithm, comprising:

[0049] Signal processing module, used to collect the current signal of the DC line in real time and perform preprocessing;

[0050] The wavelet decomposition module is used to perform multi-scale wavelet decomposition on the pre-processed current signal and extract high-frequency detail coefficients at multiple scales;

[0051] The feature calculation module is used to calculate the multi-scale energy entropy that characterizes the randomness of the signal energy distribution and the time domain kurtosis that characterizes the signal impact characteristics based on the high-frequency detail coefficients of the selected scale layer, and to weightedly fuse the two to form a composite feature quantity;

[0052] A threshold setting module is used to use a sliding time window to count the historical data of the composite feature quantity, calculate the average value and dispersion of the composite feature quantity in the window in real time, and dynamically set the dynamic threshold of the arc based on the average value and dispersion;

[0053] The fault judgment module determines that it is an arc fault if the composite feature quantity continuously exceeds the dynamic threshold within a preset time window, and the proportion of the composite feature quantity exceeding the dynamic threshold to the total number of collected samples within the time window exceeds the proportion threshold.

[0054] Furthermore, in the signal processing module, the preprocessing includes:

[0055] Eliminate baseline drift in current signals through differential filtering;

[0056] Adaptive notch filtering is used to suppress periodic interference of specific frequencies in the current signal.

[0057] Furthermore, in the wavelet decomposition module, the multi-scale wavelet decomposition adopts Daubechies wavelet basis function.

[0058] Furthermore, in the feature calculation module, the calculation of the multi-scale energy entropy includes:

[0059] Calculate the energy value of the high-frequency detail coefficient of each selected scale layer , the energy value is the sum of the squares of all detail coefficient amplitudes of this layer, and the formula is:

[0060] ;

[0061] in, Represents the i-th high-frequency detail coefficient of the k-th layer;

[0062] Calculate the sum of the energy values ​​of each selected scale layer to obtain the total energy , the formula is:

[0063] ;

[0064] Calculate the energy value of each selected scale layer Total energy proportion;

[0065] The information entropy is calculated based on the ratio to obtain the multi-scale energy entropy , the formula is:

[0066]

[0067] Where m is the number of selected scales;

[0068] The calculation of the time domain kurtosis includes:

[0069] Calculate the high-frequency detail coefficient sequence obtained in all selected scale layers, remove the DC component and obtain the target signal x[n]; set the sliding time window length to N, and calculate the average value of the target signal x[n] ,variance , the formula is:

[0070] ,

[0071] Calculate the fourth-order center distance , the formula is:

[0072]

[0073] Calculate time domain kurtosis , the formula is:

[0074]

[0075] The multi-scale energy entropy and time domain kurtosis weighted fusion form a composite feature quantity, including:

[0076] Multiscale energy entropy and temporal kurtosis Perform linear normalization and Z-score standardization respectively to obtain and ;

[0077] Based on the historical data in the sliding time window, the variance of multi-scale energy entropy and time domain kurtosis is dynamically calculated, and the weight coefficient is adaptively determined according to the "inverse variance" principle. , , the formula is:

[0078] ,

[0079] in, is the variance of the multi-scale energy entropy in the sliding time window, is the variance of the temporal kurtosis within the sliding time window;

[0080] Linear combination is used to form composite feature quantities, the formula is:

[0081] .

[0082] Furthermore, in the threshold setting module, the dynamic threshold is set as follows:

[0083] The dynamic threshold is equal to the average value of the composite feature quantity in the current sliding time window plus a floating amount. The floating amount is obtained by multiplying the discrete degree of the composite feature quantity in the current sliding time window by an adjustable adaptive coefficient. The dynamic threshold formula is:

[0084]

[0085] in, is the average value of the composite feature quantity in the current sliding time window, is the standard deviation of the composite feature quantity in the current sliding time window, is the adaptive coefficient.

[0086] Beneficial effect: The present invention provides a DC fault arc identification method and device based on the wavelet transform algorithm. The method collects current signals in real time and performs differential filtering to eliminate baseline drift and adaptive notch filtering to suppress specific frequency interference in sequence. The pure signal is decomposed into multi-scale wavelet and the 1st to 3rd layer high-frequency detail coefficients are optimized. The multi-scale energy entropy representing the randomness of energy distribution and the time domain kurtosis representing transient impact are integrated to form a composite feature. The sliding time window is combined with statistical historical data to dynamically generate a judgment threshold that is adaptively adjusted with the load characteristics (the resistive load takes a low coefficient of 1.8-2.2, and the inductive / capacitive load takes a high coefficient of 2.5-3.5). When the composite feature continuously exceeds the dynamic threshold within 0.5ms-5ms, and the proportion of the composite feature exceeding the dynamic threshold in the total number of collected samples in the time window is ≥85%, the fault arc is determined, which effectively solves the technical defects of the traditional method with a high misjudgment rate due to poor adaptability of the fixed threshold, insufficient sensitivity of the single frequency domain feature, and noise interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 This is a block diagram of the system hardware structure of the present invention;

[0088] Figure 2 This is a timing diagram of signal processing of the present invention;

[0089] Figure 3 This is a schematic diagram for comparing characteristic quantities of the present invention;

[0090] Figure 4 A schematic diagram of dynamic threshold adjustment according to the present invention;

[0091] Figure 5 This is a diagram of the module collaborative workflow of the present invention. DETAILED DESCRIPTION

[0092] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0093] Example 1

[0094] See also Figures 1 to 5 As shown, the present invention provides a DC fault arc identification method based on a wavelet transform algorithm, comprising:

[0095] Step 1: real-time acquisition of DC line current signals and preprocessing;

[0096] Step 2: Perform multi-scale wavelet decomposition on the pre-processed current signal to extract high-frequency detail coefficients at multiple scales;

[0097] Step 3: Based on the high-frequency detail coefficients of the selected scale layer, the multi-scale energy entropy representing the randomness of the signal energy distribution and the time domain kurtosis representing the signal impact characteristics are calculated, and the two are weighted and fused to form a composite feature quantity;

[0098] Step 4: Count the historical data of the composite feature quantity using a sliding time window, calculate the average value and dispersion of the composite feature quantity in the window in real time, and dynamically set the dynamic threshold of the arc based on the average value and dispersion;

[0099] Step 5: Within the preset time window, if the composite feature quantity continuously exceeds the dynamic threshold, and the proportion of the composite feature quantity exceeding the dynamic threshold to the total number of collected samples within the time window exceeds the proportion threshold, it is determined to be a fault arc.

[0100] In this embodiment, step 1 uses a Hall effect sensor to collect the DC line current signal in real time. The millivolt differential mode signal output by the sensor first enters a signal conditioning circuit consisting of a programmable operational amplifier and an anti-aliasing low-pass filter. The operational amplifier is responsible for amplifying the signal to the operating range of the analog-to-digital converter (ADC) and correcting the sensor zero point. The anti-aliasing filter performs second-order or third-order active filtering in the analog domain with a cutoff frequency of approximately 0.45 times the sampling rate, providing an amplitude roll-off of more than 40 dB above the Nyquist frequency, effectively suppressing aliasing noise.

[0101] The conditioned analog signal is sampled by the analog-to-digital converter (ADC) at a resolution of 14 bits or higher and a rate of approximately 100 kS / s before being fed into the field-programmable gate array (FPGA). The FPGA uses on-chip resources to perform real-time differential operations to quickly eliminate slow-varying baseline drift, such as maximum power point tracking (MPPT). The data stream is then transmitted to the digital signal processor (DSP), where an adaptive notch filter algorithm library is invoked to precisely suppress periodic interference, such as pulse-width modulation (PWM) harmonics.

[0102] The differential filter is implemented using an FPGA, capable of performing sliding differential calculations in microseconds. The adaptive notch filter, based on a DSP, utilizes an adaptive filtering algorithm library to update the notch frequency online. The algorithm hardware is deployed on a high-reliability printed circuit board (PCB), preferably constructed from high-glass-transition-temperature FR-4 or polyimide to meet high-temperature operating requirements. Critical signal traces are shielded with copper foil to effectively suppress electromagnetic interference.

[0103] In step 2, the multi-scale wavelet decomposition uses Daubechies wavelet basis function, the decomposition layers of the Daubechies wavelet basis function are 3 to 8 layers, and the high-frequency detail coefficients of the first to third layers after decomposition are selected for feature calculation in subsequent steps.

[0104] In step 3, the calculation of the multi-scale energy entropy includes:

[0105] Calculate the energy value of the detail coefficient of each selected scale layer. The energy value is the sum of the squares of the amplitudes of all detail coefficients of the layer. The formula is:

[0106] ;

[0107] in, Represents the i-th high-frequency detail coefficient of the k-th layer;

[0108] Calculate the sum of the energy values ​​of each selected scale layer to obtain the total energy. The formula is:

[0109] ;

[0110] Calculate the proportion of the energy value of each selected scale layer to the total energy;

[0111] The information entropy is calculated based on the ratio to obtain the multi-scale energy entropy , the formula is:

[0112]

[0113] Here, m is the number of selected scales.

[0114] The time domain kurtosis is used to quantify the transient impact characteristics of the signal. The higher the value, the more transient impact components there are in the signal. The calculation of the time domain kurtosis includes:

[0115] Calculate the high-frequency detail coefficient sequence obtained in all selected scale layers, remove the DC component and obtain the target signal x[n]; set the sliding time window length to N, and calculate the average value of the target signal x[n] ,variance , the formula is:

[0116] ,

[0117] Calculate the fourth-order center distance , the formula is:

[0118]

[0119] Calculate time domain kurtosis , the formula is:

[0120]

[0121] The kurtosis value of the Gaussian normal distribution is 3, which is a typical characteristic of a stationary random process. When the kurtosis value of the target signal deviates from 3, it indicates that there are non-stationary transient impact components in the signal.

[0122] In order to quantify the degree to which the signal deviates from stationary noise, excess kurtosis is introduced The formula is:

[0123]

[0124] when = 0, the distribution of the target signal is close to the Gaussian normal distribution, indicating that the target signal is stable; when When it deviates significantly from 0, it indicates that there are non-stationary transient impact components in the target signal.

[0125] The multi-scale energy entropy and time domain kurtosis weighted fusion form a composite feature quantity, including:

[0126] Multiscale energy entropy and temporal kurtosis Perform linear normalization and Z-score standardization respectively to obtain and ;

[0127] Based on the historical data in the sliding time window, the variance of multi-scale energy entropy and time domain kurtosis is dynamically calculated, and the weight coefficient is adaptively determined according to the "inverse variance" principle. , , the formula is:

[0128] ,

[0129] in, is the variance of the multi-scale energy entropy in the sliding time window, is the variance of the temporal kurtosis within the sliding time window;

[0130] Linear combination is used to form composite feature quantities, the formula is:

[0131] .

[0132] In step 4, the dynamic threshold is set as follows:

[0133] The dynamic threshold is equal to the average value of the composite feature quantity in the current sliding time window plus a floating amount. The floating amount is obtained by multiplying the discrete degree of the composite feature quantity in the current sliding time window by an adjustable adaptive coefficient. The dynamic threshold formula is:

[0134]

[0135] in, is the average value of the composite feature quantity in the current sliding time window, is the standard deviation of the composite feature quantity in the current sliding time window, The adaptive coefficient ranges from 1.8 to 3.5 and is dynamically adjusted according to the DC line load type and rated current. For resistive loads, a lower value in the range of 1.8 to 2.2 is used. When the rated current is hour, ; When rated current hour, ; When rated current hour, For inductive or capacitive loads, take the higher value between 2.5 and 3.5. hour, ; When rated current hour, ; When rated current hour, .

[0136] In step 5, the fault determination introduces a time window and a ratio threshold. If the composite feature value continuously exceeds the dynamic threshold within 0.5ms-5ms, and the composite feature value exceeding the dynamic threshold accounts for ≥85% of the total number of samples collected during this time, it is determined to be an arc fault. This avoids misjudgments caused by transient interference and keeps the false alarm rate to an extremely low level. Fault isolation is achieved by a trip actuator, which uses a high-speed relay.

[0137] The wavelet transform algorithm and decision logic described in steps 2 to 5 are executed by a microprocessor, which uses an ARM Cortex-M7.

[0138] With respect to the above technical solution of this embodiment, the following describes in detail an experiment of applying the above technical solution to a photovoltaic power generation DC system, an electric vehicle DC charging system and a DC microgrid system.

[0139] The experiment uses Hall effect sensors to collect current signals from the DC lines of photovoltaic power generation systems, electric vehicle DC charging systems, and DC microgrid systems to ensure reliable acquisition of weak arc signals.

[0140] Applying a combination of differential filtering and adaptive notching to the current signal to improve the signal-to-noise ratio by approximately 12-15 dB;

[0141] The pre-processed current signal is subjected to multi-scale wavelet decomposition, and Daubechies wavelet basis function is used to decompose the signal into 3 to 8 layers. The high-frequency detail coefficients of the first to third layers after decomposition are selected to accurately capture the 1-10kHz high-frequency components generated by the arc.

[0142] Based on the high-frequency detail coefficients of the selected scale layer, the multi-scale energy entropy, which characterizes the randomness of the signal energy distribution, and the time-domain kurtosis, which characterizes the signal's impact characteristics, are calculated and fused to form a composite feature. The multi-scale energy entropy quantifies the randomness of the arc energy distribution, while the time-domain kurtosis highlights the impact characteristics of the arc. The composite feature formed by the fusion of the two has unique sensitivity to fault arcs, resulting in an area under the receiver operating characteristic (ROC) curve of 0.987.

[0143] A sliding time window is used to adaptively track load changes. A lower adaptive coefficient (1.8-2.2) is used for resistive loads to ensure detection sensitivity for small current arcs, while a higher coefficient (2.5-3.5) is used for inductive / capacitive loads to avoid misjudgment of switching transients. Historical data of the composite feature is collected, and the average value and dispersion of the composite feature within the window are calculated in real time. The dynamic threshold of the arc is dynamically set based on the average value and dispersion. Within a preset time window, if the composite feature continuously exceeds the dynamic threshold, and the proportion of composite features exceeding the dynamic threshold to the total number of samples collected within the time window exceeds the proportional threshold, the fault is determined to be an arc fault. Experiments have shown that the dual constraint mechanism of the time window and the proportional threshold reduces the false alarm rate from 23.7% of the traditional method to below 1.3%.

[0144] Experiments have shown that this method achieves arc detection accuracy of ≥99.2% in photovoltaic power generation systems, electric vehicle charging facilities, and DC microgrids, with a fault response time of ≤5ms, significantly exceeding the IEC62616 standard requirement of ≤100ms. The collaborative innovation of these modules effectively addresses the poor adaptability, low robustness of single features, and weak anti-interference capabilities of traditional fixed-threshold algorithms, providing a reliable arc fault protection solution for DC systems.

[0145] Example 2

[0146] See also Figures 1 to 5 As shown, based on the first embodiment, the present invention provides a DC arc fault identification device based on a wavelet transform algorithm, comprising:

[0147] Signal processing module, used to collect the current signal of the DC line in real time and perform preprocessing;

[0148] The wavelet decomposition module is used to perform multi-scale wavelet decomposition on the pre-processed current signal and extract high-frequency detail coefficients at multiple scales;

[0149] The feature calculation module is used to calculate the multi-scale energy entropy that characterizes the randomness of the signal energy distribution and the time domain kurtosis that characterizes the signal impact characteristics based on the high-frequency detail coefficients of the selected scale layer, and to weightedly fuse the two to form a composite feature quantity;

[0150] A threshold setting module is used to use a sliding time window to count the historical data of the composite feature quantity, calculate the average value and dispersion of the composite feature quantity in the window in real time, and dynamically set the dynamic threshold of the arc based on the average value and dispersion;

[0151] The fault judgment module determines that it is an arc fault if the composite feature quantity continuously exceeds the dynamic threshold within a preset time window, and the proportion of the composite feature quantity exceeding the dynamic threshold to the total number of collected samples within the time window exceeds the proportion threshold.

[0152] In this embodiment, the signal processing module uses a Hall-effect sensor to collect the DC line current signal in real time. The millivolt differential-mode signal output by the sensor first enters a signal conditioning circuit consisting of a programmable operational amplifier and an anti-aliasing low-pass filter. The operational amplifier is responsible for amplifying the signal to the operating range of the analog-to-digital converter (ADC) and correcting the sensor zero point. The anti-aliasing filter performs second-order or third-order active filtering in the analog domain with a cutoff frequency of approximately 0.45 times the sampling rate, providing an amplitude roll-off of more than 40 dB above the Nyquist frequency, effectively suppressing aliasing noise.

[0153] The conditioned analog signal is sampled by the analog-to-digital converter (ADC) at a resolution of 14 bits or higher and a rate of approximately 100 kS / s before being fed into the field-programmable gate array (FPGA). The FPGA uses on-chip resources to perform real-time differential operations to quickly eliminate slow-varying baseline drift, such as maximum power point tracking (MPPT). The data stream is then transmitted to the digital signal processor (DSP), where an adaptive notch filter algorithm library is invoked to precisely suppress periodic interference, such as pulse-width modulation (PWM) harmonics.

[0154] The differential filter is implemented using an FPGA, capable of performing sliding differential calculations in microseconds. The adaptive notch filter, based on a DSP, utilizes an adaptive filtering algorithm library to update the notch frequency online. The algorithm hardware is deployed on a high-reliability printed circuit board (PCB), preferably constructed from high-glass-transition-temperature FR-4 or polyimide to meet high-temperature operating requirements. Critical signal traces are shielded with copper foil to effectively suppress electromagnetic interference.

[0155] In the wavelet decomposition module, the multi-scale wavelet decomposition uses Daubechies wavelet basis function, the decomposition layers of the Daubechies wavelet basis function are 3 to 8 layers, and the high-frequency detail coefficients of the first to third layers after decomposition are selected for feature calculation in subsequent steps.

[0156] In the feature calculation module, the calculation of the multi-scale energy entropy includes:

[0157] Calculate the energy value of the detail coefficient of each selected scale layer. The energy value is the sum of the squares of the amplitudes of all detail coefficients of the layer. The formula is:

[0158] ;

[0159] in, Represents the i-th high-frequency detail coefficient of the k-th layer;

[0160] Calculate the sum of the energy values ​​of each selected scale layer to obtain the total energy. The formula is:

[0161] ;

[0162] Calculate the proportion of the energy value of each selected scale layer to the total energy;

[0163] The information entropy is calculated based on the ratio to obtain the multi-scale energy entropy , the formula is:

[0164]

[0165] Here, m is the number of selected scales.

[0166] The time domain kurtosis is used to quantify the transient impact characteristics of the signal. The higher the value, the more transient impact components there are in the signal. The calculation of the time domain kurtosis includes:

[0167] Calculate the high-frequency detail coefficient sequence obtained in all selected scale layers, remove the DC component and obtain the target signal x[n]; set the sliding time window length to N, and calculate the average value of the target signal x[n] ,variance , the formula is:

[0168] ,

[0169] Calculate the fourth-order center distance , the formula is:

[0170]

[0171] Calculate time domain kurtosis , the formula is:

[0172]

[0173] The kurtosis value of the Gaussian normal distribution is 3, which is a typical characteristic of a stationary random process. When the kurtosis value of the target signal deviates from 3, it indicates that there are non-stationary transient impact components in the signal.

[0174] In order to quantify the degree to which the signal deviates from stationary noise, excess kurtosis is introduced The formula is:

[0175]

[0176] when = 0, the distribution of the target signal is close to the Gaussian normal distribution, indicating that the target signal is stable; when When it deviates significantly from 0, it indicates that there are non-stationary transient impact components in the target signal.

[0177] The multi-scale energy entropy and time domain kurtosis weighted fusion form a composite feature quantity, including:

[0178] Multiscale energy entropy and temporal kurtosis Perform linear normalization and Z-score standardization respectively to obtain and ;

[0179] Based on the historical data in the sliding time window, the variance of multi-scale energy entropy and time domain kurtosis is dynamically calculated, and the weight coefficient is adaptively determined according to the "inverse variance" principle. , , the formula is:

[0180] ,

[0181] in, is the variance of the multi-scale energy entropy in the sliding time window, is the variance of the temporal kurtosis within the sliding time window;

[0182] Linear combination is used to form composite feature quantities, the formula is:

[0183] .

[0184] In the threshold setting module, the dynamic threshold is set as follows:

[0185] The dynamic threshold is equal to the average value of the composite feature quantity in the current sliding time window plus a floating amount. The floating amount is obtained by multiplying the discrete degree of the composite feature quantity in the current sliding time window by an adjustable adaptive coefficient. The dynamic threshold formula is:

[0186]

[0187] in, is the average value of the composite feature quantity in the current sliding time window, is the standard deviation of the composite feature quantity in the current sliding time window, The adaptive coefficient ranges from 1.8 to 3.5 and is dynamically adjusted according to the DC line load type and rated current. For resistive loads, a lower value in the range of 1.8 to 2.2 is used. When the rated current is hour, ; When rated current hour, ; When rated current hour, For inductive or capacitive loads, take the higher value between 2.5 and 3.5. hour, ; When rated current hour, ; When rated current hour, .

[0188] The fault determination module incorporates a time window and ratio threshold. If the composite feature value continuously exceeds the dynamic threshold within 0.5ms-5ms, and the composite feature value exceeding the dynamic threshold accounts for ≥85% of the total number of samples collected during that time, it is determined to be an arc fault. This avoids misjudgments caused by transient interference and keeps the false alarm rate to an extremely low level. Fault isolation is achieved through a trip actuator, which uses a high-speed relay.

[0189] The wavelet decomposition module, the feature calculation module, the threshold setting module and the fault judgment module execute the wavelet transform algorithm and the judgment logic through a microprocessor (such as ARMCortex-M4 / M7 series or DSP).

[0190] Since the second embodiment is a modular product expression of the DC arc fault identification method in the first embodiment, the experiments corresponding to the second embodiment are the same as those in the first embodiment and are not described again here.

Claims

1. A DC fault arc identification method based on wavelet transform algorithm, characterized in that: include: Step 1: real-time acquisition of DC line current signals and preprocessing; Step 2: Perform multi-scale wavelet decomposition on the pre-processed current signal to extract high-frequency detail coefficients at multiple scales; Step 3: Based on the high-frequency detail coefficients of the selected scale layer, the multi-scale energy entropy representing the randomness of the signal energy distribution and the time domain kurtosis representing the signal impact characteristics are calculated, and the two are weighted and fused to form a composite feature quantity; Step 4: Count the historical data of the composite feature quantity using a sliding time window, calculate the average value and dispersion of the composite feature quantity in the window in real time, and dynamically set the dynamic threshold of the arc based on the average value and dispersion; Step 5: Within the preset time window, if the composite feature quantity continuously exceeds the dynamic threshold, and the proportion of the composite feature quantity exceeding the dynamic threshold to the total number of collected samples within the time window exceeds the proportion threshold, it is determined to be a fault arc.

2. The DC arc fault identification method based on wavelet transform algorithm according to claim 1 is characterized in that: In step 1, the pretreatment includes: Eliminate baseline drift in current signals through differential filtering; Adaptive notch filtering is used to suppress periodic interference of specific frequencies in the current signal.

3. The DC arc fault identification method based on wavelet transform algorithm according to claim 1, characterized in that: In step 2, the multi-scale wavelet decomposition uses Daubechies wavelet basis functions.

4. The DC arc fault identification method based on wavelet transform algorithm according to claim 1, characterized in that: In step 3, the calculation of the multi-scale energy entropy includes: Calculate the energy value of the high-frequency detail coefficient of each selected scale layer , the energy value is the sum of the squares of all detail coefficient amplitudes of this layer, and the formula is: ； in, Represents the i-th high-frequency detail coefficient of the k-th layer; Calculate the sum of the energy values ​​of each selected scale layer to obtain the total energy , the formula is: ； Calculate the energy value of each selected scale layer Total energy proportion; The information entropy is calculated based on the ratio to obtain the multi-scale energy entropy , the formula is: ， Where m is the number of selected scales; The calculation of the time domain kurtosis includes: Calculate the high-frequency detail coefficient sequence obtained in all selected scale layers, remove the DC component and obtain the target signal x[n]; set the sliding time window length to N, and calculate the average value of the target signal x[n] ,variance , the formula is: , ， Calculate the fourth-order center distance , the formula is: ， Calculate time domain kurtosis , the formula is: ， The multi-scale energy entropy and time domain kurtosis weighted fusion form a composite feature quantity, including: Multiscale energy entropy and temporal kurtosis Perform linear normalization and Z-score standardization respectively to obtain and ; Based on the historical data in the sliding time window, the variance of multi-scale energy entropy and time domain kurtosis is dynamically calculated, and the weight coefficient is adaptively determined according to the "inverse variance" principle. , , the formula is: ， ， in, is the variance of the multi-scale energy entropy in the sliding time window, is the variance of the temporal kurtosis within the sliding time window; Linear combination is used to form composite feature quantities, the formula is: 。 5. The DC arc fault identification method based on wavelet transform algorithm according to claim 1, characterized in that: In step 4, the dynamic threshold is set as follows: The dynamic threshold is equal to the average value of the composite feature quantity in the current sliding time window plus a floating amount. The floating amount is obtained by multiplying the discrete degree of the composite feature quantity in the current sliding time window by an adjustable adaptive coefficient. The dynamic threshold formula is: ， in, is the average value of the composite feature quantity in the current sliding time window, is the standard deviation of the composite feature quantity in the current sliding time window, is the adaptive coefficient.

6. A DC arc fault identification device based on wavelet transform algorithm, characterized in that: include: Signal processing module, used to collect the current signal of the DC line in real time and perform preprocessing; The wavelet decomposition module is used to perform multi-scale wavelet decomposition on the pre-processed current signal and extract high-frequency detail coefficients at multiple scales; The feature calculation module is used to calculate the multi-scale energy entropy that characterizes the randomness of the signal energy distribution and the time domain kurtosis that characterizes the signal impact characteristics based on the high-frequency detail coefficients of the selected scale layer, and to weightedly fuse the two to form a composite feature quantity; A threshold setting module is used to use a sliding time window to count the historical data of the composite feature quantity, calculate the average value and dispersion of the composite feature quantity in the window in real time, and dynamically set the dynamic threshold of the arc based on the average value and dispersion; The fault judgment module determines that it is an arc fault if the composite feature quantity continuously exceeds the dynamic threshold within a preset time window, and the proportion of the composite feature quantity exceeding the dynamic threshold to the total number of collected samples within the time window exceeds the proportion threshold.

7. The DC arc fault identification device based on wavelet transform algorithm according to claim 6, characterized in that: In the signal processing module, the preprocessing includes: Eliminate baseline drift in current signals through differential filtering; Adaptive notch filtering is used to suppress periodic interference of specific frequencies in the current signal.

8. The DC arc fault identification device based on wavelet transform algorithm according to claim 6, characterized in that: In the wavelet decomposition module, the multi-scale wavelet decomposition adopts Daubechies wavelet basis function.

9. The DC arc fault identification device based on wavelet transform algorithm according to claim 6, characterized in that: In the feature calculation module, the calculation of the multi-scale energy entropy includes: Calculate the energy value of the high-frequency detail coefficient of each selected scale layer , the energy value is the sum of the squares of all detail coefficient amplitudes of this layer, and the formula is: ； in, Represents the i-th high-frequency detail coefficient of the k-th layer; Calculate the sum of the energy values ​​of each selected scale layer to get the total energy , the formula is: ； Calculate the energy value of each selected scale layer Total energy proportion; The information entropy is calculated based on the ratio to obtain the multi-scale energy entropy , the formula is: ， Where m is the number of selected scales; The calculation of the time domain kurtosis includes: Calculate the high-frequency detail coefficient sequence obtained in all selected scale layers, remove the DC component and obtain the target signal x[n]; set the sliding time window length to N, and calculate the average value of the target signal x[n] ,variance , the formula is: , ， Calculate the fourth-order center distance , the formula is: ， Calculate time domain kurtosis , the formula is: ， The multi-scale energy entropy and time domain kurtosis weighted fusion form a composite feature quantity, including: Multiscale energy entropy and temporal kurtosis Perform linear normalization and Z-score standardization respectively to obtain and ; Based on the historical data in the sliding time window, the variance of multi-scale energy entropy and time domain kurtosis is dynamically calculated, and the weight coefficient is adaptively determined according to the "inverse variance" principle. , , the formula is: ， ， in, is the variance of the multi-scale energy entropy in the sliding time window, is the variance of the temporal kurtosis within the sliding time window; Linear combination is used to form composite feature quantities, the formula is: 。 10. The DC arc fault identification device based on wavelet transform algorithm according to claim 6, characterized in that: In the threshold setting module, the dynamic threshold is set as follows: The dynamic threshold is equal to the average value of the composite feature quantity in the current sliding time window plus a floating amount. The floating amount is obtained by multiplying the discrete degree of the composite feature quantity in the current sliding time window by an adjustable adaptive coefficient. The dynamic threshold formula is: ， in, is the average value of the composite feature quantity in the current sliding time window, is the standard deviation of the composite feature quantity in the current sliding time window, is the adaptive coefficient.

Images