Improved stochastic resonance and generation of deductive dc arc fault dynamic detection method

Abstract

The application discloses a dynamic detection method for direct current fault arc based on improved stochastic resonance and generative deduction, which comprises the following steps: sampling a direct current loop current, and dynamically strengthening a weakened fault feature by using a parameter self-adaptive adaptive improved stochastic resonance model based on an asymmetric bistable potential function; and then deducing a generalization feature sample set through a generative adversarial network, and realizing detection of the fault arc by using a random forest classification and a multi-time window joint decision. The dynamic detection method for the direct current fault arc based on the improved stochastic resonance and the generative deduction solves the problems of insufficient detection reliability and insufficient sample coverage when the fault feature is weakened in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of DC power distribution and electrical fire protection technology, specifically involving an improved method for dynamic detection of DC fault arcs based on random resonance and generative deduction. Background Technology

[0002] Existing DC fault arc detection methods mostly utilize current amplitude, frequency domain energy, or time-frequency characteristics as fault indicators, and identify them through a stochastic resonance model with fixed parameters or a conventional classifier. While these methods are effective under standard operating conditions, in practical engineering, on the one hand, fault arc characteristics are often weakened by factors such as sampling links, propagation paths, load noise, or impedance networks, resulting in insufficient feature enhancement before and after the fault or difficulty in distinguishing them from normal conditions. On the other hand, limitations such as complex experimental system setup, high testing costs, and numerous operating condition combinations mean that measured samples typically only cover a limited number of typical operating conditions, making it difficult to obtain sufficient samples of edge and transitional operating conditions, thus affecting the generalization ability of subsequent identification models.

[0003] In existing technologies, singular value decomposition (SVD) is often used for signal filtering. This method can still effectively extract arc features even when the signal-to-noise ratio is as low as -15dB, improving the feature extraction accuracy by 12.6% compared to traditional wavelet thresholding denoising methods. Building on this, an algorithm combining asymptotic SVD and fast Fourier transform is proposed to address load noise interference. This method, through filtering techniques, achieves a detection accuracy of 96.18%, reducing the false alarm rate by 54.3% compared to traditional SVD methods. For the weak arc features caused by three types of loads (resistors, electronic loads, inverters) and aluminum electrode materials under high current ratings of 16-40A, a method based on Chirplet sparse representation to extract the time-frequency information of early weak fault arcs is proposed. Through hierarchical processing of time-frequency features, this method improves the accuracy by an average of 48.22% compared to existing methods. To address the complex interference caused by impedance networks, a high-impedance fault detection method based on arc variation trends and nonlinear least squares is proposed. This method achieves a detection accuracy of 94.7% under 1000Ω high-impedance fault conditions by fusing time-frequency features with artificial intelligence. Furthermore, a non-stationary time series analysis method based on sample entropy is proposed to handle fault arc current. Through hierarchical processing that integrates time-frequency features with artificial intelligence, its detection accuracy is improved by an average of 21.08% compared to existing features and classifiers.

[0004] Existing detection methods primarily rely on feature-driven artificial intelligence frameworks for improvement, often leading to a significant increase in computational complexity and high computational costs. On the other hand, traditional stochastic resonance enhancement methods typically depend on preset potential function parameters. When external operating conditions change, fixed parameters struggle to simultaneously address the anti-maloperation requirements in low-noise scenarios and the fault enhancement requirements in weak-feature scenarios, resulting in poor generalization ability and strong parameter dependence. Furthermore, DC fault arc testing systems are complex to build, with numerous combinations of load types, line impedances, arc-generating materials, and operating currents. Limited by testing costs, risks, and repeatability constraints, actual test samples typically cover only a limited range of operating conditions, making it difficult to obtain sufficient samples of marginal operating conditions. Directly building a recognition model based on such small samples can easily lead to insufficient adaptability to uncovered conditions, thus affecting the stability and generalization performance of the fault arc detection method. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive improved stochastic resonance and generative deduction method for dynamic detection of DC faults, which solves the problems of insufficient detection reliability and insufficient sample coverage in the prior art when fault characteristics are weakened.

[0006] The technical solution adopted in this invention is an adaptive improved stochastic resonance and generative deduction method for dynamic detection of DC fault arcs. This method includes sampling the DC circuit current and then using an adaptive improved stochastic resonance model based on an asymmetric bistable state function with adaptive parameters to dynamically enhance the weakened fault features. The method further uses a generative adversarial network to deduce the generalized feature sample set and utilizes random forest classification and multi-time window joint decision-making to detect the fault arc.

[0007] The technical solution of this invention is also characterized by, Specifically, the following steps are included: Step 1: Sample the DC circuit output current to obtain the discrete time-domain current signal during the analysis period; Step 2: Construct an adaptive improved stochastic resonance model based on an asymmetric bistable state function; input the discrete time-domain current signal into the model, and adaptively adjust the parameters of the asymmetric bistable state function according to the noise intensity and local energy changes represented by the discrete time-domain current signal to obtain the enhanced time-domain current signal; Step 3: Perform time-frequency decomposition on the enhanced time-domain current signal, extract time-domain and time-frequency domain features, and normalize the extracted features to obtain multi-dimensional time-frequency feature vectors under different operating conditions. Step 4: Calculate the Feature Enhancement Ratio (RF) and Overlap Relationship (NOR) for the multi-dimensional time-frequency feature vectors under different working conditions, and generalize the feature space based on machine learning to obtain an expanded feature sample set; train the machine learning algorithm model through the expanded feature sample set. Step 5: Input the real-time multi-dimensional time-frequency feature vector into the machine learning algorithm model, and the model outputs the fault identification result of the current analysis window; Step 6: Perform multi-window joint decision-making on the fault identification results of multiple consecutive analysis windows. When the preset threshold for the number of consecutive windows is met, output the fault arc determination result.

[0008] The DC system electrical signal sampled in step 1 is a sampled electrical signal that has not undergone sensor or high-pass digital filtering. The sampling steps are as follows: the DC system output current is sampled according to the sampling frequency. f s Point-by-point sampling is performed, with the number of sampling points per analysis window. q The discrete-time domain current signal within the analysis period is obtained. x q Number of sampling points q The sampling frequency is 8000~12000. f s The range is 200kHz to 1MHz, and the duration of a single analysis window is... q / f s .

[0009] In step 2, the adaptive improved stochastic resonance model is represented by a first-order dynamic equation, and a linear asymmetric term is introduced into the first-order dynamic equation to construct an asymmetric bistable potential function; the coefficients of the quadratic term of the potential function... a ( D The Sigmoid function is used to adjust the signal based on the intensity of external noise in order to match the changes in the intensity of external noise.

[0010] The first-order dynamic equation of the adaptive improved stochastic resonance model is expressed as: (1) In the formula, For output signal; The input is the effective current component to be enhanced; It is Gaussian white noise, where For noise intensity, Standard Gaussian white noise; Then, a linear asymmetric term is introduced to construct an asymmetric bistable state function: (2) In the formula, The coefficients of the quadratic terms of the fundamental topological structure of the dominant potential function; The coefficient of the fourth term; The coefficients of the linear asymmetric terms are obtained through the coefficients. This creates a depth difference between the left and right potential wells, matching the asymmetric characteristics of the electric arc. Secondly, the potential function parameters for continuous control of the Sigmoid function are introduced. The expression is: (3) In the formula, To limit the coefficient of the largest quadratic term; The noise intensity normalization fusion threshold; The steepness coefficient is used to smooth the continuous transition.

[0011] Step 2 involves adaptively adjusting the parameters of the asymmetric bistable state function based on the noise intensity and local energy changes represented by the discrete-time current signal. The first-order dynamic equations introduce adaptive parameter updates based on local energy changes, setting the current sliding window to the [number missing]. A window is defined, and the mean square value of the current signal within that window is calculated as the local energy of the signal. Then the adaptive adjustment factor Defined as: (4) According to the adaptive adjustment factor Dynamically update parameters and : (5) Get the updated parameters for the current time. , The output of the current data point is obtained by solving the fourth-order Runge-Kutta method; after the loop ends, the output sequence is the time-domain current signal after adaptive enhancement.

[0012] The fourth-order Longo-Kuta is represented as: (6) In the formula, For deterministic dynamic functions; The input is a discrete current signal; For the enhanced system output signal; For numerical calculation step size, η This is the noise term after discretization.

[0013] In step 3, the time-domain features include the average current before and after the fault arc and the rate of change of current after the fault. The time-frequency transformation method is wavelet transform. After multi-level wavelet decomposition using the Rbio3.1 antisymmetric bioorthogonal wavelet basis function, the wavelet energy calculated from the frequency band of 54.7kHz to 62.5kHz is reconstructed. The normalization process uses the max-min normalization method.

[0014] Feature enhancement ratio in step 4 RF The indicator is expressed as: (7) In the formula, M II This represents the average value of the characteristic quantities after the fault occurred. M I This represents the average value of the characteristic quantities under normal operating conditions; due to local large pulses in the signal, the characteristic boost ratio can be affected. RF If a large indicator value leads to feature confusion, then overlap needs to be introduced. NOR The indicators form a multi-dimensional evaluation system with weakened features, as shown in formula (8): (8) In the formula, OA This refers to the overlapping area of ​​characteristic quantities between normal and fault states. EA This refers to areas that are outside the normal range. F 1. F 2 is the corresponding i Wavelet energy at any given moment NOR A larger value indicates that the fault state is more likely to overlap with the normal state. Based on RF Indicators and NOR The indicators plot feature points for different working conditions. The selected machine learning method is generative adversarial network. Through adversarial training and deduction, an expanded feature sample set covering untested edge working conditions is obtained.

[0015] The machine learning algorithm model selected in step 5 is the random forest classification model, with the number of decision trees set to 100. The maximum number of features is the arithmetic square root of the total number of key frequency band features, and the Gini impurity is selected as the evaluation index for split nodes.

[0016] In step 6, a sliding time window decision queue with a fixed length of N is used to perform multi-time window joint decision on the fault identification results of multiple consecutive analysis time windows. N is set to 3, and the preset threshold for the number of consecutive windows is set to 3. After the fault arc judgment result is output, an audible and visual alarm signal is further output, or a trip command is sent to the circuit breaker to cut off the DC circuit when necessary.

[0017] The beneficial effects of this invention are: (1) The improved random resonance and generation derivation DC fault arc dynamic detection method provided by the present invention can reduce the dependence of traditional random resonance on fixed parameters and single working conditions and improve the generalization ability of weak fault feature enhancement by simultaneously introducing a potential function based on noise intensity and a parameter adaptive update mechanism based on local energy change in the random resonance model.

[0018] (2) The improved random resonance and generative derivation DC fault arc dynamic detection method provided by the present invention can transform the weakened fault features into a more stable key frequency band representation by connecting the adaptive improved random resonance enhancement with Rbio3.1 wavelet decomposition, thereby improving the separability between the normal state and the fault state.

[0019] (3) The improved stochastic resonance and generative derivation method for dynamic detection of DC fault arcs provided by this invention introduces... RF and NOR The feature weakening evaluation system composed of indicators, combined with GAN to extrapolate and generalize the feature space, can extend the model's adaptability to uncovered working conditions under limited measured sample conditions.

[0020] (4) The improved random resonance and generative deduction DC fault arc dynamic detection method provided by the present invention combines random forest classification with multi-time window joint decision-making, which can shorten the fault confirmation time while ensuring the anti-maloperation capability. After adopting the technical route of the present invention, the average detection time is 0.024s, which shows good engineering application potential. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the improved random resonance and generative derivation method for dynamic detection of DC fault arcs according to the present invention.

[0022] Figure 2 This is a schematic diagram illustrating the application of the improved random resonance and generative deduction method for dynamic detection of DC fault arcs in Embodiment 2 of the present invention.

[0023] Figure 3 This is a generalization derivation diagram of wavelet features under different operating conditions according to the present invention.

[0024] Figure 4 This is a waveform comparison diagram of the representative weakened and enhanced features of the present invention.

[0025] Figure 5 This is a diagram showing the real-time detection results without enhanced features.

[0026] Figure 6 This is a diagram showing the real-time detection results of the enhanced features of this invention. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0028] This invention provides an improved method for dynamic detection of DC fault arcs using stochastic resonance and generative inference. The method includes sampling the DC circuit current, then dynamically enhancing weakened fault features using an adaptive improved stochastic resonance model based on an asymmetric bistable state function with adaptive parameters; next, a generative adversarial network is used to infer a generalized feature sample set, and fault arc detection is achieved using random forest classification and multi-time-window joint decision-making. Figure 1 As shown, the specific steps are as follows: Step 1: Sample the DC circuit output current to obtain the discrete time-domain current signal during the analysis period; In step 1, the DC system electrical signal sampled is a sampled electrical signal that has not undergone sensor or high-pass digital filtering. The sampling steps are as follows: a YOKOGAWA 701930 current sensor is used to sample the DC circuit output current, and the sampling frequency is... fs The sampling points are between 200 kHz and 1 MHz, and the number of sampling points per analysis window. q The duration of a single analysis window is 8000–12000. T = q / fs Discrete-time domain current signal obtained by sampling xq Data collected under one or more of the following characteristic weakening factors: load noise, high current (e.g., 16–40 A), special arc-generating materials (e.g., aluminum electrodes), and photovoltaic impedance networks. The sampled data can be stored in the HIOKI8860-50 waveform recorder.

[0029] Step 2: Construct an adaptive improved stochastic resonance model based on an asymmetric bistable state function; input the discrete time-domain current signal into the model, and adaptively adjust the parameters of the asymmetric bistable state function according to the noise intensity and local energy changes represented by the discrete time-domain current signal to obtain the enhanced time-domain current signal; In step 2, the adaptive improved stochastic resonance model is represented by a first-order dynamic equation, and a linear asymmetric term is introduced into the first-order dynamic equation to construct an asymmetric bistable potential function; the coefficients of the quadratic term of the potential function... a ( D The Sigmoid function is used to adjust the signal based on the intensity of external noise in order to match the changes in the intensity of external noise.

[0030] The first-order dynamic equation of the adaptive improved stochastic resonance model is expressed as: (1) In the formula, For output signal; The input is the effective current component to be enhanced; It is Gaussian white noise, where For noise intensity, Standard Gaussian white noise; The asymmetric bistable state function constructed by introducing linear asymmetric terms into the first-order dynamic equation is expressed as: (2) In the formula, The coefficients of the quadratic terms of the fundamental topological structure of the dominant potential function; The coefficient of the fourth term; The coefficients of the linear asymmetric terms are obtained through the coefficients. This creates a depth difference between the left and right potential wells, matching the asymmetric characteristics of the electric arc. Introducing Sigmoid for continuously modulated potential function parameters The expression is: (3) In the formula, To limit the coefficient of the largest quadratic term; The noise intensity normalization fusion threshold; The steepness coefficient is used to smooth the continuous transition.

[0031] Step 2 involves adaptively adjusting the parameters of the asymmetric bistable state function based on the noise intensity and local energy changes represented by the discrete-time current signal. The first-order dynamic equations introduce adaptive parameter updates based on local energy changes, setting the current sliding window to the [number missing]. A window is defined, and the mean square value of the current signal within that window is calculated as the local energy of the signal. Then the adaptive adjustment factor Defined as: (4) According to the adaptive adjustment factor Dynamically update parameters and : (5) in, b 0 = 0.5 c 0 = 0.5. (Updated) b i , c i It reflects the intensity of changes in the local energy of the signal.

[0032] Get the updated parameters for the current time. , The output of the current data point is obtained by solving the fourth-order Runge-Kutta algorithm; after the loop ends, the resulting output sequence is the time-domain current signal after adaptive enhancement. The fourth-order Runge-Kutta algorithm is expressed as: (6) In the formula, For deterministic dynamic functions; The input is a discrete current signal; For the enhanced system output signal; For numerical calculation step size, η This is the noise term after discretization.

[0033] Step 3: Perform time-frequency decomposition on the enhanced time-domain current signal from Step 2, extract time-domain and time-frequency features, and normalize the extracted features to obtain multi-dimensional time-frequency feature vectors under different operating conditions. Specifically, the time-domain features include: the average current before and after the fault arc, and the rate of change of current after the fault (i.e., the average absolute value of the difference between adjacent sampling points). The time-frequency domain features are: the wavelet energy reconstructed from the 54.7kHz to 62.5kHz frequency band after multi-level wavelet decomposition based on the Rbio3.1 antisymmetric bioorthogonal wavelet basis functions. This frequency band is the key band reflecting the high-frequency harmonic energy of the fault arc. To eliminate the differences in reference amplitude caused by different power consumption conditions and dynamic load switching, a maximum-minimum normalization method is adopted.

[0034] Step 4: Calculate the Feature Enhancement Ratio (RF) and Overlap Relationship (NOR) for the multi-dimensional time-frequency feature vectors under different working conditions, and generalize the feature space based on machine learning to obtain an expanded feature sample set; train the machine learning algorithm model through the expanded feature sample set. Feature enhancement ratio in step 4 RF The indicator is expressed as: (7) In the formula, M II This represents the average value of the characteristic quantities after the fault occurred. M I This represents the average value of the characteristic quantities under normal operating conditions; due to local large pulses in the signal, the characteristic boost ratio can be affected. RF If a large indicator value leads to feature confusion, then overlap needs to be introduced. NOR The indicators form a multi-dimensional evaluation system with weakened features, as shown in formula (8): (8) In the formula, OA This refers to the overlapping area of ​​characteristic quantities between normal and fault states. EAThis refers to areas that are outside the normal range. F 1. F 2 is the corresponding i Wavelet energy at any given moment NOR The larger the value, the more likely the fault state is to overlap with the normal state; based on RF Indicators and NOR The indicators plot feature points for different working conditions, and then an expanded feature sample set covering untested edge working conditions is obtained through adversarial training and deduction using a generative adversarial network.

[0035] Step 5: Input the real-time multi-dimensional time-frequency feature vector into the machine learning algorithm model, and the model outputs the fault identification result of the current analysis window; the machine learning algorithm model is a random forest classification model, the number of decision trees in the random forest classification model is set to 100, the maximum number of features is the arithmetic square root of the total number of key frequency band features, and the Gini impurity is selected as the evaluation index for split nodes.

[0036] Step 6: Perform multi-window joint decision-making on the fault identification results of multiple consecutive analysis windows. When the preset threshold for the number of consecutive windows is met, output the fault arc determination result.

[0037] Specifically, this involves performing multi-window joint decision-making on fault identification results from multiple consecutive analysis windows. A fixed length is used. N The sliding time window decision queue N =3, and the preset threshold for the number of consecutive windows is also set to 3. The identification results of each time window output in step 5 are sequentially sent into the queue, and the identification results of the most recent consecutive time windows in the queue are continuously counted. If and only if all three consecutive analysis time windows in the decision queue are independently determined to have a fault arc, the multi-time window joint decision criterion is triggered, and a confirmed fault arc determination signal is output.

[0038] Example 1 This invention provides an improved method for dynamic detection of DC fault arcs using stochastic resonance and generative inference. The method includes sampling the DC circuit current, then dynamically enhancing weakened fault features using an adaptive improved stochastic resonance model based on an asymmetric bistable state function with adaptive parameters; next, a generative adversarial network is used to infer a generalized feature sample set, and fault arc detection is achieved using random forest classification and multi-time-window joint decision-making. Figure 1 As shown, the specific steps are as follows: Step 1: Sample the DC circuit output current to obtain the discrete time-domain current signal during the analysis period; Step 2: Construct an adaptive improved stochastic resonance model based on an asymmetric bistable state function; input the discrete time-domain current signal into the model, and adaptively adjust the parameters of the asymmetric bistable state function according to the noise intensity and local energy changes represented by the discrete time-domain current signal to obtain the enhanced time-domain current signal; Step 3: Perform Rbio3.1 wavelet multi-level decomposition on the enhanced time-domain current signal to extract time-domain and time-frequency domain features, and normalize the extracted features to obtain multi-dimensional time-frequency feature vectors under different operating conditions. Step 4: Calculate the Feature Boosting Ratio (RF) and Overlap Relationship (NOR) indices for the multidimensional time-frequency feature vectors under different working conditions, and generalize the feature space based on the generative adversarial network to obtain an expanded feature sample set; train the random forest classification model using the expanded feature sample set. Step 5: Input the real-time multi-dimensional time-frequency feature vector into the random forest classification model, and the model outputs the fault identification result of the current analysis window; Step 6: Perform multi-window joint decision-making on the fault identification results of multiple consecutive analysis windows. When the preset threshold for the number of consecutive windows is met, output the fault arc determination result.

[0039] Example 2 Based on Example 1, the DC system electrical signal sampled in step 1 is a sampled electrical signal that has not undergone sensor or high-pass digital filtering. In this example, a YOKOGAWA 701930 current sensor is used to sample the DC circuit output current point by point, with a sampling frequency of... fs The setting is 200kHz to 1MHz. For the DC side of photovoltaic inverters with higher switching frequencies, select... fs =1MHz, to capture the high-frequency components of weak electric arcs; for general low-voltage DC power distribution systems, select fs =500kHz, balancing sampling accuracy and data processing volume. Number of sampling points per analysis window. q The range is 8000 to 12000. fs =500kHz q When =10000, duration of a single analysis window T =20ms, ensuring the real-time performance of subsequent feature extraction.

[0040] Discrete-time current signal obtained by sampling xq Data collected must be included under the following characteristics that weaken the signal: load noise (such as conducted interference from the frequency converter), high current (high arc energy but slow characteristic changes under 16-40A operating conditions), special arc-generating materials (such as small voltage fluctuations at the moment of arc ignition due to the oxide film on the surface of aluminum electrodes), and photovoltaic impedance networks (including DC / DC converters and multiple parallel components, making the fault current path complex). The sampled data is simultaneously stored in the HIOKI8860-50 waveform recorder and transmitted to the DSP processor in real time. Figure 2The diagram illustrates the hardware application of the method in this embodiment. Essentially, it integrates current sampling, signal preprocessing, adaptive improved stochastic resonance enhancement, feature extraction, fault determination, and result output sequentially into the same detection link. These algorithmic steps can be integrated into a fault arc detection circuit breaker or deployed in other DC protection terminals or upper-level monitoring platforms. The sampled signal enters the signal conditioning and adaptive improved stochastic resonance enhancement module. This embodiment, through the aforementioned sampling conditions, ensures that the original signal still possesses a usable signal-to-noise ratio under weakened operating conditions, providing a foundation for subsequent enhancement.

[0041] This embodiment provides the original signal basis for subsequent adaptive improvement of stochastic resonance enhancement by reasonably setting the sampling frequency and time window length and combining the data acquisition under the aforementioned weakening factors.

[0042] Example 3 Based on the above embodiments, the adaptive improved stochastic resonance model in step 2 is represented by a first-order dynamic equation, and a linear asymmetric term is introduced into the first-order dynamic equation to construct an asymmetric bistable potential function; the coefficient of the quadratic term of the potential function a ( D The Sigmoid function is used to adjust the signal based on the intensity of external noise in order to match the changes in the intensity of external noise.

[0043] The first-order dynamic equation of the adaptive improved stochastic resonance model is expressed as: (1) In the formula, For output signal; The input is the effective current component to be enhanced; It is Gaussian white noise, where For noise intensity, Standard Gaussian white noise; The asymmetric bistable state function constructed by introducing linear asymmetric terms into the first-order dynamic equation is expressed as: (2) In the formula, The coefficients of the quadratic terms of the fundamental topological structure of the dominant potential function; The coefficient of the fourth term; The coefficients of the linear asymmetric terms are obtained through the coefficients. This creates a depth difference between the left and right potential wells, matching the asymmetric characteristics of the electric arc. Introducing Sigmoid for continuously modulated potential function parameters The expression is: (3) In the formula, To limit the coefficient of the largest quadratic term; The noise intensity normalization fusion threshold; The steepness coefficient is used to smooth the continuous transition.

[0044] Step 2 involves adaptively adjusting the parameters of the asymmetric bistable state function based on the noise intensity and local energy changes represented by the discrete-time current signal. The first-order dynamic equations introduce adaptive parameter updates based on local energy changes, setting the current sliding window to the [number missing]. A window is defined, and the mean square value of the current signal within that window is calculated as the local energy of the signal. Then the adaptive adjustment factor Defined as: (4) According to the adaptive adjustment factor Dynamically update parameters and : (5) in, b 0 = 0.5 c 0 = 0.5. (Updated) b i , c i It reflects the intensity of changes in the local energy of the signal.

[0045] Get the updated parameters for the current time. , The output of the current data point is obtained by solving the fourth-order Runge-Kutta algorithm; after the loop ends, the resulting output sequence is the time-domain current signal after adaptive enhancement. The fourth-order Runge-Kutta algorithm is expressed as: (6) In the formula, For deterministic dynamic functions; The input is a discrete current signal; For the enhanced system output signal; For numerical calculation step size, η This represents the noise term after discretization. For example... Figure 4 As shown, in a representative weakened sampling scenario, the weakened feature waveform does not form a stable upward trend before and after the arcing point, and the fault features are not clearly distinguishable from the background state, making it difficult to directly use as a reliable fault criterion. After inputting the above-mentioned collected signals into the adaptive improved stochastic resonance model and extracting the Rbio3.1 key frequency band features, the wavelet features after the arcing point are rapidly enhanced and form a clearer distinction. Therefore, in this embodiment, the enhanced Rbio3.1 wavelet features after adaptive improved stochastic resonance are preferably used as the subsequent identification input.

[0046] Example 4 Based on the above embodiments, step 3 involves wavelet feature extraction of the enhanced current signal. Multi-level wavelet decomposition is performed using the Rbio3.1 antisymmetric bioorthogonal wavelet basis function, selecting sensitive frequency bands reflecting fault arc information. Time-domain features include the average current before and after the fault arc and the rate of change of current after the fault. Time-frequency domain features are the wavelet energy calculated based on the high-frequency detail scale coefficients after Rbio3.1 wavelet decomposition. Then, the extracted time-domain and time-frequency domain features are processed to eliminate amplitude reference differences. Specifically, max-min normalization is used to obtain a standardized feature shape across operating conditions.

[0047] Example 5 Based on the above embodiments, the feature enhancement ratio in step 4 RF The indicator is expressed as: (7) In the formula, M II This represents the average value of the characteristic quantities after the fault occurred. M I This represents the average value of the characteristic quantities under normal operating conditions; due to local large pulses in the signal, the characteristic boost ratio can be affected. RF If a large indicator value leads to feature confusion, then overlap needs to be introduced. NOR The indicators form a multi-dimensional evaluation system with weakened features, as shown in formula (8): (8) In the formula, OA This refers to the overlapping area of ​​characteristic quantities between normal and fault states. EA This refers to areas that are outside the normal range. F 1. F 2 refers to the wavelet energy at the corresponding time. NOR The larger the value, the more likely the fault state is to overlap with the normal state; based on RF Indicators and NOR The indicators plot feature points for different operating conditions, and then, through adversarial training and deduction using a generative adversarial network, an expanded feature sample set covering untested edge operating conditions is obtained. For example... Figure 3As shown, to overcome the limitations of the number of measured samples and the coverage of working conditions, this embodiment introduces a Generative Adversarial Network (GAN) to generalize and extrapolate the feature space of multiple working conditions. First, using wavelet features from a limited set of measured working conditions as the real sample set, a GAN architecture containing a generator and a discriminator is constructed. Through adversarial training, the generator learns the distribution patterns of real features, and based on existing measured data points, it extrapolates and generates feature samples covering unmeasured marginal working conditions, thus achieving a generative expansion of the wavelet feature dataset. This expanded sample set retains the feature evolution logic of real working conditions and fills in the blind spots of working conditions not covered by the measured samples, providing a more complete data foundation for subsequent adaptive improvement of stochastic resonance parameter pre-calibration, random forest classification model training, and algorithm generalization verification.

[0048] Example 6 Based on the above embodiments, real-time multi-dimensional time-frequency feature vectors are input into a random forest classification model. Nonlinear pattern recognition is performed on each analysis window, and the fault judgment result for the current window is output. The number of decision trees in the random forest classification model is set to 100, the maximum number of features is the arithmetic square root of the total number of features in the preset key frequency band, and the Gini impurity is used as the evaluation index for split nodes. The decision results of each window are jointly judged by a sliding time queue, and a confirmed fault arc judgment signal is output. In this embodiment, the length N of the sliding time window decision queue is set to 3, and the preset threshold for the number of consecutive windows is also set to 3. The multi-window joint decision criterion is triggered only when all three consecutive analysis windows in the decision queue independently determine that a fault arc exists, and a confirmed fault arc judgment signal is output. Figure 2 As shown, the determination result can further output audible and visual alarm signals, and send a trip command to the circuit breaker to disconnect the DC circuit when necessary. Figure 5 As shown, without the introduction of adaptive improved stochastic resonance enhancement, the original features can be directly input into the random forest classification model and combined with multi-time-window joint decision for real-time judgment. Due to insufficient differences in features before and after the fault, the model remains in a state of refusal to operate after the arcing point, failing to form a stable fault output result. This indicates that the original features without enhancement are insufficient to meet the requirements of this invention for rapid identification of fault arcs. Figure 6 As shown, after introducing adaptive improved stochastic resonance enhancement and completing key frequency band feature extraction, the random forest classification model can quickly output effective discrimination results after the arcing point. Multi-time-window joint decision-making issues a fault arc determination signal after meeting a preset threshold for the number of consecutive windows. The detection result can quickly cross the criterion and be stably output after the arcing point, with an average detection time of 0.024s. It is evident that this invention does not simply amplify all high-frequency components, but rather enhances the significance of weakened features through adaptive improved stochastic resonance, and then stably transforms the enhanced result into the detection output through wavelet decomposition and multi-time-window decision-making.

Claims

1. An improved method for dynamic detection of DC fault arcs based on stochastic resonance and generative deduction, characterized in that, This includes sampling the DC circuit current, then using an adaptive improved stochastic resonance model based on an asymmetric bistable state function with adaptive parameters to dynamically enhance weakened fault features; then using a generative adversarial network to infer the generalized feature sample set, and using random forest classification and multi-time window joint decision-making to detect fault arcs.

2. The improved method for dynamic detection of DC fault arcs based on stochastic resonance and generative deduction according to claim 1, characterized in that, Specifically, the following steps are included: Step 1: Sample the DC circuit output current to obtain the discrete time-domain current signal during the analysis period; Step 2: Construct an adaptive improved stochastic resonance model based on an asymmetric bistable state function; input the discrete time-domain current signal into the model, and adaptively adjust the parameters of the asymmetric bistable state function according to the noise intensity and local energy changes represented by the discrete time-domain current signal to obtain the enhanced time-domain current signal; Step 3: Perform time-frequency decomposition on the enhanced time-domain current signal, extract time-domain and time-frequency domain features, and normalize the extracted features to obtain multi-dimensional time-frequency feature vectors under different operating conditions. Step 4: Calculate the Feature Enhancement Ratio (RF) and Overlap Relationship (NOR) for the multi-dimensional time-frequency feature vectors under different working conditions, and generalize the feature space based on machine learning to obtain an expanded feature sample set; train the machine learning algorithm model using the expanded feature sample set. Step 5: Input the real-time multi-dimensional time-frequency feature vector into the machine learning algorithm model, and the model outputs the fault identification result of the current analysis window; Step 6: Perform multi-window joint decision-making on the fault identification results of multiple consecutive analysis windows. When the preset threshold for the number of consecutive windows is met, output the fault arc determination result.

3. The improved stochastic resonance and generative derivation method for dynamic detection of DC fault arcs according to claim 2, characterized in that, The sampling step described in step 1 is as follows: The DC system output current is sampled at a sampling frequency... f s Point-by-point sampling is performed, with the number of sampling points per analysis window. q The discrete-time domain current signal within the analysis period is obtained. x q The number of sampling points q The sampling frequency is 8000~12000. f s The range is 200kHz to 1MHz, and the duration of a single analysis window is... q / f s .

4. The improved stochastic resonance and generative derivation method for dynamic detection of DC fault arcs according to claim 2, characterized in that, The adaptive improved stochastic resonance model described in step 2 is represented by a first-order dynamic equation, and a linear asymmetric term is introduced into the first-order dynamic equation to construct the asymmetric bistable potential function; the coefficients of the quadratic term of the potential function... a ( D The Sigmoid function is used to adjust the signal based on the intensity of external noise in order to match changes in the intensity of external noise. The first-order dynamic equation of the adaptive improved stochastic resonance model is expressed as follows: （1） In the formula, For output signal; The input is the effective current component to be enhanced; It is Gaussian white noise, where For noise intensity, Standard Gaussian white noise; The asymmetric bistable state function constructed by introducing linear asymmetric terms into the first-order dynamic equation is expressed as: （2） In the formula, The coefficients of the quadratic terms of the fundamental topological structure of the dominant potential function; The coefficient of the fourth term; The coefficients of the linear asymmetric terms are obtained through the coefficients. This creates a depth difference between the left and right potential wells, matching the asymmetric characteristics of the electric arc. Introducing Sigmoid for continuously modulated potential function parameters The expression is: （3） In the formula, To limit the coefficient of the largest quadratic term; The noise intensity normalization fusion threshold; The steepness coefficient is used to smooth the continuous transition.

5. The improved stochastic resonance and generative derivation method for dynamic detection of DC fault arcs according to claim 4, characterized in that, The adaptive adjustment of the parameters of the asymmetric bistable state function based on the noise intensity and local energy changes characterized by the discrete-time current signal in step 2 specifically involves: The first-order dynamic equation introduces adaptive parameter updates based on local energy changes, setting the current sliding window to the [number missing]. A window is defined, and the mean square value of the current signal within that window is calculated as the local energy of the signal. Then the adaptive adjustment factor Defined as: （4） According to the adaptive adjustment factor Dynamically update parameters and : （5） Get the updated parameters for the current time. , The output of the current data point is obtained by solving the fourth-order Runge-Kutta method; after the loop ends, the output sequence is the time-domain current signal after adaptive enhancement.

6. The improved method for dynamic detection of DC fault arcs based on stochastic resonance and generative deduction according to claim 5, characterized in that, The fourth-order Runge-Kutta representation is as follows: （6） In the formula, For deterministic dynamic functions; The input is a discrete current signal; For the enhanced system output signal; For numerical calculation step size, η This is the noise term after discretization.

7. The improved method for dynamic detection of DC fault arcs based on stochastic resonance and generative deduction according to claim 2, characterized in that, The time-domain features mentioned in step 3 include the average current before and after the fault arc and the rate of change of current after the fault. The time-frequency transformation method is wavelet transform. After multi-level wavelet decomposition using the Rbio3.1 antisymmetric bioorthogonal wavelet basis function, the wavelet energy calculated from the frequency band of 54.7kHz to 62.5kHz is reconstructed. The normalization process adopts the maximum-minimum normalization method.

8. The improved method for dynamic detection of DC fault arcs based on stochastic resonance and generative deduction according to claim 2, characterized in that, The feature enhancement ratio mentioned in step 4 RF The indicator is expressed as: （7） In the formula, M II This represents the average value of the characteristic quantities after the fault occurred. M I This represents the average value of the characteristic quantities under normal operating conditions; due to local large pulses in the signal, the characteristic boost ratio can be affected. RF If a large indicator value leads to feature confusion, then overlap needs to be introduced. NOR The indicators form a multi-dimensional evaluation system with weakened features, as shown in formula (8): （8） In the formula, OA This refers to the overlapping area of ​​characteristic quantities between normal and fault states. EA This refers to areas that are outside the normal range. F 1. F 2 refers to the corresponding i Wavelet energy at any given moment NOR The larger the value, the more likely the fault state is to overlap with the normal state; based on RF Indicators and NOR The indicators plot feature points for different working conditions. The selected machine learning method is generative adversarial network. Through adversarial training and deduction, an expanded feature sample set covering untested edge working conditions is obtained.

9. The improved method for dynamic detection of DC fault arcs based on stochastic resonance and generative deduction according to claim 2, characterized in that, The machine learning algorithm model selected in step 5 is the random forest classification model. The number of decision trees in the random forest classification model is set to 100, the maximum number of features is the arithmetic square root of the total number of key frequency band features, and the Gini impurity is selected as the evaluation index for split nodes.

10. The improved method for dynamic detection of DC fault arcs based on stochastic resonance and generative deduction according to claim 2, characterized in that, Step 6 specifically involves: using a sliding time window decision queue with a fixed length of N to perform multi-time window joint decision-making on the fault identification results of multiple consecutive analysis time windows. The N is set to 3, and the preset threshold for the number of consecutive windows is set to 3. After outputting the fault arc determination result, an audible and visual alarm signal is further output, or a trip command is sent to the circuit breaker to cut off the DC circuit when necessary.

Images