Direct-current arc detection method based on combination of improved filtering and enhanced energy operator

Abstract

The invention relates to the technical field of photovoltaic direct-current arc fault detection, and particularly provides a direct-current arc detection method based on combination of improved filtering and an enhanced energy operator. According to the method, firstly, frequency domain information of fault signals is extracted through fast Fourier transform, the signals are preprocessed in combination with extreme point density and fluctuation factors, and noise interference is reduced; noise suppression and smoothing processing of signals are realized by using an adaptive filter of an improved cuckoo algorithm, and meanwhile, key spectrum characteristics are reserved. In order to solve the problem of amplitude attenuation in the signal processing process, an energy operator improved based on three-point symmetric difference is introduced, the nonlinear transient characteristic of the signal is enhanced, and frequency domain component noise is further suppressed. And finally, extracting enhanced signal features, and performing efficient identification on fault feature data by adopting a residual error connection stacked auto-encoder. By combining the adaptive spectral line enhancement technology and the deep learning model, the problem of arc fault detection in a direct current system and a photovoltaic system is effectively solved, the detection requirements under complex operation conditions such as inverter starting noise are particularly met, the accuracy and reliability of fault recognition are improved, and the method is suitable for large-scale popularization and application. The method is suitable for safety monitoring and early fault diagnosis of a photovoltaic system and a DC power system.

Description

Technical Field

[0001] The invention relates to the technical field of photovoltaic DC arc fault detection, in particular to a DC arc detection method based on improved filtering combined with an enhanced energy operator. Background Art

[0002] As the energy structure accelerates its transformation towards renewable energy, photovoltaic power generation, as the core form of solar energy development, has attracted much attention for its safety and reliability. In photovoltaic systems, DC side arc faults frequently occur due to reasons such as line insulation aging and poor contact. DC arcs have the characteristics of rich high-frequency harmonics, concentrated energy and no zero crossing, which makes it difficult to be quickly detected and removed by traditional methods. It is easy to cause serious hidden dangers such as local overheating, equipment damage and electrical fires, which significantly threatens the safe operation of the system. Although DC arc detection technology has made some progress, there are still bottlenecks in signal feature extraction and adaptability to complex environments, which limits the stability improvement of photovoltaic systems. Especially when the inverter is started, noise interference will be significantly superimposed on the fault signal, making it more complicated to accurately identify and distinguish normal startup behavior from fault status. The effective detection of DC fault arcs depends largely on the accurate extraction and analysis of its signal features, which is the core link to improve the detection performance. Existing studies have shown that the detection technology based on the steady-state mode of high-frequency electromagnetic radiation has high sensitivity and environmental adaptability, but the high requirements for the anti-interference performance of the sensor limit its practical application. In addition, improved signal processing algorithms, such as the combination of spectrum decomposition and time-frequency feature analysis, have made some progress in improving anti-interference capabilities and detection accuracy, but they are still clearly insufficient when dealing with weak signal feature extraction and diversified application scenarios. How to improve detection accuracy while reducing dependence on equipment and environmental conditions is still a core issue that needs to be solved urgently. Summary of the invention

[0003] In view of the above problems, the purpose of the present invention is to provide a DC arc detection method based on improved filtering combined with an enhanced energy operator, so as to solve the problems of insufficient signal feature extraction accuracy and poor adaptability to complex environments in the prior art, and to provide an efficient and accurate photovoltaic DC arc detection method.

[0004] To achieve the above object, the present invention adopts the following technical solution: a DC arc detection method based on improved filtering combined with enhanced energy operator, characterized in that it includes the following steps:

[0005] S1: Obtain DC arc fault data under four different scenarios: photovoltaic resistance, photovoltaic grid connection, DC resistance, and DC grid connection;

[0006] S2: Through the startup interference suppression strategy, the startup noise interference area is accurately identified and the interference components are dynamically corrected, realizing the effective distinction between the DC arc fault signal and the startup noise interference signal. This not only solves the misjudgment problem of the traditional method in the identification of startup noise interference, but also improves the ability to retain key fault features, thereby realizing accurate analysis and classification of the signal;

[0007] S3: Adaptive filters are used to suppress noise and smooth frequency domain signals. The improved cuckoo optimization algorithm is used to dynamically adjust filter parameters to improve adaptability and optimization effects. This effectively solves the problem of difficult parameter settings due to differences in noise interference levels in different scenarios, and significantly improves signal processing robustness and background noise suppression capabilities.

[0008] S4: By introducing the improved TEO energy operator, the signal spectrum is deeply enhanced, and the three-point symmetrical difference strategy is combined to optimize the core formula, which significantly amplifies the characteristic amplitude of the fault signal and suppresses the background noise interference, thereby achieving accurate analysis and efficient distinction of signal detail features;

[0009] S5: Fault identification is performed on the processed data, a multi-dimensional feature extraction scheme is designed, and an automatic detection mechanism is implemented based on the stacked autoencoder (SAE). In order to overcome the problems of gradient vanishing and low convergence efficiency of SAE during training, residual connections are further introduced to optimize the network structure, which not only improves the stability of gradient transfer, but also accelerates the convergence process of the model, significantly enhancing the accuracy of fault detection.

[0010] Furthermore, the step S1 is specifically as follows: collecting voltage data of the fault arc by building a DC arc collection device under four scenarios.

[0011] Furthermore, the step S2 is specifically: combining the dual feature analysis method of extreme point density and fluctuation factor to achieve accurate identification and correction of interference data. Let x(t) be the original signal, and define the extreme point density as:

[0012]

[0013] Where ρ(t) is the extreme point density of the signal in the window Δt, N(t) is the number of extreme points in the window Δt, Δt is the analysis window width, and the density threshold is set to identify the startup noise interference area, thereby completing the preliminary separation of the interference data.

[0014] The fluctuation factor is used to quantify the amplitude fluctuation degree in the frequency domain, capture the abnormal characteristics of the interference frequency band, and dynamically adjust the amplitude in the frequency band by comparing the fluctuation factors of the interference frequency band and the normal frequency band to correct the high amplitude fluctuation in the spectrum. The expression is:

[0015]

[0016] Where A(f) is the amplitude distribution in frequency band f; σ(A(f)) and μ(A(f)) are the standard deviation and mean of the amplitude, respectively. By adopting the startup interference suppression strategy, the startup noise interference and arc fault signal are effectively distinguished, and the stability of the spectrum and the identifiability of the fault signal are enhanced.

[0017] Furthermore, the step S3 is specifically as follows: in order to solve the problem of signal interference by background noise, an adaptive filter based on an improved cuckoo algorithm is used to achieve dynamic suppression of background noise through the adaptive filter; the step size and filter order of the filter are globally optimized using the improved cuckoo algorithm, thereby significantly improving the robustness of signal processing and the accuracy of fault detection.

[0018] The adaptive filter suppresses the background noise in the signal in real time. The input signal x(n) is introduced into the filter for reconstruction, and an input vector x containing historical related information is constructed. τ (n), its construction formula is:

[0019] x τ (n)=[x(n),x(n-τ),...,x(n-L+τ)] T

[0020] In the formula, L is the order of the filter and τ is the delay length. The cost function is defined by the mean square error to measure the deviation between the filter output signal and the target signal. The formula is expressed as:

[0021] J LMS =E[e 2 (n)]

[0022] Where E represents the mathematical expectation, and e(n) is the difference between the error function d(n) and the filter output y(n).

[0023] The weight update strategy is determined by its gradient direction, and its weight update formula is:

[0024] w(n+1)=w(n)+2μe(n)x τ (n)

[0025] In the formula, w(n) is the filter weight vector and μ is the step size factor. By adjusting the weight update and the offset filter to optimize the iteration in the negative direction of the cost function gradient, the output signal gradually removes the low-correlation background noise, while strengthening the high-correlation target signal characteristics, significantly improving the signal quality and the reliability and accuracy of fault detection. The formula is expressed as:

[0026]

[0027] Where tr(R) is the trace of the autocorrelation matrix of the input signal, and L is the filter order. By introducing the offset, the signal characteristics, step size factor and filter performance can be effectively correlated, and used as the objective function of the improved cuckoo optimization algorithm to provide theoretical support for parameter optimization and systematically quantify the steady-state performance of the output signal y(n). The output expression is:

[0028] y(n)=w T (n)x τ (n)

[0029] In the four scenarios, the background noise characteristics have different effects on the data, making it difficult to accurately set the step size and filter order parameters of the adaptive filter through traditional empirical methods. The cuckoo optimization algorithm is introduced and improved by enhancing the population distribution, optimizing the search strategy and improving the global exploration ability. The specific improvements include: using chaotic tent mapping instead of random initialization to enhance the uniformity of population distribution. The formula is expressed as:

[0030]

[0031] In the formula, x i is the current chaos value (i.e. the current nest position), x i+1 is the iterative chaotic value (i.e., the newly generated nest position), ε is the tent mapping threshold parameter, and its value is 0.499.

[0032] Introducing a dynamically adjusted discovery probability instead of a fixed value can expand the search range and improve the optimization accuracy. The formula is:

[0033]

[0034] Where P a Indicates the discovery probability value of the current iteration stage; P a,max Indicates the maximum value of the discovery probability, which is the upper limit of the dynamic adjustment function, and k represents the current number of iterations; K max is the maximum number of iterations. This method ensures that the probability P is found early in the algorithm. a Keep it small, give priority to enhancing the global exploration ability, and ensure the wide coverage of the search space; as the iteration proceeds, the probability of discovery P a Gradually increase to improve the local development ability of the algorithm.

[0035] The cosine step size is used instead of the fixed step size, and the step size is flexibly adjusted according to different search stages. The formula is expressed as:

[0036]

[0037] Where α is the search step factor of the cuckoo algorithm, α maxThe strategy uses a larger step size to enhance the global exploration capability and cover a wide solution space at the beginning of the algorithm. As the iteration proceeds, the step size gradually decreases, focusing on the fine optimization of the local area, thereby improving the optimization accuracy.

[0038] Cauchy mutation is used to replace random perturbations to enhance global exploration capabilities and avoid falling into local optimality. The formula is expressed as:

[0039] σ=η·tan(π·(rand-0.5))

[0040] In the formula, σ is the Cauchy mutation perturbation factor; η is the scaling factor used to adjust the amplitude of the mutation; rand is a random number in the interval [0,1]; tan is the tangent function. Cauchy mutation uses its long-tail distribution characteristics to enhance the exploration ability of the understanding space through large-scale jumps, effectively avoiding the local optimal trap. At the same time, Cauchy mutation and Levy flight complement each other. The former focuses on the scalability of global search, while the latter focuses on the refinement of local development. The combination of the two achieves a balance between breadth and depth in search updates.

[0041] Further, the step S4 is specifically: using the TEO operator to process the signal based on the forward difference, and the formula is expressed as:

[0042] ψ[y(n)]=[y(n)] 2 -y(n+1)·y(n-1)

[0043] Where ψ[y(n)] is the energy operator of the discrete signal at time point n. However, since only the difference between the current moment and the previous moment is considered, it is difficult to fully capture the dynamic characteristics of the signal, and it is easy to produce large errors in a fast-changing or high-frequency signal environment. At the same time, it does not support the nonlinear changes of the actual signal. In order to improve its performance, the core formula of the three-point symmetric difference improved TEO energy operator is introduced as follows:

[0044]

[0045] By combining the signal values ​​at the current moment and the two moments before and after, the instantaneous changes of the signal can be more comprehensively reflected, while smoothing the drastic changes caused by the background noise and reducing the estimation error of the difference at a single moment. On this basis, the frequency domain amplitude feature a(n) of the enhanced or suppressed signal is obtained, and its expression is:

[0046]

[0047] The spectrum enhancement method based on the improved TEO energy operator can capture the discrete characteristic changes of the signal after filtering, further suppress the residual noise and enhance the feature significance.

[0048] Further, the step S5 is specifically as follows: after enhancing the processed data, by extracting the combined feature vector X(n) of the fluctuation factor W(f) and the spectrum energy E, the fluctuation factor W(f) is used to capture the rapid change and mutation trend of the signal and reveal the dynamic characteristics of the fault signal; the spectrum energy E reflects the spectrum characteristics of the fault by analyzing the energy distribution of the signal in the frequency domain, and its spectrum energy formula is:

[0049]

[0050] In order to further improve the accuracy of fault detection, the SAE model is introduced to effectively represent the input data through multi-layer nonlinear transformation and gradually extract high-order features. Its formula is expressed as:

[0051] Y(n)=ζ(W e X(n)+b e )

[0052] Z(n)=ζ(W d Y(n)+b d )

[0053] In the formula, ζ is a nonlinear mapping function, Y(n) is the intermediate layer vector, Z(n) is the output vector, and W e is the weight matrix of the encoder; b e is the encoder bias vector, W d is the weight matrix of the decoder; b d is the bias vector of the decoder and is updated through the back-propagation algorithm. In order to solve the possible "degradation problem" of the model, residual connections are used to optimize the stacked autoencoder. Through residual path optimization, the formula is expressed as:

[0054] Z(n)=F(X(n))+X(n)

[0055] Where F is a nonlinear transformation function. This method can effectively alleviate the "degradation problem" of the model, improve the convergence speed and training stability of the network, and thus significantly improve the reliability and robustness of fault detection.

[0056] The beneficial effects of the present invention are:

[0057] (1) By combining the startup noise suppression strategy, adaptive filter and improved cuckoo algorithm, the feature extraction problem of DC arc signal under the interference of startup noise and background noise is effectively solved. The startup noise suppression strategy uses the dual feature analysis of extreme point density and fluctuation factor to accurately identify and dynamically correct noise interference, significantly enhancing the recognition of fault signals and the stability of the spectrum. At the same time, the adaptive filter dynamically suppresses the background noise, and the improved cuckoo algorithm is combined to optimize the filtering parameters, further improving the robustness and adaptability of signal processing, thereby significantly improving the fault signal analysis effect in complex environments.

[0058] (2) By introducing the spectrum enhancement method and SAE model based on the improved TEO energy operator, the present invention realizes the deep enhancement and automatic detection of signal features. The improved TEO operator combined with three-point symmetrical difference not only amplifies the characteristic amplitude of the fault signal, but also suppresses the residual noise interference, providing high-quality input data for subsequent feature extraction. The SAE-based detection framework uses fluctuation factors and spectrum energy to construct a combined feature vector, optimizes the network structure through residual connection, effectively improves the convergence speed and detection accuracy of the model, and further enhances the reliability and adaptability of DC arc fault detection under multiple scenarios and multiple interference conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0060] Figure 1 A flow chart of a DC arc detection method based on improved filtering combined with an enhanced energy operator provided in an embodiment of the present application;

[0061] Figure 2 A flow chart of an improved cuckoo optimization algorithm provided in an embodiment of the present application;

[0062] Figure 3 A schematic diagram of a residual block structure provided in an embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of this application clearer, the relevant technical solutions will be described in detail below with reference to the accompanying drawings. It should be pointed out that the embodiments described here are only part of the content of this application, not all embodiments. In this field, all other embodiments that can be obtained based on the embodiments in this application by ordinary technicians without creative work should be included in the scope of protection of this application.

[0064] Figure 1 A flow chart of a DC arc detection method based on improved filtering combined with an enhanced energy operator provided in an embodiment of the present application, a DC arc detection method based on improved adaptive filtering combined with an enhanced energy operator and a stacked autoencoder with residual connection specifically comprises the following steps:

[0065] Step S101, using the constructed DC arc acquisition device to input voltage data under four scenarios.

[0066] As an embodiment, the collected voltage data may include normal voltage data in four scenarios, arc fault voltage data, and voltage data of startup noise interference in photovoltaic grid-connected and DC grid-connected scenarios, etc.

[0067] Step S102, performing startup noise suppression processing on the voltage data.

[0068] As an embodiment, step S102 includes: using extreme point density to perform preliminary separation of interference data on the original signal; and using fluctuation factors to dynamically adjust and correct abnormal amplitude fluctuations in the frequency domain.

[0069] Among them, the extreme point density is calculated by comparing the number of extreme points in the statistical analysis window with the set threshold (ρ(t)>0.009) to identify the startup noise interference area and calculate the fluctuation factor value in the identified data frequency band. The calculation formula is:

[0070]

[0071] Where A(f) is the amplitude distribution on frequency band f; σ(A(f)) and μ(A(f)) are the standard deviation and mean of the amplitude, respectively. By comparing the fluctuation factors of the interference frequency band and the normal frequency band, the abnormal characteristics of the interference frequency band can be determined and eliminated.

[0072] Step S103: input the signal after the noise suppression process is started into an adaptive filter to further suppress the noise.

[0073] As an embodiment, step S103 includes: constructing an input vector of history-related information for the input signal; defining a cost function by a mean square error to measure the deviation between the filter output signal and the target signal; and adjusting the filter weight based on a gradient descent method, wherein the gradient descent formula is as follows:

[0074] ▽J LMS = -2E[e(n)x τ (n)]

[0075] In the formula, E represents the mathematical expectation, ▽J LMS is the gradient vector of the desired signal d(n) (i.e., the input data x(n)), e(n) is the difference between the error function d(n) and the filter output y(n), μ is the step size factor, x τ (n) is the reconstructed input data.

[0076] The offset reflects the relationship between the autocorrelation characteristics of the signal and the filter parameters, and is used as the objective function of the improved cuckoo optimization algorithm.

[0077] Step S104, introducing an offset in the filter optimization process to dynamically correlate signal characteristics and filter parameters, the formula is:

[0078]

[0079] Where tr(R) is the trace of the autocorrelation matrix of the input signal, L is the filter order, and the step size factor μ. The offset reflects the global characteristics of the signal through the trace of the autocorrelation matrix. After balancing with the step size factor μ and the filter order L, the performance of the filter is effectively improved. The offset is used as the objective function of the improved cuckoo algorithm, and the global search and local development capabilities of the algorithm are used to further optimize the step size and order parameters of the filter.

[0080] As an embodiment, step S104 includes: selecting a reasonable range.

[0081] The selection range should ensure that the filter can take into account both convergence speed and optimization accuracy. When the offset is too small, the filter's ability to extract the characteristics of the target signal is weakened, which may lead to reduced optimization efficiency; while excessive offset may cause unstable filter convergence or overfitting of noise. According to experimental measurements, the range of offset can be selected in a reasonable range of [1e-5, 1e-3] to ensure a balance between filter performance and optimization effect.

[0082] Step S105, using the improved cuckoo algorithm to perform global search and local development on the filter parameter optimization problem in step S103, and outputting the optimized filter parameters.

[0083] Among them, the structural diagram of the improved cuckoo algorithm is as follows Figure 2 shown.

[0084] As an embodiment, step S105 includes: initializing a population based on Tent mapping; dynamically adjusting the discovery probability; dynamically adjusting the step size; and Cauchy mutation optimization.

[0085] The population initialization refers to generating the initial population position using the Tent chaotic map to enhance the uniformity and diversity of population distribution, avoid local deviations that may be caused by random initialization, and thus improve the global search capability. The formula is expressed as follows:

[0086]

[0087] In the formula, x i is the current nest position, x i+1 is the newly generated nest position, and ε is the tent mapping threshold parameter. After initialization, the objective function value corresponding to the position of each nest (i.e., the misalignment in step S104) is calculated, and the optimal solution of the current population is recorded. Subsequently, the nest position is updated based on the Levy flight formula, and the global exploration ability of the algorithm is improved through large-scale random jumps, and compared and replaced with the optimal value of the previous generation population to ensure that the population continues to evolve towards the optimal solution.

[0088] The dynamic adjustment of the discovery probability is achieved by introducing an adaptive mechanism to adjust the discovery probability value at different iteration stages, so that the algorithm can expand the search scope for global exploration in the early stage and gradually focus on local development in the later stage. The formula is expressed as:

[0089]

[0090] Where Pa represents the probability of finding a bird's nest in the current iteration stage; P a,max represents the maximum probability of finding a bird's nest, k represents the current number of iterations; K max is the maximum number of iterations.

[0091] The cosine adjustment of the search step size refers to dynamically adjusting the step size using the cosine function, using a larger step size to cover a wider solution space in the global search phase, and shrinking the step size in the local development phase to focus on optimizing the potential optimal solution area. The formula is expressed as:

[0092]

[0093] In the formula, α max is the maximum step size, and α is the step size factor.

[0094] In the local search stage, Cauchy mutation optimization is further introduced to improve the global exploration ability by adding large disturbances, while avoiding falling into the local optimum and maintaining the diversity of the population. The adjustment mechanism of Cauchy mutation enables the mutation amplitude to adapt to the needs of different optimization stages, and its formula is expressed as

[0095] σ=η·tan(π·(rand-0.5))

[0096] In the formula, σ is the Cauchy variation disturbance factor; η is the scaling factor used to adjust the variation amplitude; rand is a random number; and tan is the tangent function.

[0097] In the iterative process of the algorithm, multiple rounds of global search and local development are performed alternately to continuously update the current optimal solution. When the iteration termination conditions are met, the optimized filter parameters are output, including the optimal step size and order of the filter.

[0098] Step S106: Use the improved TEO energy operator to perform noise suppression and feature enhancement on the data that meets the adjustment requirements described in step S104.

[0099] As an embodiment, step S106 includes: improving the TEO energy operator by using three-point symmetric difference

[0100] The three-point symmetrical difference is to comprehensively reflect the instantaneous change characteristics of the signal by combining the signal values ​​at the current moment and the two moments before and after, and effectively smooth the drastic changes caused by background noise, thereby reducing the error of single moment difference estimation. The formula is expressed as follows:

[0101]

[0102] Wherein, y(n)' is the data after three-point symmetrical difference processing, and y(n) is the output data of step S104.

[0103] The improved TEO energy operator uses three-point symmetric difference to modify its core formula, and the modified formula is:

[0104]

[0105] Where ψ[y(n)] is the energy operator of the discrete signal at time point n.

[0106] The improved TEO energy operator can capture the discrete characteristic changes of the filtered signal, significantly improve the frequency domain amplitude characteristics of the signal through spectrum enhancement, suppress residual noise and highlight the characteristic significance of the signal.

[0107] Step S107, by introducing data enhancement and feature extraction technology, combined with the stacked autoencoder model and residual connection mechanism, the performance and robustness of fault detection are further improved.

[0108] Among them, the structural diagram of the residual block is as follows Figure 3 shown.

[0109] As an embodiment, step S108 includes: constructing a combined feature vector, using the fluctuation factor to capture the dynamic change characteristics of the signal, and reflecting the frequency domain characteristics of the fault signal through the spectrum energy; using a stacked autoencoder to extract high-order features through multi-layer nonlinear mapping; introducing a residual connection mechanism, and realizing jump connection through a nonlinear transformation function, which is expressed as follows:

[0110] Z(n)=F(X(n))+X(n)

[0111] In the formula, Z(n) is the output vector, X(n) is the input vector, and F is a nonlinear transformation function. The processed data can effectively alleviate the gradient attenuation problem in the deep network, thereby improving the network convergence speed and training stability, forming a stacked autoencoder model based on residual connection, and its output expression becomes two cases:

[0112] Y ie (n) = ζ(W e X i-1 (n)+b e )+X i-1 (n)

[0113] Y id (n) = ζ(W d X i-1 (n)+b d )+X i-1 (n)

[0114] Where Y ie (n) is the output of the i-th encoder, Y id (n) is the output of the i-th decoder.

[0115] Step S108, outputting the final DC arc detection result for use in fault detection and early warning of power equipment.

[0116] Finally, it should be noted that the above is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A DC arc detection method based on improved filtering combined with enhanced energy operator, characterized in that: The method comprises: Obtain DC arc fault data in four different scenarios: photovoltaic resistance, photovoltaic grid connection, DC resistance, and DC grid connection; Using a startup interference suppression strategy on the DC arc fault data, accurately identifying the startup noise interference area and dynamically correcting the interference component; An adaptive filter model based on the improved cuckoo optimization algorithm is constructed, and background noise suppression processing of DC arc fault data is performed; For the processed data, the improved TEO energy operator is used for feature enhancement; The enhanced data feature values ​​are extracted, and direct current arc fault detection is performed according to the feature values.

2. The method according to claim 1, characterized in that The data preprocessing includes the following steps: The data is differentiated and noise suppression is initiated using a startup interference suppression strategy, which is implemented by the following steps: Using the dual characteristic analysis method of extreme point density and fluctuation factor, let x(t) be the original signal and define the extreme point density as: Where ρ(t) is the extreme point density of the signal in the window Δt, N(t) is the number of extreme points in the window Δt, Δt is the analysis window width, and the density threshold is set to identify the startup noise interference area. The fluctuation factor is used to quantify the amplitude fluctuation degree in the frequency domain, capture the abnormal characteristics of the interference frequency band, dynamically adjust the amplitude in the frequency band, and correct the high amplitude fluctuation in the spectrum. The expression is: Where A(f) is the amplitude distribution in frequency band f; σ(A(f)) and μ(A(f)) are the standard deviation and mean of the amplitude, respectively. The startup interference suppression strategy is used to effectively distinguish the startup noise interference from the arc fault signal.

3. The method according to claim 1, characterized in that: The background denoising process comprises the following steps: Adaptive filters are used to dynamically suppress background noise. The input data is reconstructed through adaptive filters to construct an input vector x containing historical related information. τ (n), the cost function is defined by the mean square error to measure the deviation between the filter output signal and the target signal. Its formula is expressed as: J LMS =E[e 2 (n)] In the formula, E represents the mathematical expectation, and e(n) is the error function, which is the difference between the objective function d(n) and the filter output y(n). On this basis, the offset is derived. This parameter can effectively associate the signal characteristics, step size factor and filter performance, and is used as the objective function of the improved cuckoo optimization algorithm. Its formula is expressed as: In the formula, tr(R) is the trace of the autocorrelation matrix of the input signal, and L is the filter order. Combined with the improved cuckoo optimization algorithm, the step size and filter order of the filter are globally optimized with the offset as the objective function to improve the performance of the adaptive filter.

4. The method according to claim 3, characterized in that The modified cuckoo optimization algorithm is implemented by the following steps: Chaotic Tent mapping is used to replace random initialization to enhance the uniformity of population distribution. The formula is expressed as: In the formula, x i is the current chaos value (i.e. the current nest position), x i+1 is the iterative chaotic value (i.e., the newly generated nest position), ε is the tent mapping threshold parameter, and its value is 0.

499. Introducing a dynamically adjusted discovery probability instead of a fixed value can expand the search range and improve the optimization accuracy. The formula is: Where P a Indicates the discovery probability value of the current iteration stage; P a,max Indicates the maximum value of the discovery probability, which is the upper limit of the dynamic adjustment function, and k represents the current number of iterations; K max is the maximum number of iterations. The cosine step size is used instead of the fixed step size, and the step size is flexibly adjusted according to different search stages. The formula is expressed as: Where α is the search step factor of the cuckoo algorithm, α max is the maximum step length; Cauchy mutation is used to replace random perturbations to enhance global exploration capabilities and avoid falling into local optimality. The formula is expressed as: σ=η·tan(π·(rand-0.5)) Where σ is the Cauchy variation disturbance factor; η is the scaling factor used to adjust the variation amplitude; rand is a random number in the interval [0,1]; and tan is the tangent function.

5. The method according to claim 1, characterized in that The method of processing data based on the improved TEO energy operator comprises the following steps: The adopted TEO energy operator processes the data processed by the adaptive filtering model based on forward difference processing, and the improved TEO energy operator is improved by three-point symmetric difference, and the improved formula is expressed as: Where ψ[y(n)] is the energy operator of the discrete signal at time point n; The improved TEO energy operator is used to process the data processed by the adaptive filtering model, and the frequency domain amplitude a(n) is expressed as:

6. The method according to claim 1, characterized in that The feature extraction process is achieved by the following steps: By analyzing the data processed by the modified TEO energy operator, the combined feature vector X(n) of the fluctuation factor W(f) and the spectrum energy E is extracted according to its frequency domain characteristics. The spectrum energy E formula is: Based on this feature vector X(n), a stacked autoencoder model based on residual connection is used for fault detection.

7. The method according to claim 6, characterized in that The stacked autoencoder model process based on residual connection is implemented by the following steps: The stacked autoencoder effectively represents the input data through multiple layers of nonlinear transformation and gradually extracts high-order features. Its formula is expressed as: Y(n)=ζ(W e X(n)+b e ) Z(n)=ζ(W d Y(n)+b d ) In the formula, ζ is a nonlinear mapping function, Y(n) is the intermediate layer vector, Z(n) is the output vector, and W e is the weight matrix of the encoder; b e is the encoder bias vector, W d is the weight matrix of the decoder; b d is the bias vector of the decoder; Residual connections are used to optimize stacked autoencoders. Through residual path optimization, the possible "degradation problem" of the model is solved. The residual connection formula is expressed as: Z(n)=F(X(n))+X(n) Where F is a nonlinear transformation function, forming a stacked autoencoder model based on residual connection, and its output expression becomes two cases: Y ie (n)=ζ(W e X i-1 (n)+b e )+X i-1 (n) Y id (n)=ζ(W d X i-1 (n)+b d )+X i-1 (n) Where Y ie (n) is the output of the i-th encoder, Y id (n) is the output of the i-th decoder.

Images