A noise reduction processing method based on high-frequency test signals of a distribution box

Abstract

The present application belongs to the technical field of data processing, and particularly relates to a power distribution box high-frequency test signal denoising processing method, which comprises the following steps: obtaining an original signal to perform frequency domain transformation to generate complex frequency spectrum data; dividing a concerned frequency band into sub-bands and calculating spectral entropy and spectral line gradient energy to identify signal characteristic sub-bands; adaptively determining the target mode number of variational mode decomposition according to the number of characteristic sub-bands; calculating noise injection strength in combination with background noise power and average spectral entropy; synthesizing auxiliary noise sequences by using target amplitude spectrum and chaotic phase spectrum and superimposing the auxiliary noise sequences to the test signal to excite weak physical components; performing variational mode decomposition on the superimposed signal according to the target number to extract multiple intrinsic mode functions; calculating double-domain composite weights and performing weighted fusion to reconstruct a denoised signal. The present application effectively solves technical problems such as difficulty in parameter configuration and distortion of reconstructed waveforms under extremely low signal-to-noise ratio and poor working conditions.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a method for noise reduction processing of high-frequency test signals from a distribution box. Background Technology

[0002] Distribution boxes are core node equipment in power distribution systems. Early faults inside distribution boxes often manifest as externally measurable high-frequency electromagnetic signals. By collecting and analyzing the corresponding signals through high-frequency sensors, status monitoring and early warning can be achieved. However, distribution boxes operate under strong electromagnetic interference for many years, resulting in low signal-to-noise ratios of the high-frequency test signals collected by sensors. The original test signals not only contain weak fault characteristic components representing the equipment status, but also contain a large amount of broadband white noise and random pulse interference. Strong electromagnetic noise can easily drown out weak fault information, leading to misjudgments of equipment status.

[0003] In existing technologies, fault features are extracted by introducing basis-free signal decomposition algorithms such as variational mode decomposition (VMD) and performing adaptive decomposition based on time-scale characteristics. Referring to Chinese invention patent document CN113887398B, a GPR signal denoising method based on variational mode decomposition and singular spectrum analysis is disclosed. This method uses VMD to extract signal modal components, calculates the Pearson correlation coefficient for each component, and constructs a judgment threshold relationship. In this technical solution, when the correlation coefficient is greater than the judgment threshold, the corresponding component is retained; otherwise, it is discarded. Based on the above technical means, effective signal components can be screened and some background noise can be filtered out, achieving basic noise reduction processing of the test signal.

[0004] However, while the above-mentioned technical solutions can achieve basic noise reduction of test signals to a certain extent, the variational mode decomposition effect is highly dependent on the number of target modes. Under harsh conditions with low signal-to-noise ratio, the modal energy corresponding to weak fault features is much smaller than that of strong noise modes. The lack of parameter adaptive configuration mechanism in existing technologies often makes it impossible to effectively remove the corresponding weak physical components. Secondly, the existing technologies disclose the corresponding hard threshold reconstruction formula by using a single correlation coefficient to perform a one-size-fits-all mode truncation. This rigid approach is prone to losing the corresponding weak physical details hidden inside the noise modes. At the same time, the direct participation of binarization coefficients in multiplication operations can easily introduce serious step errors and spectral energy leakage into the final reconstructed waveform. Summary of the Invention

[0005] To address the technical problems of adaptive configuration of variational mode decomposition parameters in low signal-to-noise ratio environments and waveform distortion caused by traditional hard threshold mode reconstruction, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a noise reduction processing method for high-frequency test signals from a distribution box, comprising: acquiring a high-frequency test signal and performing frequency domain transformation, dividing the frequency band of interest into multiple frequency sub-bands, and calculating the frequency domain statistical characteristics and energy distribution characteristics of each frequency sub-band; determining the attributes of each frequency sub-band based on a preset threshold, identifying frequency sub-bands that meet preset conditions as signal feature sub-bands, identifying frequency sub-bands that do not meet preset conditions as noise-dominant sub-bands, calculating the average background noise power based on the noise-dominant sub-bands, and adaptively determining the target number of variational mode decomposition based on the number of signal feature sub-bands; determining the noise injection intensity based on the average background noise power and the characteristics of the signal feature sub-bands, and generating a target amplitude spectrum based on the noise injection intensity; synthesizing an auxiliary noise sequence by combining a non-periodic random phase spectrum, and superimposing the auxiliary noise sequence with the high-frequency test signal to generate a superimposed signal for jitter injection; performing variational mode decomposition on the superimposed signal with the target number of modes to extract multiple intrinsic mode functions, and performing weighted summation and reconstruction on the multiple intrinsic mode functions by fusing multi-domain composite weights to generate a noise-reduced test signal.

[0007] This invention accurately locates signal subbands by utilizing frequency domain statistical features and energy distribution features, and determines the number of target modes accordingly, effectively avoiding the risks of over-decomposition or under-decomposition. At the same time, it introduces a jitter injection mechanism, which adaptively adjusts the noise injection intensity based on background noise and signal features, and generates an auxiliary noise sequence in conjunction with a non-periodic random phase spectrum, thereby effectively helping the algorithm to escape local minima and enhancing the sensitivity to weak features. Finally, through multi-domain composite weighting for weighted reconstruction, it can smoothly preserve fault details and eliminate waveform distortion caused by traditional hard threshold truncation.

[0008] Preferably, the frequency domain transformation includes: performing sampling processing on the high-frequency test signal using a sampling rate of not less than 2.56 MHz; performing ChirpZ transform on the high-frequency test signal and setting the number of calculation points to 8192; and extracting all values ​​within the frequency band of interest with a frequency point spacing of less than 50 Hz to form complex spectrum data.

[0009] Preferably, the calculation of the frequency domain statistical characteristics and energy distribution characteristics of each frequency sub-band, and the determination of the attributes of each frequency sub-band, includes: calculating the spectral entropy and spectral gradient energy of each frequency sub-band as the frequency domain statistical characteristics and the energy distribution characteristics, respectively; calculating the sum of the mean spectral entropy of all frequency sub-bands and twice the standard deviation as the entropy threshold, and calculating the sum of the mean spectral gradient energy of all frequency sub-bands and three times the standard deviation as the energy threshold, wherein the entropy threshold and the energy threshold constitute the preset threshold; when the spectral entropy of a certain frequency sub-band is less than the entropy threshold and the spectral gradient energy of the certain frequency sub-band is greater than the energy threshold, the certain frequency sub-band is determined to be the signal feature sub-band.

[0010] This invention utilizes spectral entropy to assess the determinism of the spectral structure and spectral line gradient to assess the smoothness of the spectrum. By constructing a statistically based adaptive threshold system, it can dynamically adapt to changes in background noise according to different electromagnetic environments, effectively eliminating the risk of judgment failure caused by over-reliance on setting fixed thresholds based on human experience.

[0011] Preferably, the adaptive determination of the target number of variational mode decomposition includes: calculating the total number of signal feature subbands to generate the total number of signal feature subbands; and setting the target number of modes to be equal to the sum of the total number of signal feature subbands and the value 2.

[0012] This invention establishes a direct mapping relationship between physical information components and the number of decomposed modes. By introducing margin modes to carry full-band background noise and deal with potential mode aliasing, this structured derivation method can significantly reduce the amount of computation in massive iterative optimization and reduce hardware computing power consumption.

[0013] Preferably, determining the noise injection intensity based on the average background noise power and the average spectral entropy of the signal characteristic subband includes: setting a reference intensity parameter equal to the average background noise power multiplied by 0.1; Based on the mapping function Calculate the noise injection intensity; In the formula, Indicates the noise injection intensity. This represents an exponential coefficient that is greater than zero, ensuring that the noise injection intensity amplifies nonlinearly with the increase of the average spectral entropy of the signal's characteristic subbands. The average spectral entropy of the characteristic subband of the signal.

[0014] The exponential mapping function constructed in this invention can achieve adaptive and precise tuning of the injected energy: when the signal characteristics are not obvious, the injection intensity is amplified exponentially to give full play to the gain effect of jitter injection; when the characteristics are obvious, a very small reference noise is output to avoid introducing additional electromagnetic interference.

[0015] Preferably, generating the target amplitude spectrum includes: extracting the frequency band boundaries of each signal feature sub-band to determine the corresponding adjacent transition frequency bands; generating an amplitude distribution structure in all the adjacent transition frequency bands using a Gaussian window function; and performing a normalization operation on the amplitude distribution structure of all the adjacent transition frequency bands to obtain the target amplitude spectrum.

[0016] Preferably, the step of combining the aperiodic random phase spectrum to synthesize the auxiliary noise sequence includes generating a chaotic sequence based on the Logistic mapping, linearly mapping the chaotic sequence to a specified interval to generate a chaotic phase spectrum, using the chaotic phase spectrum as the aperiodic random phase spectrum; assigning the frequency component corresponding to the target amplitude spectrum to the aperiodic random phase spectrum to synthesize a noise complex spectrum; and performing an inverse Fourier transform on the noise complex spectrum to generate the auxiliary noise sequence.

[0017] This invention utilizes a non-periodic random phase spectrum generated by chaos theory to endow the auxiliary noise sequence with excellent non-correlation, ensuring that it can more uniformly perturb the solution space of the objective function when performing jitter injection, thereby more effectively assisting the variational mode decomposition algorithm in capturing hidden weak physical components.

[0018] Preferably, the fusion of multi-domain composite weights includes calculating frequency domain similarity weights. The steps for calculating frequency domain similarity weights are as follows: performing Fourier transforms on each intrinsic mode function to generate a first power spectrum; calculating the Pearson correlation coefficient between each first power spectrum and the original power spectrum corresponding to the high-frequency test signal; extracting each Pearson correlation coefficient, comparing it with the value 0, and taking the maximum value to generate the frequency domain similarity weights.

[0019] Preferably, the fusion of multi-domain composite weights includes calculating time-domain feature gain weights. The steps for calculating time-domain feature gain weights are: calculating the Shannon entropy of the amplitude distribution of the time-domain sampling points corresponding to each intrinsic mode function, and performing normalization processing on all the Shannon entropies to generate normalized time-domain entropy; Calculate the time-domain feature gain weights based on the relational formula:

[0020] In the formula, Represents the time-domain feature gain weights. This represents the gain sensitivity coefficient and is greater than zero, ensuring that the time-domain feature gain weights smoothly decrease with increasing normalized time-domain entropy. This represents the normalized time-domain entropy.

[0021] This invention introduces Shannon entropy to evaluate the orderliness of modal functions and combines it with a nonlinear smoothing function to construct gain weights. It achieves smooth transition weight allocation for modal components with different characteristic attributes, effectively avoiding time-domain waveform step errors and spectral energy leakage problems caused by one-size-fits-all modal interception.

[0022] Preferably, the synthesis of the auxiliary noise sequence by combining aperiodic random phase spectrum includes: Set the control parameters and initial values ​​for the Logistic mapping based on the iterative formula:

[0023] A chaotic sequence is generated, wherein the control parameter $r$ causes the Logistic mapping to be in a chaotic state; the chaotic sequence is mapped to a phase interval to generate a chaotic phase spectrum; the chaotic phase spectrum is combined with the target amplitude spectrum to construct a conjugate symmetric complete complex spectrum; an inverse fast Fourier transform is performed on the complete complex spectrum, and the real part is taken to generate the auxiliary noise sequence.

[0024] The beneficial effects of this invention are as follows: This invention constructs a full-chain noise reduction system that spans signal feature perception, dithering injection, and multi-dimensional feature weighted fusion. Addressing the problem of interference in complex electromagnetic environments, it introduces high-resolution frequency domain transformation combined with an adaptive feature subband discrimination mechanism, establishing a dynamic derivation model for the number of decomposed modes. This effectively overcomes the over- or under-decomposition defects caused by the reliance on manual experience in conventional methods. Secondly, to address the problem that weak fault features are easily submerged by background noise under extremely low signal-to-noise ratios, and that the variational mode decomposition process is prone to getting trapped in local optima, this invention innovatively introduces a dithering injection mechanism. By synthesizing an auxiliary noise sequence with excellent non-periodicity and accurately projecting it onto the signal edge transition frequency band, the appropriate edge frequency band noise energy alters the solution space of the target optimization, assisting the decomposition algorithm in escaping local minima, thereby stably extracting hidden weak physical components and significantly enhancing the system's sensitivity to high-frequency weak features.

[0025] Furthermore, the hard threshold screening strategy of crude truncation is abandoned in the reconstruction stage. Instead, a dual-domain composite weight evaluation function that covers frequency domain structural similarity and time domain amplitude distribution characteristics is derived. The composite weight is used to perform smooth weighted fusion on each intrinsic mode function. While filtering out broadband white noise and narrowband interference, it effectively preserves the high-frequency change details corresponding to transient faults. Thus, high-quality test data with high waveform fidelity and strong feature recognition can be stably output even under harsh operating conditions. Attached Figure Description

[0026] Figure 1 This is a flowchart of a noise reduction processing method for high-frequency test signals from a distribution box. Figure 2 This is a line chart comparing feature extraction methods from different approaches. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0028] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] This invention discloses a noise reduction method for high-frequency test signals from a distribution box, referring to... Figure 1 This includes steps S1-S9: S1. Acquire the test signal and generate spectrum data.

[0030] A high-bandwidth data acquisition card was used to sample the high-frequency test signal from the distribution box; the sampling rate was set to no less than 2.56 MHz according to the Nyquist sampling theorem. The Nyquist sampling rate, from the perspective of sampling requirements, is the minimum sampling frequency required to completely reconstruct the signal, i.e., twice the highest frequency of the signal. The range of interest was set to 100 kHz to 500 kHz, with the preset sampling rate much greater than twice the highest frequency of the interest band, thus avoiding spectral aliasing and preserving sufficient signal details. A discrete-time domain sequence was obtained after sampling; the total length of the discrete-time domain sequence was set to 10240 points, and a Chirp-Z transform was performed on the discrete-time domain sequence, with the number of calculation points for the Chirp-Z transform set. Set the starting frequency to 8192 points. Set the stop frequency to 100 kHz. 500 kHz, combined with sampling frequency Calculate starting point parameters and the helix ratio parameter .

[0031] Starting point parameters The calculation satisfies the following relationship:

[0032] Helix ratio parameter The calculation satisfies the following relationship:

[0033] in, Represents the mathematical imaginary unit. Represents pi, initial frequency With termination frequency All data were acquired through a manually preset method, with a sampling frequency of [missing information]. The number of data points is obtained and calculated through the parameter configuration of the data acquisition card. Determined through a manual setting method.

[0034] For example, the discrete-time sequence is compared with the starting point parameter. and the helix ratio parameter Input the data into the standard algorithm library for calculation, and provide an example of relational calculation based on the preset data mentioned above.

[0035] Substitute the starting frequency value of 100000 and the sampling frequency value of 2560000 into the formula to obtain the starting point parameter. The specific process is as follows:

[0036] Substituting the difference of 400,000 between the termination frequency and the starting frequency into the formula, we obtain the helix ratio parameter. The specific process is as follows:

[0037] The discrete time-domain sequence, along with the starting point parameter A and the spiral ratio parameter W, is input into a standard algorithm library for calculation. The output is a complex spectrum sequence with a length of 8192. The complex spectrum sequence contains 8192 frequency points that are evenly distributed in the range of 100 kHz to 500 kHz. At this time, the specific frequency resolution is 48.83 Hz. All values ​​with a frequency point spacing of less than 50 Hz within the band of interest are extracted to form the complex spectrum data.

[0038] S2. Divide the frequency bands of interest and determine the attributes of the frequency sub-bands.

[0039] The amplitude spectrum data corresponding to the complex spectral data is obtained. This amplitude spectrum data includes the amplitude values ​​corresponding to all frequency points. The frequency band of interest is divided into multiple non-overlapping and non-intersecting frequency sub-bands, specifically 80 consecutive non-intersecting frequency sub-bands. Each frequency sub-band corresponds to a bandwidth of 5 kHz. 102 corresponding spectral amplitude points are extracted to form an amplitude sequence. For example, the first frequency sub-band covers the range of 100 kHz to 105 kHz. The spectral entropy and spectral line gradient energy corresponding to each frequency sub-band are calculated in parallel. Shannon entropy theory is introduced to construct the spectral entropy calculation formula. Shannon entropy is used to evaluate the determinism of the spectral structure; the smaller the spectral entropy value, the more deterministic the spectral structure. The probability distribution variable corresponding to the amplitude sequence is calculated. Probability distribution variables Calculate the expression that satisfies the following relation:

[0040] In the formula, This represents the specific amplitude value in the amplitude sequence. Based on probability distribution variables. Calculate the spectral entropy corresponding to each frequency sub-band. Spectral entropy Calculate the expression that satisfies the following relation:

[0041] A differential theory is introduced to construct a formula for calculating spectral line gradient energy. The differential gradient is used to evaluate spectral smoothness; steep spectral peaks in the signal will lead to inflated gradient energy values. The amplitude sequence is calculated to correspond to the first-order differential gradient sequence. First-order difference gradient sequence Calculate the expression that satisfies the following relation:

[0042] In the formula, Represents the index parameter. Represents the amplitude. Calculate the energy parameters corresponding to the difference gradient sequence. Energy parameters Calculate the expression that satisfies the following relation:

[0043] Calculate the mean spectral entropy of all frequency sub-bands. and standard deviation Set the entropy threshold. Calculate the expression that satisfies the following relation:

[0044] Calculate the mean spectral gradient energy for all frequency subbands. and standard deviation Set energy threshold The calculation formula is as follows:

[0045] The above statistical threshold setting method is derived and constructed based on the normal distribution theory, and logical conditions are used to determine each frequency sub-band; When the spectral entropy corresponding to a certain frequency sub-band is less than the entropy threshold The first condition is triggered at that time; When the spectral gradient energy corresponding to this frequency subband is greater than the energy threshold The second condition is triggered at that time; If both the first and second conditions are met, the frequency sub-band is determined to be a signal characteristic sub-band. If neither of the above two conditions is met simultaneously, it is determined to be a noise-dominant sub-band, and the index information corresponding to all signal characteristic sub-bands is recorded.

[0046] S3. Determine the number of target variables for variational mode decomposition.

[0047] Count the total number of all signal feature subbands, and denote the total number of signal feature subbands as a parameter. According to parameters Determine the number of target modes corresponding to variational mode decomposition. The calculation process for the number of target modes is a conventional logic setting process in this field, used to prevent excessive or insufficient decomposition of the signal.

[0048] Calculate the number of target modes Satisfying the relation: .

[0049] The value 2 in the formula represents the preset number of margin modes, one of which is dedicated to detecting the background noise component that runs through the entire frequency band, and the other margin mode is used to handle potentially complex signal situations.

[0050] For example, to address mode aliasing caused by the splitting of specific signal features, such as identifying that frequency sub-bands 15, 28, 29, and 56 possess signal features, the parameter... The value is set to 4. Substituting this value into the formula for the number of target modes, the number of variational mode decomposition modes is obtained. It is 6.

[0051] S4. Calculate the average spectral entropy and noise intensity.

[0052] Obtain the corresponding spectral points within all noise-dominant subbands, calculate the average power of these spectral points, and then average the average power of all noise-dominant subbands again to obtain the average background noise power. Set the reference strength parameters Reference strength parameters The calculation satisfies the following relation:

[0053] In the formula, the value 0.1 is a preset scaling factor, used to calculate the average spectral entropy corresponding to all signal feature sub-bands. Average spectral entropy This indicates the overall clarity of the test signal. A lower average spectral entropy value indicates more prominent signal features, based on the average power of the background noise. With average spectral entropy Determine the noise injection intensity The formula for calculating the noise injection intensity is derived based on a jitter injection mechanism. It aims to introduce an appropriate amount of non-periodic random noise to alter the solution space of the objective function, thereby assisting the variational mode decomposition algorithm in escaping local minima and achieving stable capture of weak signal characteristics. Noise injection intensity Calculate the expression that satisfies the following relation:

[0054] In the formula, This represents the exponential coefficient. For example, assuming there are four signal feature sub-bands with corresponding spectral entropy values ​​of 1.8, 2.1, 2, and 1.9, the average spectral entropy is calculated. The specific value is 1.95, assuming the average power of the background noise is obtained. The specific value is 0.000005. Substitute this value into the formula to calculate the reference strength parameter. The specific value is 0.0000005. In this embodiment, the exponent coefficient is... The value is set to 1.2, and the result is obtained by substituting it into the mapping function to calculate the noise injection intensity. The specific result is 0.00000519.

[0055] S5. Generate the target amplitude spectrum and allocate energy.

[0056] The frequency band boundaries of each signal characteristic sub-band are extracted. Based on these boundaries, corresponding adjacent transition frequency bands are determined. A Gaussian window function or a Tukey window function is used to generate an amplitude distribution structure within the corresponding adjacent transition frequency bands, concentrating energy in the edge regions near the signal characteristic sub-bands. A normalization operation is performed on the amplitude distribution structures corresponding to all adjacent transition frequency bands to obtain the target amplitude spectrum. The normalization operation ensures that the total energy of the target amplitude spectrum is equal to the noise injection intensity obtained in the previous steps. For example, assuming the calculated noise injection intensity is 5, the sum of squares of all amplitudes within the target amplitude spectrum after normalization is also equal to 5.

[0057] S6. Synthesize the noise complex spectrum and generate the superimposed signal.

[0058] A non-periodic random phase spectrum is added to the target amplitude spectrum to synthesize a complex noise spectrum. By introducing phase randomness, the periodicity of the noise is destroyed. In this embodiment, a chaotic phase spectrum is preferred to ensure the uniformity and non-correlation of energy distribution during jitter injection. A chaotic sequence is generated according to the Logistic mapping relationship. The corresponding Logistic mapping iterative relationship is as follows:

[0059] In the formula Represents the value of the current iteration. Represents the value from the previous iteration. This represents the control parameter. In this embodiment, the control parameter is set. The value is 3.99, set the initial value. The value is set to 0.5, and the first iteration value is 0.9975 when substituted into the relational expression.

[0060] Let the chaotic sequence be denoted as a sequence. ,sequence All internal values ​​fall within the range of 0 to 1. A transformation formula is applied to linearly map the chaotic sequence to a specified interval, from negative pi to positive pi. The transformation formula is as follows:

[0061] In the formula, Representative sequence Specific point values ​​inside, This represents the numerical value of the corresponding phase spectrum sequence. For example, when... When the value is 0.5, the corresponding mapped value is 0. When the value is 1, the corresponding mapped value is ,when When the value is 0, the corresponding mapped value is - After traversing the mapping, a chaotic phase spectrum sequence with a length of 8192 is obtained.

[0062] To ensure that the inverse transform outputs a real number sequence, a conjugate symmetric spectrum needs to be constructed, and the number of calculation points corresponding to the fast Fourier transform is set. Set the number of calculation points. The specific value is 16384. This value is set to fully cover the frequency range corresponding to the preset sampling rate, combining the target amplitude spectrum with the chaotic phase spectrum to construct a complete complex spectrum. Initialize the complete complex spectrum All internal values ​​are 0; calculate the corresponding target index for each frequency point within the target amplitude spectrum. The target amplitude spectrum is denoted as Based on the target index Performing a spectrum assignment operation satisfies the following relation:

[0063] In the formula, Represents the original index within the target amplitude spectrum. Represents the imaginary unit, for the complete complex number spectrum. Construct the conjugate symmetric part; the construction relation for conjugate symmetry is as follows:

[0064] Letters in the formula Represents the spectrum index parameter, function This represents the operation of finding the complex conjugate while ensuring that the imaginary parts corresponding to the zero frequency and the Nyquist frequency are both zero. It calls standard algorithm libraries to calculate the complete complex spectrum. Perform an inverse fast Fourier transform (IFFT), take the real part of the IFFT result to generate an auxiliary noise sequence, extract the first 10240 data points of the auxiliary noise sequence to match the length of the high-frequency test signal, and perform time-domain point-by-point superposition of the extracted auxiliary noise sequence and the high-frequency test signal to finally obtain the superimposed signal required for noise reduction processing.

[0065] S7. Perform variational mode decomposition to extract mode functions.

[0066] Variational mode decomposition is used to adaptively decompose the signal based on its own time-scale characteristics, extract the previously determined number of target modes, take the superimposed signal as input parameters, perform variational mode decomposition on the superimposed signal with the target number of modes, and extract multiple intrinsic mode functions after the decomposition operation.

[0067] S8. Calculate the frequency domain similarity and time domain gain weight.

[0068] Constructing the power spectrum of signal feature templates The template power spectrum has a total length of 8192 points. Values ​​within the frequency range corresponding to the signal characteristic sub-bands are set to the original signal power spectrum values, while values ​​within the frequency range corresponding to the noise-dominant sub-bands are set to 0. Fourier transforms are then performed on each intrinsic mode function to generate the first power spectrum. Pearson correlation coefficient is introduced to assess spectral structure similarity, and the first power spectrum is calculated. With template power spectrum Pearson correlation coefficient between The Pearson correlation coefficient is extracted and compared with 0, and the maximum value is taken to generate the frequency domain similarity weight. .

[0069] For example, suppose a certain eigenmode function corresponds to the Pearson correlation coefficient. The specific value is 0.92, which corresponds to the frequency domain similarity weight of this intrinsic mode function. The specific value is 0.92; assuming the other eigenmode function corresponds to the Pearson correlation coefficient. The specific value is -0.3, which corresponds to the frequency domain similarity weight. The specific value is 0.

[0070] Calculate the amplitude distribution and Shannon entropy of each intrinsic mode function at time-domain sampling points based on time-domain characteristics. Generate a set of Shannon entropies containing multiple intrinsic mode functions, and denote the maximum value within the set as . Let the minimum value inside the set be denoted as Shannon entropy corresponding to all intrinsic mode functions Perform normalization processing to generate normalized time-domain entropy The normalized calculation formula is as follows:

[0071] For example, suppose the maximum value inside the set is... The specific value is 3.5, the minimum value within the set. The specific value is 0.7. When the specific value of a certain Shannon entropy is 3.5, the normalized time-domain entropy is obtained by substituting it into the relational formula. The specific value is 1; when a certain Shannon entropy When the specific value is 0.7, the normalized time-domain entropy is obtained by substituting it into the relational expression. The specific value is 0.

[0072] Constructing time-domain characteristic gain weights based on the sigmoid gain function Calculate the relational expression and the time-domain feature gain weight. The calculation formula is as follows:

[0073] In the formula Represents the time-domain feature gain weights. Represents the gain sensitivity coefficient. This represents the normalized time-domain entropy.

[0074] For example, suppose the gain sensitivity coefficient The specific value is 1.5, when the normalized time-domain entropy... When the specific value is 1, the temporal feature gain weight is obtained by substituting it into the relational formula. =0; When normalized time-domain entropy When the value is 0, substitute it into the relational expression to calculate and obtain the time-domain feature gain weight. =1; When normalized time-domain entropy When the value is 0.5, the temporal feature gain weight is obtained by substituting it into the relational expression. It is 0.68.

[0075] S9. Perform weighted summation to reconstruct the noise reduction test signal.

[0076] An all-zero array is initialized as the initial reconstructed signal. The frequency-domain similarity weights and time-domain feature gain weights are multiplied to obtain a dual-domain composite weight. This dual-domain composite weight is then used to perform a weighted summation of each intrinsic mode function. The reconstructed signal sequence is then obtained. The calculation formula is as follows:

[0077] In the formula Represents the reconstructed signal sequence. Represents the sequence of eigenmode functions. For the first The dual-domain composite weights corresponding to each intrinsic mode function, the dual-domain composite weights The calculation satisfies: .

[0078] The final noise reduction test signal is obtained by traversing all intrinsic mode functions and performing cumulative reconstruction.

[0079] A verification scheme was adopted, which included datasets corresponding to three typical electrical fault signals. The frequencies of these signals were located near 150 kHz, 280 kHz, and 410 kHz, respectively. Gaussian white noise of different intensities was injected into the clean signals to form three test sets with signal-to-noise ratios of 5 dB, 0 dB, and -5 dB, respectively. Each set contained 500 samples.

[0080] In a preferred embodiment of the present invention, a heterogeneous hardware acceleration and sliding window downsampling mechanism are introduced at the system's underlying layer to address the aforementioned computationally intensive steps. Specifically, a field-programmable gate array (FPGA) or a dedicated digital signal processor (DSP) is used to construct the underlying data pipeline. A hardware multiply-accumulator array embedded in the FPGA is specifically responsible for performing intensive matrix operations such as real-time caching of the 2.56MHz sampling stream, Chirp-Z transform, and inverse IFFT transform. Subsequently, the extracted high signal-to-noise ratio (SNR) feature subband data is reported in batches to the main control microprocessor using a time-sliding window mechanism to perform variational mode decomposition and weight reconstruction calculations. This hardware-software co-operational microservice architecture ensures complete capture of broadband electromagnetic signals under extremely low SNR conditions while keeping the transient computing load of the edge computing terminal within safe boundaries, thereby guaranteeing the long-term stable operation of the high-frequency noise reduction system in industrial settings.

[0081] The complete scheme of this invention is designated as Scheme 1, with three sets of control experiments. Scheme 2 uses a fixed number of modes to perform variational mode decomposition without subband identification or noise injection. Scheme 3 removes the noise injection step. Scheme 4 replaces it with mode reconstruction based solely on the frequency domain correlation coefficient. The accuracy of fault feature extraction for each scheme under different signal-to-noise ratios is recorded below: When the signal-to-noise ratio is 5 dB, the accuracy rates of schemes one through four are 98.6%, 92.3%, 96.8%, and 97.5%, respectively. When the signal-to-noise ratio drops to 0 dB, the accuracy rates of the four schemes are 95.2%, 83.1%, 89.2%, and 91.4%, respectively. In a strong noise environment of -5 dB, the accuracy rates of the four schemes decreased to 91.5%, 71.6%, 80.5%, and 85.3%, respectively. Scheme 1 achieved the highest accuracy under all signal-to-noise ratio conditions, with corresponding accuracy improvements of 6.3%, 12.1%, and 19.9% ​​compared to Scheme 2, respectively, verifying the effectiveness of each improvement step.

[0082] like Figure 2 As shown, the feature extraction performance of each scheme is compared under different signal-to-noise ratios.

[0083] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A method for noise reduction processing of high-frequency test signals from a distribution box, characterized in that, include: Acquire high-frequency test signals and perform frequency domain transformation to divide the frequency band of interest into multiple frequency sub-bands, and calculate the frequency domain statistical characteristics and energy distribution characteristics of each frequency sub-band; The attributes of each frequency sub-band are determined based on a preset threshold. Frequency sub-bands that meet the preset conditions are identified as signal feature sub-bands, and frequency sub-bands that do not meet the preset conditions are identified as noise-dominant sub-bands. The average background noise power is calculated based on the noise-dominant sub-bands, and the target number of variational mode decomposition is adaptively determined based on the number of signal feature sub-bands. The noise injection intensity is determined based on the average power of the background noise and the characteristics of the signal characteristic subband, and the target amplitude spectrum is generated based on the noise injection intensity. An auxiliary noise sequence is synthesized by combining a non-periodic random phase spectrum, and the auxiliary noise sequence is superimposed with the high-frequency test signal to generate a superimposed signal to perform jitter injection; The superimposed signal is subjected to variational mode decomposition to extract multiple intrinsic mode functions based on the target number of modes. The multiple intrinsic mode functions are then weighted and reconstructed by multi-domain composite weights to generate a noise-reduced test signal.

2. The noise reduction method for high-frequency test signals from a distribution box according to claim 1, characterized in that, The frequency domain transformation includes: The high-frequency test signal is sampled using a sampling rate of not less than 2.56 MHz; Perform Chirp-Z transform on the high-frequency test signal and set the number of calculation points to 8192. Extract all values ​​within the band of interest whose frequency spacing is less than 50 Hz to form complex spectrum data.

3. The noise reduction method for high-frequency test signals based on a distribution box according to claim 1, characterized in that, The calculation of the frequency domain statistical characteristics and energy distribution characteristics of each frequency sub-band, and the determination of the attributes of each frequency sub-band, includes: calculating the spectral entropy and spectral line gradient energy of each frequency sub-band as the frequency domain statistical characteristics and the energy distribution characteristics, respectively. The sum of the mean spectral entropy of all frequency sub-bands and twice the standard deviation is used as the entropy threshold, and the sum of the mean spectral gradient energy of all frequency sub-bands and three times the standard deviation is used as the energy threshold. The entropy threshold and the energy threshold constitute the preset threshold. When the spectral entropy of a certain frequency sub-band is less than the entropy threshold and the spectral gradient energy of the certain frequency sub-band is greater than the energy threshold, the certain frequency sub-band is determined to be the signal characteristic sub-band.

4. The noise reduction method for high-frequency test signals based on a distribution box according to claim 1, characterized in that, The adaptive determination of the target number of modes for variational mode decomposition includes: Calculate the total number of signal feature sub-bands to generate the total number of signal feature sub-bands; The number of target modes is set to be equal to the sum of the number of signal feature subbands plus the value 2.

5. The noise reduction method for high-frequency test signals based on a distribution box according to claim 1, characterized in that, The step of determining the noise injection intensity based on the average background noise power and the average spectral entropy of the signal characteristic subbands includes: Set the reference strength parameter to be equal to the average power of the background noise multiplied by 0.1; Based on the mapping function Calculate the noise injection intensity; In the formula, Indicates the noise injection intensity. Represents the reference strength parameter. This represents an exponential coefficient that is greater than zero, ensuring that the noise injection intensity amplifies nonlinearly with the increase of the average spectral entropy of the signal's characteristic subbands. The average spectral entropy of the characteristic subband of the signal.

6. The noise reduction method for high-frequency test signals based on a distribution box according to claim 1, characterized in that, The generation of the target amplitude spectrum includes: Extract the frequency band boundaries of each signal feature sub-band to determine the corresponding adjacent transition frequency band; A Gaussian window function is used to generate an amplitude distribution structure in all the adjacent transition frequency bands; The target amplitude spectrum is obtained by performing a normalization operation on the amplitude distribution structure of all adjacent transition frequency bands.

7. The noise reduction method for high-frequency test signals based on a distribution box according to claim 1, characterized in that, The auxiliary noise sequence synthesized by combining aperiodic random phase spectrum includes: A chaotic sequence is generated based on the Logistic mapping, and the chaotic sequence is linearly mapped to a specified interval to generate a chaotic phase spectrum, which is then used as the non-periodic random phase spectrum. The non-periodic random phase spectrum is assigned the frequency component corresponding to the target amplitude spectrum to synthesize a noise complex spectrum; The auxiliary noise sequence is generated by performing an inverse Fourier transform on the noise complex spectrum.

8. The noise reduction method for high-frequency test signals based on a distribution box according to claim 1, characterized in that, The fusion of multi-domain composite weights includes calculating the frequency domain similarity weight. The steps for calculating the frequency domain similarity weight are as follows: Perform Fourier transforms on each intrinsic mode function to generate the first power spectrum; Calculate the Pearson correlation coefficient between each first power spectrum and the original power spectrum corresponding to the high-frequency test signal; Each Pearson correlation coefficient is extracted and compared with the value 0, and the maximum value is taken to generate the frequency domain similarity weight.

9. The noise reduction method for high-frequency test signals based on a distribution box according to claim 1, characterized in that, The fusion of multi-domain composite weights includes calculating the temporal feature gain weights. The steps for calculating the temporal feature gain weights are as follows: Calculate the Shannon entropy of the amplitude distribution of each intrinsic mode function at the time-domain sampling point, and perform normalization on all the Shannon entropies to generate normalized time-domain entropy; Calculate the time-domain feature gain weights based on the relational formula: ； In the formula, Represents the time-domain feature gain weights. This represents the gain sensitivity coefficient and is greater than zero, ensuring that the time-domain feature gain weights smoothly decrease with increasing normalized time-domain entropy. This represents the normalized time-domain entropy.

10. The noise reduction method for high-frequency test signals from a distribution box according to claim 1, characterized in that, The auxiliary noise sequence synthesized by combining aperiodic random phase spectrum includes: Set the control parameters and initial values ​​for the Logistic mapping based on the iterative formula: Generate a chaotic sequence, wherein the control parameter $r$ causes the Logistic map to be in a chaotic state; The chaotic sequence is mapped to a phase interval to generate a chaotic phase spectrum; By combining the chaotic phase spectrum with the target amplitude spectrum, a conjugate symmetric complete complex spectrum is constructed. Perform an inverse fast Fourier transform on the complete complex spectrum and extract the real part to generate the auxiliary noise sequence.

Images