A wind power weak feature extraction method based on cascaded double feedback stochastic resonance

Abstract

This invention provides a method for extracting weak features of wind power based on cascaded dual-feedback stochastic resonance, belonging to the field of wind turbine condition monitoring and fault early warning technology. It can at least partially solve the problems of existing technologies, such as linear filtering's inability to simultaneously suppress weak harmonics and common-mode features, poor adaptability of single-stage stochastic resonance to high-frequency signals, and severe parameter coupling. The invention includes: synchronously acquiring multi-channel electrical signals from the wind turbine and preprocessing them to obtain standardized signals; performing bandpass tracking filtering on the standardized signals; performing piecewise folding time-scale compression on the component signals to map high-frequency components to the effective frequency band of the stochastic resonance; constructing a cascaded coupling between the pre-stage bistable potential and the post-stage Gaussian potential, and achieving a secondary transfer of noise energy to the target frequency band through real-time optimization using an online optimization algorithm; and achieving early fault detection through self-learning. This invention can effectively enhance weak features and achieve reliable online early warning of early defects in wind power converters under extremely low signal-to-noise ratio conditions.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine condition monitoring and fault early warning technology, specifically involving a method for extracting weak features of wind power based on cascaded dual feedback random resonance. Background Technology

[0002] During high-frequency switching, wind turbine converters inevitably generate harmonic currents and common-mode voltages. When power devices age, filter inductors become magnetically saturated, or grounding brushes experience uneven wear, weak sidebands appear in the harmonic spectrum, and the common-mode current pulse amplitude only increases by tens of milliamperes. These early characteristic frequencies often fall in the 2 to 50 kHz range, with amplitudes more than 60 dB lower than the fundamental frequency, and are severely masked by switching noise, electromagnetic interference, and environmental noise. If these frequencies are not identified in time, they will lead to secondary faults such as insulation breakdown and bearing erosion, resulting in downtime losses and high maintenance costs.

[0003] Existing research has the following limitations.

[0004] In terms of linear filtering methods, current engineering practices commonly employ bandpass filters combined with Fast Fourier Transform (FFT) or wavelet packet decomposition (WPD) for harmonic and common-mode feature extraction. However, while linear filters suppress noise, they also attenuate weak harmonics and common-mode pulses in the same frequency band, resulting in a simultaneous suppression phenomenon. When the signal-to-noise ratio (SNR) is below -20 dB, the harmonic distortion rate detection error exceeds 30%, and the common-mode pulse detection rate is below 62%, making it difficult to meet early warning requirements. Furthermore, wavelet methods rely on basis function selection, have poor adaptability to high-frequency non-stationary pulses, and involve large computational loads, making them unsuitable for embedded online implementation.

[0005] In terms of stochastic resonance methods, existing stochastic resonance methods mostly employ single-stage bistable or Gaussian potentials, requiring secondary sampling for frequency reduction, resulting in low time resolution. The potential well depth and width parameters are heavily coupled, making it difficult to simultaneously accommodate both harmonic and common-mode frequencies. The output is prone to saturation, lacks a feedback memory mechanism, and exhibits output fluctuations exceeding 3 dB below -25 dB, with low utilization of heavy-tail noise.

[0006] Regarding the cascaded stochastic resonance method, although cascaded dual-feedback stochastic resonance has achieved a certain gain in low-frequency impulse signals, it has not solved the problems of high-frequency mapping, multi-channel synchronization, and quantification of harmonic distortion rate and common-mode energy ratio, and cannot be directly used for online monitoring of wind power harmonics and common modes.

[0007] Therefore, how to design a method that can simultaneously enhance the weak characteristics of harmonics and common modes under extremely low signal-to-noise ratio conditions and realize online reliable early warning of early aging defects in wind power converters is a technical problem that urgently needs to be solved in the field of wind power condition monitoring. Summary of the Invention

[0008] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a method for extracting weak features of wind power based on cascaded double feedback stochastic resonance.

[0009] To achieve the above objectives, this invention provides a method for extracting weak features of wind power based on cascaded dual-feedback stochastic resonance, comprising: The multi-channel electrical signals of the wind turbine are synchronously acquired and preprocessed to obtain standardized signals. The multi-channel electrical signals include three-phase voltage, three-phase current, DC bus voltage and grounding common-mode current. The standardized signal is subjected to bandpass tracking filtering to extract the harmonic component signal of the harmonic frequency band and the common-mode component signal of the common-mode pulse frequency band, respectively. The harmonic component signal and the common-mode component signal are subjected to piecewise folding time-scale compression, and the high-frequency components are mapped to the effective frequency band of random resonance to obtain the compressed signal. A cascaded double-feedback stochastic resonance model is constructed. The front-stage subsystem of the cascaded double-feedback stochastic resonance model adopts a bistable potential function, and the rear-stage subsystem adopts a Gaussian potential function. The front-stage subsystem and the rear-stage subsystem are cascaded and coupled. A first negative feedback coefficient is introduced into the front-stage subsystem and a second negative feedback coefficient is introduced into the rear-stage subsystem to form a double-loop negative feedback structure. An online optimization algorithm is used to optimize the parameters of the first negative feedback coefficient, the second negative feedback coefficient, the bistable potential function, and the Gaussian potential function in real time, so that the noise energy is transferred to the target frequency band in a secondary manner, resulting in an enhanced cascaded output signal. The cascaded output signals are subjected to parallel Fast Fourier Transform and S-Transform to extract the harmonic distortion rate and common-mode energy ratio. When the harmonic distortion rate or the common-mode energy ratio exceeds the dynamic threshold, an early fault alarm is triggered.

[0010] Furthermore, the preprocessing includes: The multi-channel electrical signal is subjected to sliding window median filtering. When the deviation between the sampling point and the median in the window exceeds the product of a preset deviation multiple and the standard deviation, the sampling point is replaced by the median in the window. The DC component in the multi-channel electrical signal is calculated and subtracted in real time according to a preset time period, so that the mean of the multi-channel electrical signal is zero. Each electrical quantity is divided by its corresponding rated base value and then converted to per-unit value to obtain the standardized signal whose amplitude range does not exceed the preset per-unit upper limit.

[0011] Furthermore, the bandpass tracking filter includes: The standardized signal is simultaneously fed into a harmonic bandpass filter bank and a common-mode bandpass filter bank. The passband range of the harmonic bandpass filter bank is from a preset lower harmonic frequency to a preset upper harmonic frequency, and the passband range of the common-mode bandpass filter bank is from a preset lower common-mode frequency to a preset upper common-mode frequency. The harmonic bandpass filter bank outputs the harmonic component signal, the common-mode bandpass filter bank outputs the common-mode component signal, and the timestamp alignment error between the harmonic component signal and the common-mode component signal does not exceed one sampling point.

[0012] Furthermore, the segmented folding timescale compression includes: The harmonic component signal and the common-mode component signal are respectively truncated by sliding window with a preset window length and a preset step size to obtain the data segment within the window; The amplitude spectrum is obtained by performing a fast Fourier transform on the data segment within the window; Based on the online-set low-frequency compression ratio threshold and high-frequency compression ratio threshold, retain the spectral lines in the amplitude spectrum that satisfy the following conditions: ; in, The low-frequency compression ratio threshold is... The high-frequency compression ratio threshold is... For frequency The amplitude at that point, The peak value of the amplitude spectrum is set to zero for spectral lines that do not meet the conditions. The compressed signal is obtained by performing an inverse fast Fourier transform on the preserved spectral lines.

[0013] Furthermore, the online update method for the low-frequency compression ratio threshold and the high-frequency compression ratio threshold is as follows: When the average wind speed change within a preset time period reaches a preset wind speed threshold, or the percentage change in active power reaches a preset power threshold, or the change in converter casing temperature reaches a preset temperature threshold, the threshold update is initiated. The low-frequency compression ratio threshold and the high-frequency compression ratio threshold are adjusted by increasing or decreasing according to a preset adjustment step size. In the high-power region, the low-frequency compression ratio threshold is decreased and the high-frequency compression ratio threshold is increased, and the opposite adjustment is made in the low-power region. After each adjustment, ensure that the low-frequency compression ratio threshold and the high-frequency compression ratio threshold both fall within their respective preset value ranges, and that the difference between the two is not less than the preset minimum difference. If a preset number of consecutive updates causes the harmonic distortion rate or the common-mode energy ratio to jump beyond a preset jump threshold, it will automatically revert to the previous stable value.

[0014] Furthermore, the expression for the bistable state function is: ; in, Let be the potential well depth parameter of the bistable state function. Let be the potential well width parameter of the bistable state function. These are the state variables of the preceding subsystem; The expression for the Gaussian potential function is: ; in, Let be the potential well depth parameter of the Gaussian potential function. Let be the potential well width parameter of the Gaussian potential function. These are the state variables of the subsequent subsystem; First negative feedback coefficient The output of the subsequent subsystem is fed back to the input of the preceding subsystem, and the second negative feedback coefficient... The output of the subsequent subsystem is fed back to the input of the subsequent subsystem itself. The value range of the first negative feedback coefficient is from a preset first feedback lower limit to zero, and the value range of the second negative feedback coefficient is from a preset second feedback lower limit to zero.

[0015] Furthermore, the online optimization algorithm is executed as follows: An improved particle swarm optimization algorithm is used to simultaneously optimize six parameters: the first negative feedback coefficient, the second negative feedback coefficient, the potential well depth parameter and the potential well width parameter of the bistable potential function, and the potential well depth parameter and the potential well width parameter of the Gaussian potential function. The objective function of the improved particle swarm optimization algorithm is: ; in, For harmonic signal-to-noise gain, Common-mode energy ratio, and For preset weighting coefficients and ; The improved particle swarm algorithm sets the number of particles to a preset number and the maximum number of iterations to a preset iteration limit.

[0016] Furthermore, the harmonic distortion rate is calculated as follows: ; in, Frequency within the harmonic band The amplitude at that point, and These are the lower and upper limits of the harmonic frequency band, respectively. The fundamental amplitude; The common-mode energy ratio is calculated as follows: ; in, Frequency within the common-mode band The amplitude at that point, and These are the lower and upper limits of the common-mode frequency band, respectively. This represents the total energy.

[0017] Furthermore, the dynamic threshold is generated as follows: Take historical data of the harmonic distortion rate and the common-mode energy ratio within a preset sliding window period; Calculate the preset quantile value and standard deviation of the historical data, and multiply the preset quantile value and preset standard deviation by the sum of the standard deviations to obtain the dynamic threshold. The sliding window slides forward once according to a preset sliding period to achieve self-learning update of the dynamic threshold.

[0018] Furthermore, the triggering condition for the early fault alarm is: The harmonic distortion rate exceeds the corresponding dynamic threshold, or the common-mode energy ratio exceeds the corresponding dynamic threshold; Furthermore, after the aforementioned over-limit state continues for several preset consecutive calculation cycles, the early fault alarm is triggered; The information in the early fault alarm includes the type of out-of-limit indicator, the out-of-limit amplitude, and the suggested repair location.

[0019] The beneficial effects of this invention are as follows: This invention constructs a cascaded double-feedback stochastic resonance model, employing a cascaded topology of a pre-stage bistable potential function and a post-stage Gaussian potential function, combined with a dual-loop negative feedback mechanism, to achieve a secondary transfer of noise energy to the target signal frequency band. This effectively solves the problem of suppressing weak harmonics and common-mode characteristics simultaneously in strong noise environments using traditional linear filtering methods.

[0020] This invention employs a segmented folding time-scale compression technique to uniformly map high-frequency harmonics and common-mode pulses to the effective frequency band of stochastic resonance, thus overcoming the technical bottleneck of poor adaptability of single-level stochastic resonance models to high-frequency signals and severe parameter coupling.

[0021] This invention optimizes six-dimensional parameters in real time using an online optimization algorithm, and combines a dynamic threshold self-learning mechanism to accurately extract harmonic distortion rate and common-mode energy ratio even under extremely low signal-to-noise ratio conditions. This provides reliable online early warning for early defects such as aging of power devices, saturation of filter inductors, and wear of grounding brushes in wind power converters. Attached Figure Description

[0022] Figure 1 This is a flowchart of a wind power weak feature extraction method based on cascaded dual feedback stochastic resonance, according to an embodiment of the present invention. Figure 2 This is a functional architecture diagram of the signal processing platform according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the principle of segmented folding timescale compression in an embodiment of the present invention. Figure 4 This is a logic diagram for feature extraction and dynamic threshold diagnosis in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and beneficial effects of this application clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, further illustrates this application. It should be understood that the specific embodiments described in this specification are merely for explaining this application and are not intended to limit it.

[0024] The wind power weak feature extraction method based on cascaded double-feedback stochastic resonance of the present invention is implemented on an embedded signal processing platform. The signal processing platform adopts a DSP and FPGA collaborative architecture design. The FPGA is responsible for front-end signal preprocessing tasks such as multi-channel synchronous acquisition, bandpass tracking filtering, and segmented folding compression, while the DSP is responsible for core algorithm tasks such as cascaded double-feedback stochastic resonance model operation, online parameter optimization, and feature index calculation.

[0025] The signal processing platform's hardware configuration includes a multi-channel synchronous acquisition board and an algorithm processing board. The multi-channel synchronous acquisition board is equipped with a high-precision analog-to-digital converter, anti-aliasing filter circuitry, and a time synchronization module. The algorithm processing board is equipped with a digital signal processor and a programmable gate array (PGA), and data exchange is achieved through a dual-port random access memory (RAM). Communication interfaces utilize Gigabit Ethernet and industrial fieldbus to ensure real-time and reliable data transmission.

[0026] See Figure 2 The functional architecture of the signal processing platform includes five core components: a multi-channel synchronous acquisition module, a bandpass tracking filter module, a time-scale compression module, a cascaded dual-feedback stochastic resonance module, and a feature extraction and diagnosis module.

[0027] The multi-channel synchronous acquisition module is responsible for synchronously sampling the three-phase voltage, three-phase current, DC bus voltage, and grounding common-mode current of the wind turbine. It performs preprocessing operations such as digital filtering, zero drift correction, and per-unit transformation on the acquired raw signals and outputs a standardized signal matrix in a unified format.

[0028] The bandpass tracking filter module is responsible for feeding the standardized signal into the harmonic bandpass filter bank and the common-mode bandpass filter bank respectively, separating the component signals of the harmonic frequency band and the common-mode pulse frequency band, and providing input data for frequency band separation for subsequent time-scale compression and stochastic resonance processing.

[0029] The time-scale compression module is responsible for performing segmented folding compression operations, mapping high-frequency harmonics and common-mode pulses to the effective frequency band of random resonance, and generating structured feature matrices according to operating conditions.

[0030] The cascaded dual-feedback stochastic resonance module is the core component of this invention. It includes a front-stage bistable stochastic resonance unit, a rear-stage Gaussian stochastic resonance unit, a dual-loop negative feedback network, and an online parameter optimization unit. It achieves weak signal enhancement through secondary transfer of noise energy.

[0031] See Figure 4 The feature extraction and diagnosis module is responsible for performing fast Fourier transform and S-transform on the cascaded output signal, calculating the harmonic distortion rate and common-mode energy ratio, and outputting early fault identification results in combination with the dynamic threshold self-learning mechanism.

[0032] Example 1 This embodiment uses a 2.5 MW doubly-fed induction generator (DFIG) full-power converter as an application scenario. The rated line voltage of the unit is 3kV, the DC bus voltage is 3.3kV, and the switching frequency is 2.45 kHz. During normal operation, the converter power devices may age, the filter inductor may become magnetically saturated, and the grounding brush may experience uneven wear. The harmonic sidebands and common-mode pulses generated by these early defects have small amplitudes and are easily overwhelmed by the strong noise environment. Therefore, the method of this invention is needed for effective extraction and early warning.

[0033] See Figure 1 The weak feature extraction method in this embodiment includes the following steps: Step S1: Synchronously acquire multi-channel electrical signals from the wind turbine and perform preprocessing.

[0034] A multi-channel synchronous acquisition module configures the sensing layer. For voltage acquisition, a 10 kV capacitive voltage divider sensor with an accuracy of 0.1% and a bandwidth of 100 MHz is deployed on the three-phase AC side; a 50 kV isolated differential probe with a bandwidth of 50 MHz is used on the DC bus. For current acquisition, a Rogowski coil with a frequency response range of 0.1 Hz to 100 MHz is used on the three-phase AC side; a high-frequency current clamp with a frequency response range of 1 MHz to 100 MHz is used for ground common-mode current, with a minimum measurable current of 1 mA. The impedance between all sensor housings and the cabin grounding point is less than 2 milliohms to suppress common-mode to differential-mode conversion.

[0035] The synchronous acquisition hardware uses a 24-bit Delta-Sigma analog-to-digital converter with a maximum single-channel sampling rate of 256 kS / s. In this embodiment, the sampling rate is uniformly downsampled to 10 kS / s, which can accommodate the acquisition requirements of the 50th harmonic component and 50 kHz switching ripple. The FPGA is locally configured with a 10 MHz temperature-controlled crystal oscillator with a frequency stability of ±0.05 ppm. Time synchronization is performed using the IEEE 1588 v2 precise time protocol, with a synchronization error of less than 200 ns, ensuring that the phase error between channels within the same fundamental period is less than 0.01 degrees.

[0036] The analog front-end is equipped with a third-order Butterworth anti-aliasing filter with a cutoff frequency of 4 kHz and out-of-band attenuation greater than 55dB. The analog-to-digital converter pre-stage is equipped with a 5V bidirectional transient voltage suppressor array, which can withstand 4 kV combined wave impacts to prevent damage to the acquisition link from lightning strikes or switching surges.

[0037] Digital preprocessing includes two steps: sliding window median filtering and zero drift correction. A 1024-point sliding window is used for median filtering, when the sampling points... With window value Replacement is performed when the deviation meets the following conditions: ; in, The standard deviation of the signal within the window is used. Sampling points that meet the above conditions are replaced with the window value to eliminate spike interference. For zero drift correction, the DC component of each channel is calculated and subtracted in real time every 10 seconds to ensure that the mean of the input signal of the subsequent stochastic resonance model is zero.

[0038] Per-unit conversion divides each electrical quantity by its corresponding rated base value to obtain a standardized signal in per-unit form. The rated base value is set as follows: the voltage base value is taken as the peak value of the unit's rated line voltage, i.e. kV; the base current value is taken as the peak value of the unit's rated current, i.e. A; The power base value is taken as the rated apparent power of the unit, i.e. kVA. The data after per-unit transformation is stored in single-precision floating-point format, with an amplitude range not exceeding ±2 pu, to preserve overload condition information. The mean of the per-unit transformed data is zero, and the variance is no greater than 1.5, meeting the requirements of the subsequent stochastic resonance model for zero-mean and bounded input.

[0039] For data packetization and caching, one data block is generated per second. Each data block contains single-precision floating-point data from 6 channels multiplied by 10,000 sampling points, an 8-byte GPS timestamp, and a 2-byte quality tag. A quality tag value of 0 indicates normal data, 1 indicates missing data, and 2 indicates abnormal data. Data blocks are uploaded in real-time via Gigabit Ethernet in UDP frame format. When the communication link is interrupted, the FPGA board's onboard 8 GB embedded multimedia storage card performs rolling caching for at least 72 hours. The final result is a unified format 6-column raw multi-channel broadband signal matrix. : ; In this matrix, the first column is the time label, the second column is the positive sequence voltage data, the third column is the active power data, the fourth column is the reactive power data, the fifth column is the active current data, and the sixth column is the reactive current data. There are a total of [number missing]. One sampling point.

[0040] Step S2: Perform bandpass tracking filtering on the standardized signal.

[0041] The normalized signal matrix obtained in step S1 Two sets of digital bandpass filters are simultaneously fed in. The harmonic bandpass filter bank has a passband range of 2 kHz to 50 kHz, an 8th order filter, an in-band ripple of no more than 0.2 dB, and an out-of-band attenuation of no less than 60 dB / dec. The common-mode bandpass filter bank has a passband range of 1 kHz to 10 kHz, a 6th order filter, an in-band ripple of no more than 0.3 dB, and an out-of-band attenuation of no less than 60 dB / dec.

[0042] After filtering, the harmonic sub-matrices are obtained respectively. With common modulus submatrix The dimensions of both submatrices are maintained at 1. The timestamp alignment error does not exceed one sampling point. Harmonic submatrix Includes harmonic component signals in the frequency band from 2 kHz to 50 kHz, common mode sub-matrix It includes common-mode pulse component signals in the frequency band from 1 kHz to 10 kHz.

[0043] Step S3: Perform segmented folding time-scale compression on the component signals.

[0044] First, the compression ratio threshold is set online. The system provides a low-frequency compression ratio threshold. With high-frequency compression ratio threshold Two control parameters. The value range is from 0.001 to 0.1. The value ranges from 0.5 to 2.0, and always satisfies... When the unit is powered on, the factory default values ​​are read from the non-volatile memory. and This default value has been type-tested and calibrated to be suitable for most rated operating conditions.

[0045] During operation, when the 10-minute average wind speed change is not less than 1 m / s, or the active power percentage change is not less than 5%, or the converter casing temperature change is not less than 5 degrees Celsius, or a remote command is received, the edge computing unit initiates a threshold update in the next 1-second data block interval. The update is performed with a default step size of 0.002. and Adjustments should be made accordingly. In areas with high wind speeds and high power output, reduce... And increase synchronously To preserve high-frequency impact characteristics; in low-wind-speed and low-power operating areas, reverse adjustments are made to suppress low-frequency drift. Ensure after each adjustment... and The values ​​still fall within their respective ranges, and the difference between them is not less than 0.1. The update cycle is 5 minutes by default, which is shortened to 1 minute during gusts or severe power fluctuations. The new threshold is immediately written to the double-buffered register and takes effect in the next data block. If three consecutive updates cause subsequent harmonic distortion rate or common-mode energy ratio to abnormally jump by more than 50%, the system will automatically revert to the previous stable value and report the anomaly.

[0046] Then, the harmonic submatrix and common modulus submatrix Parallel segmented compression is performed. A sliding window is used with a window length of 1 second and a step size of 0.5 seconds to capture the data segments within the window. .right Perform a 4096-point Fast Fourier Transform to obtain the amplitude spectrum. According to online updates and Spectral lines that satisfy the following conditions are retained: ; in, The peak value of the amplitude spectrum is used. Spectral lines that do not meet the above conditions are set to zero. An inverse fast Fourier transform is performed on the retained spectral lines to obtain a compressed time-domain segment. The length remains 1 s, and the sampling rate remains 10 kS / s.

[0047] Compress the time domain segment The active power of the generating units is divided into three levels based on the percentage of real-time active power: the first steady-state level corresponds to active power between 20% and 40% of the rated power, the second steady-state level corresponds to active power between 40% and 70% of the rated power, and the third steady-state level corresponds to active power between 70% and 100% of the rated power. At least 10 valid windows are accumulated in each level before proceeding to the next step of feature matrix generation.

[0048] For each steady-state compression segment, the per-unit averages of voltage, active power, and reactive power are calculated and arranged in frequency band order to form the harmonic compression data matrix HCMT and the common-mode compression data matrix CCMT. All matrix elements are 32-bit floating-point data, with storage space compressed by at least 85% compared to the original data, and are updated continuously according to operating conditions. HCMT and CCMT serve as the input features for the subsequent cascaded dual-feedback stochastic resonance model.

[0049] Step S4: Construct a cascaded double-feedback stochastic resonance model and perform online optimization.

[0050] The cascaded dual-feedback stochastic resonance model comprises a two-stage cascaded structure consisting of a front-stage subsystem and a rear-stage subsystem. The front-stage subsystem employs a bistable state function, the expression of which is: ; in, Let be the potential well depth parameter. The potential well width parameter, These are the state variables of the preceding subsystem. In this embodiment, The initial value is set to 1.2. The initial value is set to 0.8. The bistable state function is... Two symmetrical potential wells are formed at the location, with a barrier height of . When the input signal and noise work together, the transition probability of particles between the two potential wells increases, producing a stochastic resonance effect and achieving initial signal enhancement.

[0051] The subsequent subsystem uses a Gaussian potential function, the expression of which is: ; in, Let be the potential well depth parameter. The potential well width parameter, These are the state variables of the subsequent subsystem. In this embodiment, The initial value is set to 0.45 squared, which is 0.2025. The initial value is set to 0.45. The Gaussian potential function forms a single potential well at the origin, with a potential well depth of... The effective width of the potential well is determined by Control. The signal output from the front-end subsystem is processed by a Gaussian potential function to achieve secondary energy accumulation, further improving the signal-to-noise ratio.

[0052] The front-end subsystem and the back-end subsystem are connected by unidirectional coupling, and the coupling coefficient is... The initial value is set to 0.32. A first negative feedback coefficient is introduced into the preceding subsystem. The output of the subsequent subsystem is fed back to the input of the preceding subsystem. A second negative feedback coefficient is introduced into the subsequent subsystem. The output of the subsequent subsystem is fed back to the input of the subsequent subsystem itself. The value range is from -0.5 to 0, and the initial value is set to -0.35. The value range is from -0.1 to 0, and the initial value is set to -0.06. First negative feedback coefficient. Its function is to adjust the gain and phase characteristics of the preceding subsystem, affecting the system's stability and response speed. The second negative feedback coefficient... Its function is to fine-tune the dynamic behavior of subsequent subsystems, ensuring that the system can accurately respond to changing input conditions. The dual-loop negative feedback structure enables the cascaded system to better adapt to different operating conditions and environmental changes.

[0053] The online optimization employs an improved particle swarm optimization algorithm. The parameters to be optimized include... , , , , , It has six dimensions. The number of particles is set to 30, and the maximum number of iterations is set to 50. The objective function is defined as follows: ; in, For harmonic signal-to-noise gain, This refers to the common-mode energy ratio. In this embodiment, , The improved particle swarm optimization algorithm performs parameter optimization within each computation cycle, with a single computation time of less than 5 ms, meeting the requirements of 1 kHz real-time closed-loop control. After online optimization, an enhanced harmonic sequence is obtained. With common mode sequence The signal-to-noise ratio input is improved by no less than 20 dB.

[0054] Step S5: Extract the harmonic distortion rate and common-mode energy ratio and perform early fault detection.

[0055] For enhanced harmonic sequences Perform a 4096-point Fast Fourier Transform to obtain the harmonic amplitude spectrum. For enhanced common-mode sequences Perform a 4096-point Fast Fourier Transform to obtain the common-mode amplitude spectrum. Simultaneously, an S-transform is performed on the same data segment to obtain the time-frequency matrix. It is used to observe instantaneous frequency drift.

[0056] The formula for calculating harmonic distortion rate is: ; in, The fundamental amplitude.

[0057] The formula for calculating the common-mode energy ratio is: ; in, This represents the total energy.

[0058] The dynamic threshold is generated using a self-learning method. Historical data on harmonic distortion rate and common-mode energy ratio within a 24-hour sliding window are used to calculate the 80th percentile value, which is then multiplied by three standard deviations to form the dynamic alarm threshold. The sliding window advances every 5 minutes, enabling continuous self-learning and updating of the threshold.

[0059] When harmonic distortion rate Exceeding the dynamic threshold or common-mode energy ratio Exceeding the dynamic threshold If the above-mentioned over-limit state continues for three consecutive calculation cycles, an early fault alarm will be triggered. The alarm information includes the type of over-limit indicator, the over-limit amplitude, the frequency band energy distribution, and the recommended maintenance parts. The rules for determining the recommended maintenance parts are as follows: when the harmonic distortion rate exceeds the limit, it is recommended to check the converter power module and filter inductor; when the common-mode energy ratio exceeds the limit, it is recommended to check the grounding brush and generator winding.

[0060] In the normal operating condition verification of this embodiment, the average wind speed was 6.8 m / s, and the unit output power was 1.5 MW, accounting for 60% of the rated power. The original signal-to-noise ratio was -29 dB in the harmonic band and -32 dB in the common-mode band. After cascaded dual-feedback random resonance enhancement, the harmonic signal-to-noise gain was 14.5 dB, and the common-mode signal-to-noise gain was 13.0 dB. The extracted harmonic distortion rate was 2.0%, and the common-mode energy ratio was 1.3%, both lower than the dynamic threshold. The system determined it to be a normal operating condition and did not trigger an alarm.

[0061] In the fault injection experiment of this embodiment, a 0.18 μF capacitor with a withstand voltage of 630 V was connected in parallel with phase B of the generator rotor winding to simulate early inter-turn insulation degradation; a 47 ohm resistor with a power of 200 W was connected in series between the converter positive bus and ground to simulate uneven wear of the grounding brush. Ten minutes after the fault was injected, the harmonic sideband amplitude increased by only 9 mV, corresponding to approximately -33 dB; the common-mode pulse increased by 16 mA. The traditional 8th-order bandpass filter combined with the fast Fourier transform method measured a harmonic distortion rate of 2.7%, which did not reach the alarm threshold, resulting in a missed alarm. The method of this invention measured a harmonic distortion rate of 6.1% and a common-mode energy ratio of 4.0%, exceeding the dynamic threshold for 3 consecutive seconds, accurately triggering an early fault alarm, and pointing the recommended maintenance location to the generator winding and the converter power module.

[0062] Example 2 This embodiment uses a 3.0 MW direct-drive permanent magnet wind turbine full-power converter as an application scenario. The unit's rated line voltage is 690 V, DC bus voltage is 1100 V, and switching frequency is 3.0 kHz. The unit employs a permanent magnet synchronous generator, and the converter is a full-power back-to-back structure, using silicon carbide metal-oxide-semiconductor field-effect transistors (MOSFETs) as power devices. Due to the faster switching speed of silicon carbide devices, they generate richer high-frequency harmonic components, requiring more stringent extraction of weak features.

[0063] See Figure 1 The weak feature extraction method in this embodiment includes the following steps: Step S1: Synchronously acquire multi-channel electrical signals from the wind turbine and perform preprocessing.

[0064] For voltage acquisition, a 1 kV capacitive voltage divider sensor with an accuracy of 0.05% and a bandwidth of 200 MHz is used on the three-phase AC side; a 10 kV isolated differential probe with a bandwidth of 100 MHz is used on the DC bus. For current acquisition, a Rogowski coil with a frequency response range of 0.1 Hz to 200 MHz is used on the three-phase AC side; a high-frequency current clamp with a frequency response range of 500 kHz to 200 MHz is used for ground common-mode current, with a minimum measurable current of 0.5 mA. The impedance between all sensor housings and the cabin grounding point is less than 1 milliohm.

[0065] The synchronous acquisition hardware uses a 24-bit Delta-Sigma analog-to-digital converter with a maximum single-channel sampling rate of 512 kS / s. In this embodiment, the sampling rate is uniformly downsampled to 20 kS / s, which can accommodate the acquisition requirements of the 100th harmonic component and 100 kHz switching ripple generated by the silicon carbide device. The FPGA is locally configured with a 20 MHz temperature-controlled crystal oscillator with a frequency stability of ±0.02 ppm. Time synchronization is performed using the IEEE 1588 v2 precise time protocol, with a synchronization error of less than 100 ns, ensuring that the phase error between channels within the same fundamental period is less than 0.005 degrees.

[0066] The analog front-end is equipped with a 4th-order Butterworth anti-aliasing filter with a cutoff frequency of 8 kHz and out-of-band attenuation greater than 70 dB. The analog-to-digital converter pre-stage is equipped with a 3.3 V bidirectional transient voltage suppressor array, which can withstand 2 kV combined wave impacts.

[0067] Digital preprocessing employs a 2048-point sliding window for median filtering. When sampling points... With window value The deviation satisfies When the value is in the window, it is replaced. Zero drift correction calculates the DC component of each channel every 5 seconds and subtracts it in real time.

[0068] The standard base value setting method for per-unit transformation is as follows: The voltage base value is taken as the peak value of the rated line voltage of the unit. Right now V; The base current value is taken as the peak value of the unit's rated current. Right now A; The power base value is taken as the rated apparent power of the unit. Right now kVA.

[0069] The amplitude range of the data after per-unit transformation does not exceed ±2 pu, the mean is zero, and the variance is not greater than 1.5.

[0070] Regarding data packetization, one data block is generated per second. Each data block contains single-precision floating-point data from 6 channels multiplied by 20,000 sampling points, an 8-byte GPS timestamp, and a 2-byte quality tag. When the communication link is interrupted, the FPGA board's onboard 16 GB embedded multimedia storage card performs rolling caching, with a caching duration of no less than 48 hours.

[0071] Step S2: Perform bandpass tracking filtering on the standardized signal.

[0072] The standardized signal matrix is ​​simultaneously fed into two sets of digital bandpass filters. The harmonic bandpass filter bank has a passband range of 3 kHz to 100 kHz, a filter order of 10th order, an in-band ripple of no more than 0.15 dB, and an out-of-band attenuation of no less than 80 dB / dec. The common-mode bandpass filter bank has a passband range of 1.5 kHz to 15 kHz, a filter order of 8th order, an in-band ripple of no more than 0.2 dB, and an out-of-band attenuation of no less than 70 dB / dec. Harmonic sub-matrices are obtained after filtering. With common modulus submatrix The timestamp alignment error does not exceed 1 sampling point.

[0073] Step S3: Perform segmented folding time-scale compression on the component signals.

[0074] The factory default setting for the compression ratio threshold is: The update triggering conditions during operation are the same as in Example 1. The update step size is set to 0.001, and the update cycle is 3 minutes by default, which is shortened to 30 seconds when there are gusts of wind or severe power fluctuations.

[0075] harmonic submatrix and common modulus submatrix Parallel segmented compression is performed. A sliding window is used with a window length of 1 s and a step size of 0.5 s for truncation. An 8192-point Fast Fourier Transform is performed on the data segments within the window to obtain higher frequency resolution. (Based on online updates...) and The spectral lines that meet the conditions are retained, while those that do not meet the conditions are set to zero. Then, an inverse fast Fourier transform is performed to obtain a compressed time-domain segment.

[0076] The compressed time-domain segments are divided into three levels based on the percentage of the unit's real-time active power: the first steady-state level corresponds to active power between 15% and 35% of the rated power, the second steady-state level corresponds to active power between 35% and 65% of the rated power, and the third steady-state level corresponds to active power between 65% and 100% of the rated power. At least 15 valid windows are accumulated for each level before proceeding to the feature matrix generation stage.

[0077] Step S4: Construct a cascaded double-feedback stochastic resonance model and perform online optimization.

[0078] The initial values ​​of the bistable situation function parameters of the preceding subsystem are: , The initial values ​​of the Gaussian potential function parameters of the subsequent subsystem are... , Coupling coefficient The initial value is set to 0.28. First negative feedback coefficient. The initial value is set to -0.40, and the second negative feedback coefficient is... The initial value is set to -0.08.

[0079] The online optimization employs a modified particle swarm optimization algorithm, with the particle count set to 40 and the maximum number of iterations set to 60. The weight coefficients of the objective function are set as follows: , The single calculation time is less than 8 ms, meeting the requirements of real-time closed-loop control. After online optimization, the signal-to-noise ratio comparison input of the harmonic sequence and common-mode sequence is improved by no less than 22 dB.

[0080] Step S5: Extract the harmonic distortion rate and common-mode energy ratio and perform early fault detection.

[0081] The enhanced harmonic sequence was subjected to an 8192-point Fast Fourier Transform (FFT) to obtain the harmonic amplitude spectrum, and the enhanced common-mode sequence was subjected to an 8192-point FFT to obtain the common-mode amplitude spectrum. In the calculation of the harmonic distortion rate, the harmonic frequency band ranged from 3 kHz to 100 kHz. In the calculation of the common-mode energy ratio, the common-mode frequency band ranged from 1.5 kHz to 15 kHz.

[0082] The dynamic threshold is calculated using historical data within a 24-hour sliding window, with the 85th percentile value plus three times the standard deviation as the dynamic alarm threshold. The sliding window moves forward every 3 minutes. An early fault alarm is triggered when the harmonic distortion rate or common-mode energy ratio exceeds the dynamic threshold for three consecutive calculation cycles.

[0083] In the normal operating condition verification of this embodiment, the average wind speed was 8.5 m / s, and the unit output power was 2.1 MW, accounting for 70% of the rated power. The original signal-to-noise ratio was -31 dB in the harmonic band and -34 dB in the common-mode band. After cascaded dual-feedback stochastic resonance enhancement, the harmonic signal-to-noise gain was 15.8 dB, and the common-mode signal-to-noise gain was 14.2 dB. The extracted harmonic distortion rate was 1.8%, and the common-mode energy ratio was 1.1%, both lower than the dynamic threshold, and the system was determined to be operating under normal conditions.

[0084] In the fault injection experiment of this embodiment, a 0.5 A DC bias current was superimposed on the filter inductor on the grid side of the converter to simulate the early stage of magnetic saturation of the filter inductor; an additional 10 ohm resistor was connected in series in the grounding circuit to simulate the increase in the contact resistance of the grounding brush. Eight minutes after the fault was injected, the harmonic sideband amplitude increased by 12 mV and the common-mode pulse increased by 22 mA. The harmonic distortion rate measured by the conventional method was 2.4%, which did not reach the alarm threshold. The method of this invention measured a harmonic distortion rate of 5.7% and a common-mode energy ratio of 3.8%, which exceeded the dynamic threshold for 3 consecutive seconds, accurately triggering an early fault alarm. It is recommended to locate the filter inductor and the grounding brush for inspection.

[0085] Example 3 This embodiment uses a 6.0 MW semi-direct drive offshore wind turbine medium-voltage converter as the application scenario. The unit's rated line voltage is 3.3 kV, DC bus voltage is 5.2 kV, and switching frequency is 1.8 kHz. The unit is installed on an offshore platform, operating in an environment with strong electromagnetic interference and marine salt spray corrosion. The environmental noise includes white noise and... Stable impulse noise requires greater robustness for weak feature extraction.

[0086] See Figure 1 The weak feature extraction method in this embodiment includes the following steps: Step S1: Synchronously acquire multi-channel electrical signals from the wind turbine and perform preprocessing.

[0087] For voltage acquisition, a 20 kV capacitive voltage divider sensor with an accuracy of 0.08% and a bandwidth of 150 MHz is used on the three-phase AC side. The sensor housing is made of 316L stainless steel to resist marine salt spray corrosion. An 80 kV isolated differential probe with a bandwidth of 80 MHz is used on the DC bus. For current acquisition, a Rogowski coil with a frequency response range of 0.1 Hz to 150 MHz is used on the three-phase AC side, and the coil housing is IP67 protected. A high-frequency current clamp with a frequency response range of 800 kHz to 150 MHz is used for ground common-mode current acquisition, with a minimum measurable current of 0.8 mA. The impedance between all sensor housings and the cabin grounding point is less than 1.5 milliohms.

[0088] The synchronous acquisition hardware uses a 24-bit Delta-Sigma analog-to-digital converter with a maximum single-channel sampling rate of 256 kS / s. In this embodiment, the sampling rate is uniformly downsampled to 10 kS / s. The FPGA is locally configured with a 10 MHz temperature-controlled crystal oscillator with a frequency stability of ±0.03ppm. Time synchronization is performed using the IEEE 1588 v2 precise time protocol, with a synchronization error of less than 150 ns, ensuring that the phase error between channels within the same fundamental period is less than 0.008 degrees.

[0089] The analog front-end is equipped with a third-order Butterworth anti-aliasing filter with a cutoff frequency of 4 kHz and out-of-band attenuation greater than 55 dB. The analog-to-digital converter pre-stage is equipped with a 5 V bidirectional transient voltage suppressor array, which can withstand 6 kV combined wave impacts to adapt to the harsher lightning environment at sea.

[0090] Digital preprocessing employs a 1024-point sliding window for median filtering. When the deviation between the sampling point and the median in the window meets... Replacement is performed periodically. Zero-drift correction calculates and subtracts the DC component of each channel every 8 seconds. The rated base value for per-unit conversion is set as: voltage base value. V, base current value A, power base value kVA.

[0091] For data packetization, one data block is generated per second, containing single-precision floating-point data from 6 channels multiplied by 10,000 sampling points. Communication employs dual-channel redundant transmission via satellite link and fiber optic submarine cable. When all communication links are disconnected, the FPGA board's onboard 32GB embedded multimedia storage card performs rolling caching with a caching duration of no less than 120 hours to accommodate the long maintenance cycles at sea.

[0092] Step S2: Perform bandpass tracking filtering on the standardized signal.

[0093] The harmonic bandpass filter bank has a passband range of 1.8 kHz to 50 kHz, an 8th-order filter, in-band ripple no greater than 0.2 dB, and out-of-band attenuation no less than 60 dB / dec. The common-mode bandpass filter bank has a passband range of 800 Hz to 10 kHz, a 6th-order filter, in-band ripple no greater than 0.3 dB, and out-of-band attenuation no less than 60 dB / dec. Harmonic sub-matrices are obtained after filtering. With common modulus submatrix The timestamp alignment error does not exceed 1 sampling point.

[0094] Step S3: Perform segmented folding time-scale compression on the component signals.

[0095] The factory default setting for the compression ratio threshold is: and The update step size is set to 0.001, and the update cycle is 5 minutes by default. Due to the more drastic changes in wind speed at sea, the threshold for the 10-minute average wind speed change in the update trigger condition has been adjusted to 0.8 m / s.

[0096] Piecewise folding compression is performed in parallel on the harmonic submatrix and the common mode submatrix. A sliding window is used with a window length of 1 s and a step size of 0.5 s for truncation. A 4096-point Fast Fourier Transform is performed on the data segment within the window. Based on... and After retaining the spectral lines that meet the conditions, an inverse fast Fourier transform is performed to obtain a compressed time-domain segment.

[0097] The compressed time-domain segments are divided into three levels based on the percentage of active power: the first steady-state level corresponds to active power between 25% and 45% of the rated power, the second steady-state level corresponds to active power between 45% and 75% of the rated power, and the third steady-state level corresponds to active power between 75% and 100% of the rated power. At least 12 valid windows are accumulated for each level before proceeding to the feature matrix generation stage.

[0098] Step S4: Construct a cascaded double-feedback stochastic resonance model and perform online optimization.

[0099] The initial values ​​of the bistable situation function parameters of the preceding subsystem are: , The initial values ​​of the Gaussian potential function parameters of the subsequent subsystem are... , Coupling coefficient The initial value is set to 0.35. First negative feedback coefficient. The initial value is set to -0.30, and the second negative feedback coefficient is... The initial value is set to -0.05.

[0100] The online optimization employs a modified particle swarm optimization algorithm, with the particle count set to 35 and the maximum number of iterations set to 50. The weight coefficients of the objective function are set as follows: , Due to the more complex harmonic characteristics of offshore wind turbines, the weight of harmonic signal-to-noise ratio gain is appropriately increased. The single calculation time is less than 5 ms, meeting the requirements of real-time closed-loop control. After online optimization, the signal-to-noise ratio comparison input of the harmonic sequence and common-mode sequence is improved by no less than 18 dB.

[0101] Step S5: Extract the harmonic distortion rate and common-mode energy ratio and perform early fault detection.

[0102] The enhanced harmonic sequence was subjected to a 4096-point Fast Fourier Transform (FFT) to obtain the harmonic amplitude spectrum, and the enhanced common-mode sequence was subjected to a 4096-point FFT to obtain the common-mode amplitude spectrum. In the calculation of the harmonic distortion rate, the harmonic frequency band ranged from 1.8 kHz to 50 kHz. In the calculation of the common-mode energy ratio, the common-mode frequency band ranged from 800 Hz to 10 kHz.

[0103] The dynamic threshold is calculated using historical data within a 36-hour sliding window, with the 80th percentile value plus 3.5 times the standard deviation as the dynamic alarm threshold. The sliding window is extended to 36 hours to accommodate greater variations in sea conditions and to increase the historical sample size, thus improving threshold stability. The sliding window advances every 5 minutes. An early fault alarm is triggered when the harmonic distortion rate or common-mode energy ratio exceeds the dynamic threshold and remains so for four consecutive calculation cycles. Increasing the number of consecutive cycles to four is to reduce the false alarm rate caused by electromagnetic interference in the marine environment.

[0104] In the normal operating condition verification of this embodiment, the average wind speed was 10.2 m / s, and the unit output power was 4.2 MW, accounting for 70% of the rated power. The original signal-to-noise ratio was -30 dB in the harmonic band and -33 dB in the common-mode band. After cascaded dual-feedback stochastic resonance enhancement, the harmonic signal-to-noise gain was 13.8 dB, and the common-mode signal-to-noise gain was 12.5 dB. The extracted harmonic distortion rate was 2.2%, and the common-mode energy ratio was 1.5%, both lower than the dynamic threshold, and the system was determined to be operating under normal conditions.

[0105] In the fault injection experiment of this embodiment, a weak leakage current of 5 mA was injected between the positive terminal of the converter DC bus and the heat sink to simulate the insulation aging of power devices. A 50 milliohm additional resistor was injected relative to ground in the generator stator winding A to simulate the initial stage of insulation degradation. Fifteen minutes after the fault injection, the harmonic sideband amplitude increased by 7 mV, and the common-mode pulse increased by 12 mA. The traditional method measured a harmonic distortion rate of 2.1%, which did not reach the alarm threshold. The method of this invention measured a harmonic distortion rate of 5.3% and a common-mode energy ratio of 3.5%, exceeding the dynamic threshold for four consecutive calculation cycles, accurately triggering an early fault alarm. It is recommended that the inspection site be located in the converter power module and the generator winding.

[0106] In the extreme signal-to-noise ratio test of this embodiment, a signal source was used in the laboratory with white noise superimposed on it. Stabilizing the impulse noise to reconstruct a -30 dB signal, the Monte Carlo experiment was repeated 1000 times. Test results show that the harmonic amplitude is increased by 12 to 15 dB, with an average increase of 13.3 dB; the common-mode impulse detection rate is 96%, while the detection rate of traditional methods is only 62%; the parameter optimization convergence number does not exceed 40 times, and the single operation time is 4.0 ms, meeting the 1 kHz real-time closed-loop requirement.

[0107] In summary, the embodiments of the present invention have at least the following technical effects: This invention constructs a cascaded double-feedback stochastic resonance model, which cascades and couples the pre-stage bistable potential function with the post-stage Gaussian potential function. Combined with a double-loop negative feedback structure, it achieves a secondary transfer of noise energy to the target signal frequency band, effectively solving the technical problem of suppressing weak harmonics and common-mode characteristics simultaneously in strong noise environments using traditional linear filtering methods.

[0108] This invention employs a segmented folding time-scale compression technique to uniformly map high-frequency harmonics and common-mode pulses to the effective frequency band of random resonance. This breaks through the technical bottleneck of poor adaptability of single-level random resonance models to high-frequency signals and severe parameter coupling, and is applicable to wind turbine units with different rated capacities and different switching frequencies.

[0109] This invention improves the particle swarm optimization algorithm to optimize six-dimensional parameters in real time. Combined with a dynamic threshold self-learning mechanism, it can still increase the harmonic amplitude by 12 to 15 dB under the extreme signal-to-noise ratio of -30 dB and effectively improve the common-mode pulse detection rate. It provides a reliable online early warning method for early defects such as aging of power devices, saturation of filter inductors, and wear of grounding brushes in wind power converters.

[0110] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for extracting weak features of wind power based on cascaded dual-feedback stochastic resonance, characterized in that, include: The multi-channel electrical signals of the wind turbine are collected and preprocessed to obtain standardized signals. The multi-channel electrical signals include three-phase voltage, three-phase current, DC bus voltage and grounding common-mode current. The standardized signal is subjected to bandpass tracking filtering to extract the harmonic component signal of the harmonic frequency band and the common-mode component signal of the common-mode pulse frequency band, respectively. The harmonic component signal and the common-mode component signal are subjected to piecewise folding time-scale compression, and the high-frequency components are mapped to the effective frequency band of random resonance to obtain the compressed signal. A cascaded double-feedback stochastic resonance model is constructed. The front-stage subsystem of the cascaded double-feedback stochastic resonance model adopts a bistable potential function, and the rear-stage subsystem adopts a Gaussian potential function. The front-stage subsystem and the rear-stage subsystem are cascaded and coupled. A first negative feedback coefficient is introduced into the front-stage subsystem and a second negative feedback coefficient is introduced into the rear-stage subsystem to form a double-loop negative feedback structure. The parameters of the first negative feedback coefficient, the second negative feedback coefficient, the bistable potential function, and the Gaussian potential function are optimized in real time to achieve a secondary transfer of noise energy to the target frequency band, thereby obtaining an enhanced cascaded output signal. The cascaded output signals are subjected to parallel Fast Fourier Transform and S-Transform to extract the harmonic distortion rate and common-mode energy ratio. When the harmonic distortion rate or the common-mode energy ratio exceeds the dynamic threshold, an early fault alarm is triggered.

2. The wind power weak feature extraction method based on cascaded dual-feedback stochastic resonance according to claim 1, characterized in that, The preprocessing includes: The multi-channel electrical signal is subjected to sliding window median filtering. When the deviation between the sampling point and the median in the window exceeds the product of a preset deviation multiple and the standard deviation, the sampling point is replaced by the median in the window. The DC component in the multi-channel electrical signal is calculated and subtracted in real time according to a preset time period, so that the mean of the multi-channel electrical signal is zero. Each electrical quantity is divided by its corresponding rated base value and then converted to per-unit value to obtain the standardized signal whose amplitude range does not exceed the preset per-unit upper limit.

3. The wind power weak feature extraction method based on cascaded dual-feedback stochastic resonance according to claim 1, characterized in that, The bandpass tracking filter includes: The standardized signal is simultaneously fed into a harmonic bandpass filter bank and a common-mode bandpass filter bank. The passband range of the harmonic bandpass filter bank is from a preset lower harmonic frequency to a preset upper harmonic frequency, and the passband range of the common-mode bandpass filter bank is from a preset lower common-mode frequency to a preset upper common-mode frequency. The harmonic bandpass filter bank outputs the harmonic component signal, the common-mode bandpass filter bank outputs the common-mode component signal, and the timestamp alignment error between the harmonic component signal and the common-mode component signal does not exceed one sampling point.

4. The method for extracting weak features of wind power based on cascaded dual-feedback stochastic resonance according to claim 1, characterized in that, The segmented folding timescale compression includes: The harmonic component signal and the common-mode component signal are respectively truncated by sliding window with a preset window length and a preset step size to obtain the data segment within the window; The amplitude spectrum is obtained by performing a fast Fourier transform on the data segment within the window; Based on the online-set low-frequency compression ratio threshold and high-frequency compression ratio threshold, retain the spectral lines in the amplitude spectrum that satisfy the following conditions: ； in, The low-frequency compression ratio threshold is... The high-frequency compression ratio threshold is... For frequency The amplitude at that point, The peak value of the amplitude spectrum is set to zero for spectral lines that do not meet the conditions. The compressed signal is obtained by performing an inverse fast Fourier transform on the preserved spectral lines.

5. The wind power weak feature extraction method based on cascaded dual-feedback stochastic resonance according to claim 4, characterized in that, The online update method for the low-frequency compression ratio threshold and the high-frequency compression ratio threshold is as follows: When the average wind speed change within a preset time period reaches a preset wind speed threshold, or the percentage change in active power reaches a preset power threshold, or the change in converter casing temperature reaches a preset temperature threshold, the threshold update is initiated. The low-frequency compression ratio threshold and the high-frequency compression ratio threshold are adjusted by increasing or decreasing according to a preset adjustment step size. In the high-power region, the low-frequency compression ratio threshold is decreased and the high-frequency compression ratio threshold is increased, and the opposite adjustment is made in the low-power region. After each adjustment, ensure that the low-frequency compression ratio threshold and the high-frequency compression ratio threshold both fall within their respective preset value ranges, and that the difference between the two is not less than the preset minimum difference. If a preset number of consecutive updates causes the harmonic distortion rate or the common-mode energy ratio to jump beyond a preset jump threshold, it will automatically revert to the previous stable value.

6. The method for extracting weak features of wind power based on cascaded dual-feedback stochastic resonance according to claim 1, characterized in that, The expression for the bistable state function is: ； in, Let be the potential well depth parameter of the bistable state function. Let be the potential well width parameter of the bistable state function. These are the state variables of the preceding subsystem; The expression for the Gaussian potential function is: ； in, Let be the potential well depth parameter of the Gaussian potential function. Let be the potential well width parameter of the Gaussian potential function. These are the state variables of the subsequent subsystem; First negative feedback coefficient The output of the subsequent subsystem is fed back to the input of the preceding subsystem, and the second negative feedback coefficient... The output of the subsequent subsystem is fed back to the input of the subsequent subsystem itself. The value range of the first negative feedback coefficient is from a preset first feedback lower limit to zero, and the value range of the second negative feedback coefficient is from a preset second feedback lower limit to zero.

7. The method for extracting weak features of wind power based on cascaded dual-feedback stochastic resonance according to claim 1, characterized in that, The parameters of the first negative feedback coefficient, the second negative feedback coefficient, the bistable potential function, and the Gaussian potential function are optimized using an optimization algorithm to achieve real-time optimization. The online optimization algorithm is executed as follows: An improved particle swarm optimization algorithm is used to simultaneously optimize six parameters: the first negative feedback coefficient, the second negative feedback coefficient, the potential well depth parameter and the potential well width parameter of the bistable potential function, and the potential well depth parameter and the potential well width parameter of the Gaussian potential function. The objective function of the improved particle swarm optimization algorithm is: ； in, For harmonic signal-to-noise gain, Common-mode energy ratio, and For preset weighting coefficients and ; The improved particle swarm algorithm sets the number of particles to a preset number and the maximum number of iterations to a preset iteration limit.

8. The method for extracting weak features of wind power based on cascaded dual-feedback stochastic resonance according to claim 1, characterized in that, The harmonic distortion rate is calculated as follows: ； in, Frequency within the harmonic band The amplitude at that point, and These are the lower and upper limits of the harmonic frequency band, respectively. The fundamental amplitude; The common-mode energy ratio is calculated as follows: ； in, Frequency within the common-mode band The amplitude at that point, and These are the lower and upper limits of the common-mode frequency band, respectively. This represents the total energy.

9. The method for extracting weak features of wind power based on cascaded dual-feedback stochastic resonance according to claim 1, characterized in that, The dynamic threshold is generated as follows: Take historical data of the harmonic distortion rate and the common-mode energy ratio within a preset sliding window period; Calculate the preset quantile value and standard deviation of the historical data, and multiply the preset quantile value and preset standard deviation by the sum of the standard deviations to obtain the dynamic threshold. The sliding window slides forward once according to a preset sliding period to achieve self-learning update of the dynamic threshold.

10. The method for extracting weak features of wind power based on cascaded dual-feedback stochastic resonance according to any one of claims 1 to 9, characterized in that, The triggering conditions for the early fault alarm are as follows: The harmonic distortion rate exceeds the corresponding dynamic threshold, or the common-mode energy ratio exceeds the corresponding dynamic threshold; Furthermore, after the aforementioned over-limit state continues for several preset consecutive calculation cycles, the early fault alarm is triggered; The information in the early fault alarm includes the type of out-of-limit indicator, the out-of-limit amplitude, and the suggested repair location.

Images