Voiceprint noise reduction method based on multiphase filtering and complex variational mode decomposition

Abstract

The invention discloses a voiceprint noise reduction method based on multiphase filtering and complex variational mode decomposition, and the method comprises the steps: carrying out the multiphase filtering of a collected voiceprint signal, and dividing the voiceprint signal into a plurality of sub-band signals; calculating the average power of the sub-band signals, comparing the average power with a dynamic threshold value, and screening out the sub-band signals of which the average power is greater than or equal to the dynamic threshold value; and performing complex variational mode decomposition on the screened sub-band signals of which the average power is greater than or equal to a dynamic threshold value to obtain a mode component after noise reduction. According to the wind turbine generator voiceprint signal acquired in a complex environment, through cooperative multiphase filtering, dynamic threshold value-based sub-band signal screening and complex variational mode decomposition, the sub-band signal-to-noise ratio can be effectively improved, weak fault features such as blade cracks can be reserved, and the accuracy of the wind turbine generator voiceprint signal is improved. The accuracy of wind turbine generator fault detection according to voiceprint signal features can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of wind turbine condition monitoring and fault diagnosis technology, and in particular to a noise reduction method for acoustic signature signals of wind turbines. Background Technology

[0002] Wind turbines are a crucial component of modern power systems, and their reliable operation is essential for wind power grid connection and grid stability. Fault diagnosis of wind turbines is a key aspect of ensuring their long-term stable operation. In recent years, acoustic signals have been widely used for the early diagnosis of wind turbine faults because sound waves can reflect the operating status and potential faults of wind turbines. However, complex environmental noises (such as wind noise, mechanical vibration, and birdsong) can interfere with effective acoustic feature extraction, thus affecting the accuracy of fault diagnosis.

[0003] Existing noise suppression techniques, such as recursive averaging filtering, wavelet transform, and empirical mode decomposition (EMD), can reduce noise interference to some extent, but they still have many problems. For example, traditional filtering methods are not effective at suppressing impulse interference, EMD may cause mode aliasing, variable mode decomposition (VMD) is difficult to process complex signals, and while neural network algorithms can provide relatively accurate diagnostic results, they usually require a large amount of sample data and computational resources. Especially in multi-source environments, existing noise reduction algorithms struggle to effectively separate different sound sources and perform accurate noise reduction, thus limiting the accuracy and real-time performance of fault diagnosis. Summary of the Invention

[0004] In view of this, the present invention provides a method for acoustic text noise reduction based on multiphase filtering and complex variational mode decomposition to solve the technical problems of insufficient performance and high computational complexity of existing acoustic noise reduction technologies in complex environments.

[0005] The present invention, based on multiphase filtering and complex variational mode decomposition, includes the following measures for acoustic text noise reduction: The system collects acoustic fingerprint signals from wind turbines during operation, performs multiphase filtering on the collected acoustic fingerprint signals, and divides the acoustic fingerprint signals into several sub-band signals. The average power of the sub-band signal is calculated, and compared with a dynamic threshold. Sub-band signals with an average power greater than or equal to the dynamic threshold are selected. The average power of the sub-band signal is calculated using the following formula:

[0006] in, This represents the average power of the k-th sub-band signal. This represents the index of the signal sample in the sub-band signal, and N is the total length of the sequence of the k-th sub-band signal. This represents the signal value of the nth signal sample within the kth sub-band signal. The square of the modulus of the signal value represents the instantaneous energy of the signal. The dynamic threshold is obtained by the following method: collecting ambient wind speed, determining the dynamic adjustment coefficient β based on the ambient wind speed, and then calculating the dynamic threshold according to the following formula:

[0007] In the formula, This represents the dynamic threshold set for the k-th sub-band signal; This represents the average power of the noise in the k-th subband. The power standard deviation of the k-th subband noise is represented by the subband noise obtained by polyphase filtering of a pure noise signal without a target sound source. Complex variational mode decomposition is performed on the selected subband signals whose average power is greater than or equal to the dynamic threshold to obtain the denoised mode components.

[0008] Furthermore, the range of the dynamic adjustment coefficient β determined based on the ambient wind speed is 1.5-2.0.

[0009] Furthermore, determining the dynamic adjustment coefficient β based on the ambient wind speed includes: When the ambient wind speed is 3~10m / s, β is taken as 1.5~1.6; when the ambient wind speed is 10~18m / s, β is taken as 1.7~1.8; when the ambient wind speed is 18~25m / s, β is taken as 1.9~2.0.

[0010] Further measures The polyphase filtering of the acquired voiceprint signal described in the document includes: setting the number K of sub-band signals, and grouping the coefficients of the prototype low-pass filter to obtain the sub-filter coefficients. ,

[0011] In the formula, This represents the coefficient sequence of the prototype low-pass filter. This represents the coefficient index of the sub-filter. This represents the index of the sub-filter; Voiceprint signal Segmented signals are obtained by segmentation. :

[0012] In the formula: m represents the time-domain sampling point index of the sub-band signal. Then, the sub-band signal is obtained by processing the segmented signal as follows. ,

[0013] In the formula, j is the imaginary unit, j 2 =-1, Index of the signal in the indices The length of a single sub-filter is represented by , and N represents the total order of the prototype low-pass filter.

[0014] Further measures The calculation of the average power of the sub-band signal described in the article is as follows: only the average power of the sub-band signal from the 0th to the K / 2th sub-band is calculated, and the average power of the remaining sub-band signals is derived based on the signal symmetry.

[0015] Further measures The complex variational mode decomposition described in the document includes: Constructing a constrained variational model:

[0016] In the formula: It is the time-domain representation of the subband signal, where ∂ t Represents the time derivative. As a frequency shift operator, the modal spectrum is shifted to the baseband for bandwidth calculation; Represents modal components, It is an amplitude function. For phase function, The center frequency; By introducing the Lagrange multiplier λ(t) and the penalty factor α, the enhanced Lagrange function is constructed as follows:

[0017] By enhancing the Lagrangian function, the constrained problem is transformed into an unconstrained optimization problem. The modal components are solved using the following steps. : (1) Initialization: Set the number of modes I, the penalty factor α, and the convergence tolerance ε, and randomly initialize the modes u. i (0) Center frequency w i (0) and multiplier λ (0) ; (2) Modal update: Performed in the frequency domain, with the center frequency and Lagrange multipliers fixed, the spectrum of each mode is updated; the modal update formula for the (n+1)th iteration is:

[0018] in, , and They are respectively , Fourier transform of λ(t); (3) Center frequency update: With the modes and Lagrange multipliers fixed, update the center frequency of each mode to minimize the bandwidth, using the following formula:

[0019] (4) Multiplier update: Update the Lagrange multipliers to strengthen the constraints:

[0020] Where τ is the step size parameter; (5) Convergence criterion: Calculate the relative error. If the following conditions are met:

[0021] Then stop the iteration; otherwise return to step (2). (6) The frequency domain modal components obtained after the iteration are processed. The signal is converted back to the time domain by inverse Fourier transform, and the denoised modal components are obtained.

[0022] Furthermore, the aforementioned speaker noise reduction method based on multiphase filtering and complex variational mode decomposition also includes measures ): also includes measures ): Measures The obtained modal components are input into a polyphase synthesis filter bank for signal synthesis, including: a) Perform synthesizer filtering on each modal component separately:

[0023] In the formula, The output of the k-th mode component after being filtered by the l-th synthesizer filter. It is the discrete output signal of the k-th subband signal after being denoised by complex variational mode decomposition; It is the l-th synthesizer sub-filter in the polyphase synthesizer filter bank. This refers to the p-th coefficient of the l-th synthesized sub-filter; b) Perform interpolation on each of the filtered signals to restore the sampling rate of each signal to the original full-bandwidth signal sampling rate:

[0024] In the formula, n is the time index of the original sub-band signal sampling rate, n=K·m + r, r=0,1,...,K-1;

[0025] c) Perform delay shifting and superposition on each of the interpolated signals to achieve image subspectral suppression:

[0026]

[0027] In the formula, This is the final high-sampling-rate filtered output of the k-th signal; The value of the interpolated signal of the k-th channel and the l-th multiphase component is delayed by l full-bandwidth sampling points;

[0028] d) Combine and reconstruct the signals from each path into a full-bandwidth signal. Post-output:

[0029] .

[0030] The beneficial effects of this invention are:

[0031] 1. High noise reduction performance: For wind turbine acoustic signals collected in complex environments, this invention improves the sub-band signal-to-noise ratio through collaborative multiphase filtering, sub-band signal screening based on dynamic thresholds, and complex variational mode decomposition, effectively preserving weak fault features such as blade cracks, and can effectively improve the accuracy of subsequent wind turbine fault detection based on acoustic signal features.

[0032] 2. Multi-source separation capability: The method of this invention performs multi-phase filtering on the acoustic signal, which can accurately isolate the fan sound source (such as blade sound, mechanical sound); and selects sub-band signals according to the dynamic threshold that changes with wind speed, which can effectively reduce environmental interference.

[0033] 3. Low computational complexity: The method of this invention avoids redundant calculations by reusing multiphase structure resources and complex variational mode decomposition, which effectively improves the computational efficiency compared with traditional methods and meets the low latency requirements of wind turbine units.

[0034] 4. Strong robustness and adaptability: Complex variational mode decomposition is robust to non-stationary signals, dynamic parameter adjustment adapts to the operation of the fan under different operating conditions, and threshold detection copes with noise fluctuations, enhancing the stability of the method in harsh environments.

[0035] 5. Practicality and Integrability: The noise-reduced signal can be directly integrated with existing diagnostic systems without hardware modifications, reducing operation and maintenance costs. This method does not require extensive sample training and is suitable for large-scale deployment in wind farms. Attached Figure Description

[0036] Figure 1 This is a flowchart of the voiceprint noise reduction method. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0038] As shown in the figure, the acoustic text denoising method based on multiphase filtering and complex variational mode decomposition in this embodiment includes the following measures:

[0039] The system collects acoustic fingerprint signals from wind turbines during operation, performs multiphase filtering on the collected acoustic fingerprint signals, and divides the acoustic fingerprint signals into several sub-band signals.

[0040] Acoustic signals from wind turbines can be collected using high-precision acoustic sensors (which can be deployed at the blade roots or inside the nacelle, etc.). The sampling rate can be set to 8-16kHz according to the wind turbine's operating frequency to cover the dominant blade vibration frequency (50-2000 Hz). The collected signals contain multiple sound sources and noise, and this step involves multi-phase filtering for frequency division processing.

[0041] The polyphase filtering of the acquired voiceprint signal includes: setting the number K of sub-band signals, and grouping the coefficients of the prototype low-pass filter (in this embodiment, the prototype low-pass filter is an FIR filter) to obtain the sub-filter coefficients. ,

[0042]

[0043] In the formula, The coefficient sequence of the prototype low-pass filter (i.e., the impulse response sequence of the prototype low-pass filter) is represented, and the coefficients of the sub-filters are obtained by grouping them together. The coefficient index represents the coefficient sequence number within each sub-filter (used to identify the l-th coefficient in a given sub-filter). To represent the index of the sub-filter, r takes values ​​ranging from 0, 1, ..., K-1, where K is the number of sub-band signals.

[0044] Voiceprint signal Segmented signals are obtained by segmentation. :

[0045]

[0046] In the formula: m represents the time-domain sampling point index of the sub-band signal (also known as the signal block index after segmentation), corresponding to the time sequence number of each sub-band signal. Then, the sub-band signal is obtained by processing the segmented signal as follows. ,

[0047]

[0048] In the formula, j is the imaginary unit, j 2 =-1, The index of the sub-band signal (k ranges from 0, 1, ..., K-1). The length of a single sub-filter (i.e., the number of coefficients) is represented by N, which represents the total order of the prototype low-pass filter, i.e., the total number of coefficients in the prototype low-pass filter.

[0049] This embodiment targets wind turbine acoustic signatures, setting K=64 (sub-band bandwidth approximately 125 Hz) to ensure that blade fault characteristics (such as 100-300 Hz anomalies caused by cracks) fall within an independent sub-band. Multiphase filtering improves hardware efficiency through resource reuse (such as sharing a low-pass filter), making it suitable for embedded wind turbine systems.

[0050] The average power of the sub-band signal is calculated, and compared with a dynamic threshold. Sub-band signals with an average power greater than or equal to the dynamic threshold are selected. The average power of the sub-band signal is calculated using the following formula:

[0051]

[0052] in, This represents the average power of the k-th sub-band signal. The sequence number of the signal sample in the sub-band signal is represented by N; N is the total sequence length of the k-th sub-band signal. The total sequence length N is set reasonably according to the sampling rate (e.g., N=256 when the sampling rate is 8 kHz) to balance the requirements of calculation accuracy and real-time performance. This represents the signal value of the nth signal sample within the kth sub-band signal. The square of the modulus of the signal value represents the instantaneous energy of the signal.

[0053] Threshold setting is the core of sound source detection, and its purpose is to distinguish the target sound source from background noise. In this embodiment, the dynamic threshold is obtained through the following method: ambient wind speed is collected, a dynamic adjustment coefficient β is determined based on the ambient wind speed, and then the dynamic threshold is calculated using the following formula:

[0054]

[0055] In the formula, This represents the dynamic threshold set for the k-th sub-band signal; This represents the average power of the noise in the k-th subband. The power standard deviation of the k-th subband noise is represented by the subband noise obtained by polyphase filtering of a pure noise signal without a target sound source.

[0056] The determination of the dynamic adjustment coefficient β based on the ambient wind speed includes:

[0057] When the ambient wind speed is 3~10m / s, β is taken as 1.5~1.6; when the ambient wind speed is 10~18m / s, β is taken as 1.7~1.8; when the ambient wind speed is 18~25m / s, β is taken as 1.9~2.0.

[0058] Power calculations are performed in the time domain. The subband signal has been downsampled after polyphase filtering, with a sampling rate of 1 / K of the original signal (where K is the number of subbands). For example, when the original sampling rate fs = 8 kHz and K = 64, the subband sampling rate is reduced to 125 Hz. Power calculations must consider the length of the downsampled data to avoid statistical errors caused by insufficient data volume.

[0059] Subband signals with an average power greater than or equal to the dynamic threshold contain target sound source signals and require subsequent complex variational mode decomposition denoising processing. Subband signals with an average power less than the dynamic threshold do not contain target sound source signals, and their signal data is reset to zero. This step identifies and selects subbands containing target sound sources from the polyphase filtered subbands, avoiding invalid processing of irrelevant noise subbands, thereby improving method efficiency.

[0060] The selected sub-band signals with average power greater than or equal to the dynamic threshold are subjected to complex variational mode decomposition to obtain the denoised mode components. The complex variational mode decomposition includes:

[0061] Constructing a constrained variational model:

[0062]

[0063] In the formula: It is the time-domain representation of the sub-band signal ( It is a discrete sampling form of the sub-band signal. and Essentially, they are different representations of the same sub-band signal, where ∂ t Represents the time derivative. As a frequency shift operator, the modal spectrum is shifted to the baseband for bandwidth calculation; Represents modal components, It is an amplitude function. For phase function, The center frequency;

[0064] By introducing the Lagrange multiplier λ(t) and the penalty factor α, we construct the enhanced Lagrange function:

[0065]

[0066] By enhancing the Lagrangian function, the constrained problem is transformed into an unconstrained optimization problem.

[0067] The modal components are solved using the following steps. :

[0068] (1) Initialization: Set the number of modes I, the penalty factor α, and the convergence tolerance ε = 10−6 And randomly initialize mode u i (0) Center frequency w i (0) and multiplier λ (0) .

[0069] For wind turbine acoustic signatures (sampling rate 8-16 kHz), I is typically set to 4-6; in this embodiment, I is specifically set to 4. Setting I too small can lead to modal mixing (e.g., blade sound and noise are not separated), while setting it too large can introduce spurious modes. Optimization can be achieved by observing the center frequency distribution: if adjacent mode frequencies are too close, I needs to be reduced.

[0070] Penalty factor α: controls modal bandwidth and affects noise reduction intensity. For wind turbines, α is typically set to 1000-2000 to balance noise suppression and feature preservation; high α values ​​enhance noise reduction but may smooth weak fault features, while low α values ​​preserve details but leave noise residue. In this embodiment, α is specifically set to 1000.

[0071] (2) Modal update: Performed in the frequency domain, with the center frequency and Lagrange multipliers fixed, the spectrum of each mode is updated; the modal update formula for the (n+1)th iteration is:

[0072]

[0073] in, , and They are respectively , The Fourier transform of λ(t) is performed. This step highlights the target frequency components and suppresses noise.

[0074] (3) Center frequency update: With the modes and Lagrange multipliers fixed, update the center frequency of each mode to minimize the bandwidth, using the following formula:

[0075]

[0076] (4) Multiplier update: Update the Lagrange multipliers to strengthen the constraints:

[0077]

[0078] Wherein, τ is the step size parameter, and in this embodiment, τ=1 is set.

[0079] (5) Convergence criterion: Calculate the relative error. If the following conditions are met:

[0080]

[0081] If the iteration stops, then return to step (2); otherwise, return to step (2).

[0082] (6) The frequency domain modal components obtained after the iteration are processed. The signal is converted back to the time domain by inverse Fourier transform, and the denoised modal components are obtained.

[0083] The sub-band signal (a complex signal) output by multiphase filtering through complex variational mode decomposition has a specific center frequency and finite bandwidth for each mode component. Complex variational mode decomposition directly optimizes the complex signal, avoiding computational redundancy in the real and imaginary solutions, and maintaining the integrity of phase information. This is crucial for the transient characteristics of wind turbine acoustic signatures (such as blade crack pulses).

[0084] As an improvement to the above embodiments, measures The calculation of the average power of the sub-band signal described in the article is as follows: only the average power of the sub-band signal from the 0th to the K / 2th sub-band is calculated, and the average power of the remaining sub-band signals is derived based on the signal symmetry.

[0085] The polyphase filter output has spectral redundancy, i.e., subband 1 to... With child belt Up to K-1 mirror symmetry (e.g., when K=64, subbands 1-31 and 33-63 are symmetrical), only non-redundant subbands (subbands 0 to K-1) are processed during detection. Including sub-band 0 and (where K is a real signal) to reduce computational load; the remaining subbands can be derived through symmetry. For example, in a 32-subband system (K=32), only subbands 0-15 need to be processed, and the remaining subbands can be derived through symmetry.

[0086] As an improvement to the above embodiments, the acoustic text denoising method based on multiphase filtering and complex variational mode decomposition further includes measures. ): Measures The obtained modal components are input into a polyphase synthesis filter bank for signal synthesis, including:

[0087] a) Perform synthesizer filtering on each modal component separately:

[0088]

[0089] In the formula, The output of the k-th mode component after being filtered by the l-th synthesizer filter. It is the discrete output signal of the k-th subband signal after being denoised by complex variational mode decomposition; It is the l-th synthesizer sub-filter in the polyphase synthesizer filter bank. Let be the p-th coefficient of the l-th synthesized sub-filter. The coefficients of the synthesized sub-filter are directly taken from the prototype filter to ensure the consistency of the filtering characteristics.

[0090] For each denoised modal component Without waiting for interpolation to increase the sampling rate, it directly performs convolution operations with the corresponding K sub-filters. Each sub-band only needs to be convolved with a sub-filter of length L (e.g., when K=64 and L=8, the sub-filter length is only 8, much shorter than the prototype filter's 512), completing the core filtering work at a low sampling rate and significantly reducing the computational load. Complex variational mode decomposition has already removed noise components from the sub-bands, and this step further optimizes the signal's spectral characteristics, laying the foundation for subsequent interpolation and reconstruction.

[0091] b) Perform interpolation on each of the filtered signals to restore the sampling rate of each signal to the original full-bandwidth signal sampling rate:

[0092]

[0093] In the formula, n is the time index of the original subband signal sampling rate, n=K·m + r, r=0,1,...,K-1.

[0094] However, interpolation will cause the spectrum to... The position has K mirror subspectrals, so they are subsequently eliminated by shifting and superimposing.

[0095] c) Perform delay shifting and superposition on each of the interpolated signals to achieve image subspectral suppression:

[0096]

[0097] In the formula, The final high-sampling-rate filtered output of the k-th signal has completely suppressed the mirror sub-spectrum and preserved the narrowband acoustic characteristics corresponding to the sub-band. This is the interpolated signal of the k-th channel and the l-th multiphase component, delayed by l full-bandwidth sampling points. The summation operation realizes the linear combination of the multiphase components, which is essentially the time-domain implementation of the synthesized filter bank system function, ensuring distortion-free filtering.

[0098] This step is the core reconstruction stage of multiphase synthesis, replacing the traditional "high sampling rate full filtering". It involves reconstructing the signal for each interpolated path. Based on its corresponding polyphase component index l, the signal is delayed by l original sampling rate sampling points, and then the K delayed signals are linearly superimposed. Since the frequency responses of the sub-filters are complementary, the superposition will produce an equivalent filtering effect completely consistent with the prototype low-pass filter, precisely removing the mirror sub-spectrum introduced by interpolation and retaining only the sub-band baseband spectrum. to (range), to avoid distortion of fault characteristics caused by mirror interference.

[0099] d) Combine and reconstruct the signals from each path into a full-bandwidth signal. Post-output:

[0100] .

[0101] Each sub-band A narrow band (bandwidth f) corresponding to the original signal s / K), and the spectra of each sub-band do not overlap, so that the superposition exactly covers the entire Nyquist range of the original signal (0 to f). s / 2). The merging process is free from frequency band interference, and the denoising effect of complex variational mode decomposition is fully preserved, ultimately outputting a high signal-to-noise ratio full-bandwidth acoustic signal. After synthesis processing, the noise reduction result is ensured to be compatible with the wind turbine condition monitoring platform, completing the complete closed loop of "frequency division processing - noise reduction - synthesis output".

[0102] In practice, the signal output can be flexibly selected according to the diagnostic needs of the wind turbine:

[0103] Full-bandwidth signal output: If subsequent processing requires the original frequency signal (such as acoustic signature extraction or visualization), the synthesized signal is output directly. For example, it can be used for Mel spectrum analysis or as input into a deep learning model to identify faults such as blade cracks.

[0104] Baseband signal output: If the diagnostic system can directly process the baseband signal (such as a real-time classification algorithm), the synthesis step can be omitted, and the measures can be used directly. Noise reduction modal components.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A speaker noise reduction method based on multiphase filtering and complex variational mode decomposition, characterized in that: Including the following measures: The system collects acoustic fingerprint signals from wind turbines during operation, performs multiphase filtering on the collected acoustic fingerprint signals, and divides the acoustic fingerprint signals into several sub-band signals. The average power of the sub-band signal is calculated, and compared with a dynamic threshold. Sub-band signals with an average power greater than or equal to the dynamic threshold are selected. The average power of the sub-band signal is calculated using the following formula: in, This represents the average power of the k-th sub-band signal. This represents the index of the signal sample in the sub-band signal, and N is the total length of the sequence of the k-th sub-band signal. This represents the signal value of the nth signal sample within the kth sub-band signal. The square of the modulus of the signal value represents the instantaneous energy of the signal; The dynamic threshold is obtained by the following method: collecting ambient wind speed, determining the dynamic adjustment coefficient β based on the ambient wind speed, and then calculating the dynamic threshold according to the following formula: In the formula, This represents the dynamic threshold set for the k-th sub-band signal; This represents the average power of the noise in the k-th subband. The power standard deviation of the k-th subband noise is represented by the subband noise obtained by polyphase filtering of a pure noise signal without a target sound source. III) Perform complex variational mode decomposition on the selected sub-band signals whose average power is greater than or equal to the dynamic threshold to obtain the denoised mode components, including: Constructing a constrained variational model: In the formula: It is the time-domain representation of the sub-band signal, where ∂ t Represents the time derivative. As a frequency shift operator, the modal spectrum is shifted to the baseband for bandwidth calculation; Represents modal components, It is an amplitude function. For phase function, The center frequency; By introducing the Lagrange multiplier λ(t) and the penalty factor α, the enhanced Lagrange function is constructed as follows: By enhancing the Lagrangian function, the constrained problem is transformed into an unconstrained optimization problem. The modal components are solved using the following steps. : (1) Initialization: Set the number of modes I, the penalty factor α, and the convergence tolerance ε, and randomly initialize the modes u. i (0) Center frequency w i (0) and multiplier λ (0) ; (2) Modal update: Performed in the frequency domain, with the center frequency and Lagrange multipliers fixed, the spectrum of each mode is updated; the modal update formula for the (n+1)th iteration is: in, , and They are respectively , Fourier transform of λ(t); (3) Center frequency update: With the modes and Lagrange multipliers fixed, update the center frequency of each mode to minimize the bandwidth, using the following formula: (4) Multiplier update: Update the Lagrange multipliers to strengthen the constraints: in, τ This is the step size parameter; (5) Convergence criterion: Calculate the relative error. If the following conditions are met: Then stop the iteration; otherwise return to step (2). (6) The frequency domain modal components obtained after the iteration are processed. The signal is converted back to the time domain by inverse Fourier transform, and the denoised modal components are obtained.

2. The acoustic signature denoising method based on multiphase filtering and complex variational mode decomposition according to claim 1, characterized in that: The value range of the dynamic adjustment coefficient β determined based on the ambient wind speed is 1.5-2.

0.

3. The acoustic signature denoising method based on multiphase filtering and complex variational mode decomposition according to claim 1, characterized in that: The determination of the dynamic adjustment coefficient β based on the ambient wind speed includes: β is 1.5 to 1.6 when the ambient wind speed is 3 to 10 m / s; β is 1.7 to 1.8 when the ambient wind speed is 10 to 18 m / s; and β is 1.9 to 2.0 when the ambient wind speed is 18 to 25 m / s.

4. The acoustic signature denoising method based on multiphase filtering and complex variational mode decomposition according to claim 1, characterized in that: measure The polyphase filtering of the acquired voiceprint signal described in the document includes: setting the number K of sub-band signals, and grouping the coefficients of the prototype low-pass filter to obtain the sub-filter coefficients. , In the formula, This represents the coefficient sequence of the prototype low-pass filter. This represents the coefficient index of the sub-filter. To represent the index of the sub-filter; the voiceprint signal Segmented signals are obtained by segmentation. : In the formula: m represents the time-domain sampling point index of the sub-band signal. Then, the sub-band signal is obtained by processing the segmented signal as follows. , In the formula, j is the imaginary unit, j 2 =-1, Indicates the index of the sub-band signal. The length of a single sub-filter is represented by , and N represents the total order of the prototype low-pass filter.

5. The acoustic signature denoising method based on multiphase filtering and complex variational mode decomposition according to claim 1, characterized in that: measure The calculation of the average power of the sub-band signal described in the article is as follows: only the average power of the sub-band signal from the 0th to the K / 2th sub-band is calculated, and the average power of the remaining sub-band signals is derived based on the signal symmetry.

6. The speaker noise reduction method based on multiphase filtering and complex variational mode decomposition according to any one of claims 1-5, characterized in that: It also includes measures ): Measures The obtained modal components are input into a polyphase synthesis filter bank for signal synthesis, including: a) performing synthesis sub-filtering on each modal component separately: In the formula, The output of the k-th mode component after being filtered by the l-th synthesizer filter. It is the discrete output signal of the k-th subband signal after being denoised by complex variational mode decomposition; It is the l-th synthesizer sub-filter in the polyphase synthesizer filter bank. a) Perform interpolation on each of the filtered signals to restore the sampling rate of each signal to the original full-bandwidth signal sampling rate: In the formula, n is the time index of the original sub-band signal sampling rate, n = K·m + r, r = 0, 1, ..., K-1; c) Perform delay shifting and superposition on each interpolated signal to achieve image subspectral suppression: In the formula, This is the final high-sampling-rate filtered output of the k-th signal; d) The interpolated signal of the k-th channel and the l-th multiphase component is delayed by l full-bandwidth sampling points; Post-output: .

Images