Propeller singing peak cluster frequency domain de-singing method and system based on equivalent reference spectrum

Abstract

The present application relates to a kind of propeller singing peak cluster frequency domain elimination method and system based on equivalent reference spectrum, wherein the method comprises: S1, obtains the time domain data of propeller noise to be processed, calculates power spectral density and frequency band sound pressure level, and identifies, confirms candidate support band;S2, according to whether to use no singing reference noise data to select reference mode or no reference mode;S3, calculate the spectral residual of the equivalent no singing reference spectrum formed by the spectrum to be processed relative to the current selected mode;S4, construct target power spectrum in linear power domain according to the selected mode;S5, according to the target power spectrum, the power of frequency band in target correction area is corrected;S6, calculate and output frequency band sound pressure level, overall sound level and 1 / 3 octave sound pressure level.The present application obtains equivalent singing elimination data with higher reliability at lower engineering cost under the premise of preserving the original noise spectral characteristics, which is used for propeller noise spectral characteristic analysis and test evaluation.

Description

Technical Field

[0001] This invention relates to the field of ship radiated noise, specifically to a method and system for frequency domain noise cancellation of propeller phonation peaks based on an equivalent reference spectrum. Background Technology

[0002] The issue of underwater radiated noise from ships has received continuous attention in recent years. Besides cavitation noise, propellers may exhibit singing noise under non-cavitation or weak cavitation conditions. In water tunnel or model tests, the time-domain signal of propeller noise acquired by hydrophones typically contains a mixture of background noise, equipment noise, propeller broadband noise, narrow-band mechanical components, and any singing noise components. In propeller water tunnel or model tests, some experimental data can be affected by propeller singing noise. Singing noise usually manifests as localized narrow-band spikes or peak clusters, interfering with the judgment of the true broadband noise level of the propeller. This effect is typically manifested in the frequency domain results as narrow-band peaks or multiple adjacent peaks that are significantly higher than the surrounding spectral levels near a certain frequency, i.e., singing peak clusters. Singing peak clusters can raise the local power spectral density, spectral sound pressure level, and related 1 / 3 octave band sound level, making the measured results unable to accurately reflect the propeller noise characteristics under non-singing conditions. In existing testing and engineering processes, propellers usually need to be repaired or reworked, which is costly and time-consuming. Furthermore, changes in blade geometry or test conditions may affect the original noise characteristics and reduce the comparability of test results.

[0003] Existing methods for processing propeller noise data typically include manual peak selection, fixed-band reduction, notch filtering, or local smoothing. While these methods can reduce local spikes, they still have shortcomings: First, the vocal melody is often not a single peak, but may consist of a cluster of vocal melody peaks, including the main peak, shoulder peaks, adjacent peaks, and related frequency ranges. Processing only the highest peak can easily miss adjacent anomalous components. Second, fixed-band or fixed-amplitude processing is difficult to adapt to the actual excess levels of different peak clusters, easily resulting in insufficient vocal melody reduction or damage to the normal spectral morphology due to unreasonable reduction range or amplitude. Third, existing methods usually lack a processing mechanism that uses the spectrum without vocal melody as a reference, making it difficult to apply a unified judgment and correction logic in cases where there is no vocal melody reference data and cases where there is no vocal melody reference data. Summary of the Invention

[0004] This invention aims to solve the problem of propeller singing interference in noise spectrum analysis. By focusing on singing peak clusters and generating corresponding equivalent reference spectra in both reference-based and non-reference-based modes, this invention reduces the singing component in propeller noise data through frequency domain correction. While preserving the original noise spectrum characteristics, it obtains highly reliable equivalent anechoic data with lower engineering costs for propeller noise spectrum characteristic analysis and experimental evaluation.

[0005] The present invention adopts the following technical solution:

[0006] The frequency domain elimination method for propeller singing peak clusters based on equivalent reference spectrum includes:

[0007] S1. Acquire the time-domain data of the propeller noise to be processed, calculate the power spectral density and frequency band sound pressure level using unified spectrum analysis parameters, and identify and confirm candidate support frequency bands;

[0008] S2, select a reference mode or a no-reference mode based on whether to use non-voiced reference noise data: in the reference mode, match the non-voiced reference spectrum to form a first equivalent non-voiced reference spectrum; in the no-reference mode, divide the confirmed candidate voiced peak clusters into repair groups to form logical repair groups, and estimate the normal spectrum trend based on the non-voiced reference frequency bands on both sides of the logical repair group to form a second equivalent non-voiced reference spectrum.

[0009] S3, calculate the spectral residual of the spectrum to be processed relative to the equivalent non-vocal reference spectrum formed by the currently selected mode; in the reference mode, determine the target correction region based on the spectral residual; in the no-reference mode, decompose the spectral residual into a broad trend residual and a local peak detail residual, determine the broad trend allowance and the local peak detail allowance based on the normal spectrum statistics on both sides of the logical repair group, and determine the target correction region based on the broad trend residual, the local peak detail residual, the corresponding allowance, and the member peak structure; set transition regions on both sides of the target correction region, and the target correction region and the transition region together constitute the correction support region;

[0010] S4. Construct the target power spectrum in the linear power domain according to the selected mode: In the reference mode, construct the target power spectrum based on the spectrum to be processed and the first equivalent non-voiced reference spectrum; In the non-reference mode, construct the target power spectrum that is no higher than the spectrum to be processed based on the second equivalent non-voiced reference spectrum, the broad trend allowance, the local peak detail allowance, and the member peak structure, and the frequency band that does not exceed the corresponding allowance remains unchanged.

[0011] S5, the frequency band power in the target correction region is corrected according to the target power spectrum, continuous connection is made in the transition region, and the original power spectrum is kept unchanged outside the correction support region to obtain the corrected power spectrum;

[0012] S6. Calculate the band sound pressure level, total sound level, and 1 / 3 octave band sound pressure level based on the corrected power spectrum, and output the corrected power spectrum, band sound pressure level results, and 1 / 3 octave band results.

[0013] Preferably, in step S2, in the reference mode, a reference matching frequency band is selected outside the candidate phonic frequency and its guard band of the candidate support frequency band. The spectral level difference between the spectrum to be processed and the user-provided phonic reference spectrum is iteratively removed for outliers and fitted with low-order polynomials to obtain a matching amount that varies with frequency. The matching amount is then applied to the phonic reference spectrum to form the first equivalent phonic reference spectrum.

[0014] Preferably, in step S2, in the no-reference mode, all candidate phonic peak clusters, member peaks and their supporting frequency bands are excluded. Non-phonic reference regions are selected on both sides of the logic repair group. The normal spectral trend within the logic repair group is estimated by robust statistics and low-order trend connections to form the second equivalent no-phonic reference spectrum. The wide-trend allowance and the local peak detail allowance are obtained by frequency interpolation based on the wide-trend statistics and normal local peak statistics in the non-phonic reference regions on both sides.

[0015] Preferably, the construction of the second equivalent reference spectrum includes: dividing the effective non-vocal reference regions on both sides of the current logical repair group into multiple frequency segments; for each frequency segment, excluding spectral values ​​higher than a preset quantile, and performing median statistics on the remaining spectral values ​​to obtain the corresponding normal spectral anchor point; and interpolating based on the multiple normal spectral anchor points located on the left and right sides of the current logical repair group to obtain the local reference baseline within the current logical repair group.

[0016] Preferably, in step S3, the difference between the sound pressure level of the frequency band to be processed and the sound pressure level of the equivalent non-voiced reference frequency band formed by the currently selected mode is used as the spectral residual. In the reference mode, the target correction region is determined based on the spectral residual of the spectrum to be processed relative to the first equivalent non-voiced reference spectrum. In the no-reference mode, the spectral residual of the spectrum to be processed relative to the second equivalent non-voiced reference spectrum is decomposed into a broad trend residual and a local peak detail residual based on local robust statistics, and the target correction region is determined based on the broad trend allowance, the local peak detail allowance, and the member peak structure. A transition region with continuous attenuation is set on both sides of the target correction region, and the target correction region and the transition region together constitute the correction support region. Frequency bands in the candidate search region that do not meet the corresponding correction conditions are not included in the target correction region.

[0017] Preferably, in step S4, in the reference mode, the target power spectrum Pt(f) is determined by the following formula:

[0018] ,,

[0019] Among them, P x (f) represents the bandwidth power of the spectrum to be processed, P refw(f) is the band power of the first equivalent non-vocal reference spectrum, and w(f) is the frequency weight; w(f) = 1 in the target correction region, w(f) = 0 outside the correction support region, and changes continuously in the transition region.

[0020] Preferably, in step S4, in the no-reference mode, the sound pressure level of the frequency band to be processed is higher than the positive residual of the second equivalent non-vocal reference spectrum. for:

[0021] ,

[0022] The positive residuals are decomposed into broad-trend residuals. and local peak detail residuals :

[0023] ,

[0024] Broad trend target residual and local peak detail target residual They are respectively:

[0025]

[0026] ,

[0027] Target frequency band sound pressure level and target power spectrum They are respectively:

[0028]

[0029] ,

[0030] Among them, L x (f) represents the sound pressure level of the frequency band to be processed, L b (f) represents the band sound pressure level of the second equivalent non-vocal reference spectrum, A board (f) represents the allowable range for a broad trend, A peak (f) represents the allowable amount of local peak detail. As the core area of ​​the member summit, p0 is the reference sound pressure level, which is a texture protection zone.

[0031] 8. The method according to claim 5, characterized in that, in step S5, P is taken within the target correction region. c (f)=P t (f) Take P outside the texture protection area and the correction support area. c (f)=P x (f) Correcting the power spectrum in the transition region. Determine using the following formula:

[0032] ,

[0033] Where g(f) is a transition weight that varies continuously between 0 and 1. .

[0034] This invention also provides a propeller-based peak cluster elimination system, comprising: a spectrum analysis module for obtaining power spectral density and frequency band sound pressure level from the time-domain data to be processed; a mode selection and equivalent reference spectrum construction module for independently executing a reference-based or non-reference-based processing path according to user selection and forming a corresponding equivalent non-singing reference spectrum; a peak cluster identification module for determining candidate singing peak clusters, member peaks, and corrected support regions; and a spectrum correction module for constructing a target power spectrum in the linear power domain and forming a continuous transition.

[0035] Compared with existing technologies, this invention offers the following advantages: It establishes a frequency domain noise reduction process for propeller singing noise based on "equivalent reference spectrum—singing peak clusters—frequency domain correction." It sets up a reference-based mode and a no-reference mode. With a reference, the no-singing spectrum is used for matching correction; without a reference, the target spectrum is constructed solely based on the statistical analysis of the local normal spectra on both sides of the singing spectrum. Both modes can generate an equivalent reference spectrum adapted to the current experimental data, improving the method's adaptability to different test conditions and data completeness. Furthermore, this invention identifies singing peak clusters rather than single spikes, and in the no-reference mode, it distinguishes between peak cluster width trend anomalies and member peak detail anomalies, correcting only the portions exceeding the locally permissible normal range, thus protecting the normal spectral texture between peaks and non-target frequency bands. The correction results are used to reduce the dominant influence of singing noise on the total sound level, spectrogram, and 1 / 3 octave band results, thereby improving the comparability of test results under different operating conditions. Attached Figure Description

[0036] Figure 1 This is a schematic diagram showing the comparison before and after cancellation in no-reference mode for Embodiment 1 of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention.

[0037] Figure 2 This is a schematic diagram showing the comparison before and after cancellation in reference mode for Embodiment 1 of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention.

[0038] Figure 3 This is a schematic diagram of the data after spectral analysis of the noise to be processed in no-reference mode, which is an embodiment 2 of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention.

[0039] Figure 4 This is a schematic diagram of the candidate peak cluster grouping and local normal spectrum determined in the no-reference mode of Embodiment 2 of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention.

[0040] Figure 5 This is a schematic diagram of the repair group frequency band range determined in the no-reference mode of Embodiment 2 of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention.

[0041] Figure 6 This is a schematic diagram showing the comparison before and after the cancellation of the propeller singing peak cluster frequency domain cancellation method based on the equivalent reference spectrum in the no-reference mode of Embodiment 2 of the present invention.

[0042] Figure 7 This is a schematic diagram showing the comparison of local peak clusters before and after the cancellation in no-reference mode, in Embodiment 2 of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention.

[0043] Figure 8 This is a schematic diagram showing the comparison of 1 / 3 octave band results before and after the completion of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum in Example 2 of the present invention in the no-reference mode.

[0044] Figure 9 This is a schematic diagram of the candidate vocal peak clusters determined in reference mode in Embodiment 2 of the frequency domain cancellation method for propeller vocal peak clusters based on equivalent reference spectrum of the present invention.

[0045] Figure 10 This is a schematic diagram showing the comparison before and after the cancellation is completed in reference mode, in Embodiment 2 of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention.

[0046] Figure 11 This is a schematic diagram showing the comparison of local peak clusters before and after the cancellation in reference mode, in Embodiment 2 of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention.

[0047] Figure 12 This is a schematic diagram showing the comparison of 1 / 3 octave band before and after the cancellation of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum in Embodiment 2 of the present invention in the reference mode.

[0048] Figure 13 This is a schematic diagram of the overall process of the propeller singing peak cluster frequency domain cancellation method based on the equivalent reference spectrum of the present invention.

[0049] Figure 14 The flowchart illustrates the process of obtaining the first equivalent non-vocal reference spectrum in the reference mode of the propeller vocal peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention.

[0050] Figure 15 This is a flowchart illustrating the construction of the target spectrum in the reference mode of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention.

[0051] Figure 16 This is a flowchart of the overall process of the propeller singing peak cluster frequency domain cancellation method based on equivalent reference spectrum in the no-reference mode of the present invention.

[0052] Figure 17 The flowchart illustrates the process of constructing and modifying a second equivalent no-voice reference spectrum in the no-reference mode of the propeller vocal peak cluster frequency domain cancellation method based on equivalent reference spectrum of the present invention. Detailed Implementation

[0053] This invention provides a frequency domain cancellation method for propeller singing noise based on an equivalent reference spectrum. The singing noise peak cluster includes the main peak, shoulder peaks, adjacent peaks, and local anomalous frequency regions within their associated frequency ranges. This invention establishes a frequency domain cancellation process for propeller singing noise around the concepts of "equivalent reference spectrum—singing noise peak cluster—frequency domain correction".

[0054] In this invention, the reference mode and the no-reference mode are two independent optional processing paths. The two modes share spectrum analysis, candidate confirmation, inverse Fourier transform and output interface, but execute corresponding processes in the construction of equivalent reference spectrum and target power spectrum respectively.

[0055] The method of the present invention specifically includes the following steps, the overall process of which is shown below. Figure 13 In reference mode, the process for obtaining the first equivalent unvoiced reference score is as follows: Figure 14 The process for constructing the target spectrum is as follows: Figure 15 For the overall process in no-reference mode, see [link to documentation]. Figure 16 The process for constructing and revising the second equivalent non-vocal reference score is described in [link to documentation]. Figure 17 .

[0056] S1. Perform frequency domain analysis on the vocal noise data. This step includes: a. Input the time-domain noise data to be processed; if a reference mode is selected, input the non-vocal reference time-domain noise data separately; if a no-reference mode is selected, do not read the non-vocal reference data; b. Read and preprocess the input data; c. Calculate the one-sided power spectral density using the Welch method; d. Integrate the power spectral density in the frequency band and convert it to the frequency band sound pressure level based on the reference sound pressure. The data to be processed and the non-vocal reference data are analyzed using the same window function, segment length, overlap rate, frequency range, and frequency band integration boundary.

[0057] The Welch method was used to calculate the one-sided power spectral density of the data to be processed. The power spectral density was then integrated over a frequency band to obtain the mean square sound pressure level (MSP) of the frequency band, which was then converted to the frequency band SPP level based on the reference SPP. The data to be processed and the reference data without vocalization were analyzed using the same window function, segment length, overlap rate, frequency range, and frequency band integration boundary.

[0058] Candidate search is conducted within a preset vocal search frequency band. Based on the local spectral baseline and residual envelope, as well as broadband uplift, local peak prominence, continuous excess interval, peak-to-valley value, frequency band occupancy, and integral excess, candidate vocal peak clusters are formed, member peaks and candidate supporting frequency bands are identified, and the candidate results are confirmed.

[0059] S2, depending on whether phonon-free reference noise data is used, selects either a reference mode or a no-reference mode to construct an equivalent reference spectrum characterizing the spectral variation trend in the phonon-free state. The two modes are independent of each other and do not call input data or intermediate results from each other in a single processing step.

[0060] S2.1 In reference mode, read the data to be processed and the user-provided non-vocal reference data. The same spectral analysis parameters are used for both the data to be processed and the non-vocal reference data. For the q-th candidate vocal peak cluster, select left and right local reference matching bands outside its candidate support band. q is the candidate vocal peak cluster number.

[0061] Calculate the spectral level difference between the spectrum to be processed and the unvoiced reference spectrum within the local reference matching frequency band. Perform a polynomial fitting of no higher than second order with frequency as the independent variable and spectral level difference as the dependent variable. Iterate and remove outlier frequency points based on the fitting residuals, and denote the set of valid reference matching frequency points as Ω. q The reference matching quantity for the q-th candidate vocal peak cluster is represented as:

[0062] ,

[0063] In the formula, ΔL q Ω is the reference matching value for the q-th candidate vocal peak cluster; q υ is the set of valid reference matching frequency points retained after iteratively removing outliers; υ is the set Ω. q Frequency within; L x (υ) and L r (υ) represents the frequency band sound pressure level of the spectrum to be processed and the reference spectrum without vocal accompaniment at frequency υ, respectively.

[0064] Within the candidate support band of the q-th candidate vocal peak cluster, the reference matching quantity ΔL q Applying this to the non-voiced reference spectrum yields the first equivalent non-voiced reference spectrum, whose frequency band sound pressure level... for:

[0065] ,

[0066] In the formula, f is the frequency within the candidate support band; L ref,q (f) represents the band sound pressure level of the first equivalent non-vocal reference spectrum at frequency f; L r (f) represents the band sound pressure level of the unaphonic reference spectrum at frequency f.

[0067] Corresponding bandwidth power for: P x (f) represents the bandwidth power of the unaphonic reference spectrum at frequency f.

[0068] S2.2 In the no-reference mode, identify member peaks for confirmed candidate singularity peak clusters; when the candidate support frequency bands of adjacent candidate singularity peak clusters are in contact or overlap with each other, the baseline anchor point is occupied by other candidate anomalous frequency bands, or they cannot form effective external references respectively, they are merged into the same logical repair group.

[0069] When constructing the second equivalent non-vocal reference spectrum, all candidate vocal peak clusters, member peak frequency bands, candidate support frequency bands, and frequency bands occupied by other logical repair groups are used as exclusion frequency bands. Valid non-vocal reference regions are searched outwards from both sides of the current logical repair group. The non-vocal reference region is divided into multiple frequency segments. For each frequency segment, spectral values ​​higher than a preset quantile are excluded. The truncated median of the spectral values ​​for each frequency segment is statistically analyzed to reduce the impact of local abnormal peaks on the reference spectrum estimation, resulting in multiple normal spectral anchor points. Interpolation is performed based on the normal spectral anchor points located on both sides of the current logical repair group to estimate the local normal spectral trend within the current logical repair group, forming the second equivalent non-vocal reference spectrum.

[0070] When both sides of the logic repair group have normal spectrum anchor points that meet the requirements, the reference spectrum estimation result is recorded as valid; when either side lacks a valid normal spectrum anchor point, the entire logic repair group is not extrapolated without constraints based on the spectrum of one side, and the corresponding logic repair group is marked as needing to be reviewed or not entering the subsequent correction.

[0071] S3, Calculate the spectral residual of the sound pressure level in the frequency band to be processed relative to the equivalent non-voiced reference frequency band sound pressure level formed by the currently selected mode. :

[0072] ,

[0073] In the reference mode, based on the spectral residual of the spectrum to be processed relative to the first equivalent non-vocal reference spectrum, the frequency bands within the confirmed candidate vocal peak clusters that meet the preset correction conditions are determined as the target correction area.

[0074] In the no-reference mode, a local low quantile or equivalent robust method is used to extract the broad-trend residual from the spectral residual of the spectrum to be processed relative to the second equivalent no-voice reference spectrum. :

[0075] ,

[0076] In the formula, Rtotal (f) represents the total spectral residual at frequency u; Q α {} represents quantile operation; α is the preset low quantile parameter, which defaults to 25%; W(f) is a local sliding frequency window centered on frequency f; u is the frequency within the local sliding frequency window.

[0077] Local peak detail residuals Defined as:

[0078] ,

[0079] In the formula, R detail (f) represents the local peak detail residual at frequency f, which is higher than the broad trend residual.

[0080] Within the non-phonic reference areas on the left and right sides of the logical restoration group, normal wide-trend and normal local peaks were statistically analyzed. The permissible amount of local peaks on the left and right sides was determined. , It can be represented as

[0081] ,

[0082] ,

[0083] In the formula, H L and H R These are the sets of local peak values ​​of the local peak detail residuals in the non-phonic reference regions on the left and right sides of the logical repair group, respectively; Q 90 () represents the 90th percentile; MAD() represents the median absolute deviation of each value in the set from its median; λ is the scaling correction factor; U L and U R These represent the allowable amount of local peak detail on the left and right sides, respectively, with a default value of 1.4826.

[0084] Based on the left and right reference statistics, frequency interpolation is performed within the repair group to obtain the broad trend allowance A. board (f) and the allowable amount of local peak detail A peak (f).

[0085] The frequency bands where the broad trend residual exceeds the broad trend allowable limit, and the frequency bands where the local peak detail residual within the core region of a member peak exceeds the local peak detail allowable limit, are defined as the target correction region in the no-reference mode. The frequency bands where the inter-peak residual does not exceed the corresponding allowable limit are defined as the texture protection region. Candidate search regions are only used to detect and organize anomalies and are not directly used as continuous correction regions. Continuously attenuating transition regions are set on both sides of the target correction region, and the target correction region and the transition region together constitute the correction support region.

[0086] S4. If there is no peak cluster, the original spectrum remains unchanged, and the frequency domain and 1 / 3 octave band results calculated from the original data are output; if there is a peak cluster, the target frequency band power is constructed in the linear power domain.

[0087] S4.1 In reference mode, the target power spectrum P t (f) Determined by the following formula:

[0088] ,

[0089] In the formula, P x (f) represents the bandwidth power of the spectrum to be processed, P ref，q (f) represents the band power of the first equivalent non-vocal reference spectrum, w q (f) represents the frequency weight, within the target correction region w q (f)=1, outside the modified support zone w q (f)=0, and changes continuously within the transition region.

[0090] S4.2 In the no-reference mode, the positive residual where the sound pressure level of the frequency band to be processed is higher than the second equivalent non-vocal reference spectrum is: ,

[0091] The positive residuals are decomposed into broad-trend residuals. and local peak detail residuals :

[0092] ,

[0093] The broad-trend target residual is:

[0094] ,

[0095] The residuals of local peak details within the core region of the member peaks are:

[0096] , ,

[0097] Taken within the texture protection zone:

[0098] , ,

[0099] The target sound pressure level and target power spectrum are as follows:

[0100] ,

[0101] ,

[0102] In the formula, L x (f) represents the sound pressure level of the frequency band to be processed, L b(f) represents the band sound pressure level of the second equivalent non-vocal reference spectrum, A board (f) represents the allowable range for a broad trend, A peak (f) represents the allowable amount of local peak detail. As the core area of ​​the member summit, p0 is the reference sound pressure level, which is a texture protection zone.

[0103] Therefore, only abnormal power exceeding the local normal allowable range is reduced, without adjusting frequency bands below the baseline, without causing the target power to exceed the power to be processed, and without using a fixed repair intensity to determine the reduction amount for all operating conditions.

[0104] S5, take P within the target correction area. c (f)= P t (f) Take P outside the texture protection area and the correction support area. c (f)= P x (f). Continuous weighted connections are used within the transition region:

[0105] ,

[0106] This yields the corrected power spectrum, and based on the corrected power spectrum, the band sound pressure level, the total sound level, and the 1 / 3 octave band sound pressure level are calculated.

[0107] S6 outputs the corrected power spectrum, frequency band sound pressure level results, and 1 / 3 octave band results.

[0108] Example 1: A set of ideal singing voice data, with a time domain data sampling rate of 51200Hz and a sampling time of 6s.

[0109] (1) In no-reference mode: the Hann window Welch method with 4s segmentation and 50% overlap was used for analysis. After executing the program, the sound pressure level of the phonograph band to be eliminated at 8500Hz was found to be 96.295dB. Figure 1 As shown, the equivalent reference sound pressure level estimated from the local frequency bands on both sides of the candidate frequency is 58.874 dB. The core frequency range of the vocal peak cluster identified by the program is 8497.5 Hz to 8502.5 Hz, and the corrected support range including the transition region is 8493.5 Hz to 8506.5 Hz. After band power correction, the band sound pressure level at 8500 Hz is 58.858 dB, with a residual of -0.016 dB relative to the second equivalent non-vocal reference spectrum; the 95th percentile of the change in band sound pressure level outside the corrected support region is 1.65 × 10⁻⁶. -8 dB.

[0110] (2) In reference mode, the data to be processed is the same as in Example 1, and the user-selected non-vocal reference time domain data is also input. The original non-vocal reference spectrum has a frequency band sound pressure level of 54.258 dB at 8500 Hz, such as... Figure 2 As shown, by excluding candidate phonic frequencies and their guard bands, the local reference matching at 8500Hz yields a matching value of 4.050dB, resulting in a first equivalent reference spectrum of 58.308dB at 8500Hz. Spectral-level matching is then performed on the phonic reference spectrum to form the first equivalent phonic reference spectrum, which is then used to construct the target power spectrum and correct the corresponding frequency band power.

[0111] Example 2: A set of time-domain data to be de-sampling was tested with a sampling rate of 262144Hz and a duration of 30s. In no-reference mode, data input and spectrum analysis were first performed, such as... Figure 3 As shown, candidate peak clusters are then grouped and local normal spectra are determined based on the relationships between candidate peak clusters, member peak frequency bands, candidate support frequency bands, and adjacent candidates, such as... Figure 4 As shown, a referenceless local baseline is constructed within the candidate search frequency band from 5000Hz to 20000Hz, and the residual of the input spectrum relative to the local baseline and its local envelope are calculated. The program detects candidate member peaks based on a 3dB wide anomalous recall threshold and a 10dB local peak recall threshold, and merges adjacent member peaks into candidate peak clusters based on the frequency spacing between member peaks and the degree of valley fall between peaks, as shown. Figure 5 As shown.

[0112] Next, the identification and correction of broad trends and member peak details are performed. For the logic repair group entering the processing stage, the spectrum to be processed is compared with the local normal baseline to identify anomalous spectral components that are higher than the local normal baseline. Based on the frequency distribution characteristics of the anomalous spectral components, they are divided into broad trend components that reflect the overall shape of the peak cluster and member peak detail components superimposed on the broad trend components.

[0113] Within the normal frequency bands on both sides of the logic repair group, the wide-trend fluctuation range and local peak fluctuation range of the normal spectrum are statistically analyzed to determine the wide-trend allowable amount and the member peak allowable amount. When the wide-trend component and the member peak detail component exceed the allowable amount, the corresponding frequency band is corrected; the spectrum components that do not exceed the corresponding allowable amount are retained; frequency components that are originally not higher than the local normal baseline are not adjusted upwards or modified additionally.

[0114] Based on the power relationship between the target spectrum and the spectrum to be processed, the adjustment amount for each frequency component is determined, and the adjustment amount is applied to the corresponding frequency range of the data to be processed. After the above processing, the corrected output spectrum is obtained. The comparison before and after cancellation is as follows: Figure 6-8 As shown, the recalculated 1Hz frequency spectrum was used as the verification basis for the sound suppression effect and the preservation of the non-target frequency band.

[0115] The process for obtaining the target map in reference mode is as follows: Figure 15 :

[0116] First, the candidate peak clusters and equivalent unaphonic reference spectra are determined. This example inputs the noise data to be processed and the unaphonic noise data. The sampling frequency for both sets of data is 262144Hz, and the sampling time is 30s. The spectral level difference between the spectrum to be processed and the unaphonic reference spectrum within the candidate search frequency band is calculated. Based on the median and absolute median difference of the spectral level difference, frequency points that significantly deviate from the overall difference level are identified and excluded, obtaining the overall spectral level matching amount of the reference spectrum. Based on this, the original reference spectrum is adjusted. The spectrum to be processed is compared with the matched reference spectrum. Based on the abnormal spectral level difference between the two, local peaks and their continuous abnormal frequency bands are identified. According to the frequency interval between adjacent peaks and the degree of spectral level decline between peaks, the interconnected local peaks are organized into candidate unaphonic peak clusters, such as... Figure 9 As shown.

[0117] Secondly, target spectrum correction and output spectrum verification are performed. The spectral level difference between the spectrum to be processed and the reference spectrum is calculated within the reference frequency bands on both sides of the candidate vocal peak clusters, and trend fitting is performed to eliminate abnormal frequency points. Then, the retained frequency points are used to refit and obtain the local spectral level matching amount corresponding to the candidate peak clusters. Based on the difference between the spectrum to be processed and the non-vocal reference spectrum, the target core region, transition region, and correction support region are determined. Within the target core region, the target power spectrum is determined based on the non-vocal reference spectrum; within the transition region, the spectrum to be processed and the non-vocal reference spectrum are smoothly connected in the power domain according to a weight that changes continuously with frequency; outside the correction support region, the original spectrum of the data to be processed is kept basically unchanged. Through the above processing, the target spectrum is adapted to the local normal spectral level of the data to be processed, and the risk of discontinuous changes at the boundary of the correction support region is reduced. The frequency range corresponding to each candidate vocal peak cluster is corrected according to the target power spectrum. After the initial correction is completed, the output spectrum is recalculated using the same spectral analysis conditions as before processing, and the difference between the actual output spectrum and the target spectrum is compared. Figure 10-12 For local frequency ranges that still deviate from the target spectrum, further corrections are made to reduce the deviation between the actual output spectrum and the target spectrum.

[0118] This invention also provides a propeller-based vocal peak cluster elimination system that can perform the above-described method. The system includes: a spectrum analysis module for obtaining power spectral density and band sound pressure level from the time-domain data to be processed; a mode selection and equivalent reference spectrum construction module for independently executing a reference-based or non-reference-based processing path according to user selection and forming a corresponding equivalent non-vocal reference spectrum; a peak cluster identification module for determining candidate vocal peak clusters, member peaks, and corrected support regions; and a spectrum correction module for constructing a target power spectrum in the linear power domain and forming a continuous transition.

[0119] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Various modifications or equivalent substitutions made by those skilled in the art to the present invention within its spirit and scope of protection also fall within the scope of protection of the present invention.

Claims

1. A method for frequency domain elimination of propeller vocal peak clusters based on equivalent reference spectrum, characterized in that, include: S1. Acquire the time-domain data of the propeller noise to be processed, calculate the power spectral density and frequency band sound pressure level using unified spectrum analysis parameters, and identify and confirm candidate support frequency bands; S2, select a reference mode or a no-reference mode based on whether to use no-voice reference noise data: in the reference mode, match the no-voice reference spectrum to form the first equivalent no-voice reference spectrum; In the no-reference mode, the confirmed candidate singular peak clusters are divided into repair groups to form logical repair groups, and the normal spectrum trend is estimated based on the non-singular reference frequency bands on both sides of the logical repair group to form a second equivalent no-singular reference spectrum. S3, calculate the spectral residual of the spectrum to be processed relative to the equivalent non-vocal reference spectrum formed by the currently selected mode; in the reference mode, determine the target correction region based on the spectral residual; In the no-reference mode, the spectral residual is decomposed into a broad trend residual and a local peak detail residual. The allowable amount of the broad trend and the allowable amount of the local peak detail are determined based on the normal spectrum statistics on both sides of the logical repair group. The target correction region is determined based on the broad trend residual, the local peak detail residual, the corresponding allowable amount, and the member peak structure. A transition region is set on both sides of the target correction region. The target correction region and the transition region together constitute the correction support region. S4. Construct the target power spectrum in the linear power domain according to the selected mode: In the reference mode, construct the target power spectrum based on the spectrum to be processed and the first equivalent non-voiced reference spectrum; In the non-reference mode, construct the target power spectrum that is no higher than the spectrum to be processed based on the second equivalent non-voiced reference spectrum, the broad trend allowance, the local peak detail allowance, and the member peak structure, and the frequency band that does not exceed the corresponding allowance remains unchanged. S5, the frequency band power in the target correction region is corrected according to the target power spectrum, continuous connection is made in the transition region, and the original power spectrum is kept unchanged outside the correction support region to obtain the corrected power spectrum; S6. Calculate the band sound pressure level, total sound level, and 1 / 3 octave band sound pressure level based on the corrected power spectrum, and output the corrected power spectrum, band sound pressure level results, and 1 / 3 octave band results.

2. The method according to claim 1, characterized in that, In step S2, in the reference mode, a reference matching frequency band is selected outside the candidate phonic frequency and its guard band of the candidate support frequency band. The spectral level difference between the spectrum to be processed and the user-provided phonic reference spectrum is iteratively removed for outliers and fitted with low-order polynomials to obtain a matching amount that varies with frequency. The matching amount is then applied to the phonic reference spectrum to form the first equivalent phonic reference spectrum.

3. The method according to claim 1, characterized in that, In step S2, in the no-reference mode, all candidate phonic peak clusters, member peaks and their supporting frequency bands are used as exclusion frequency bands. Non-phonic reference regions are selected on both sides of the logic repair group. The normal spectral trend within the logic repair group is estimated by robust statistics and low-order trend connectivity to form the second equivalent no-phonic reference spectrum. The wide-trend allowance and local peak detail allowance are obtained by frequency interpolation based on the wide-trend statistics and normal local peak statistics in the non-phonic reference regions on both sides.

4. The method according to claim 3, characterized in that, The construction of the second equivalent reference spectrum includes: dividing the effective non-singing reference regions on both sides of the current logical repair group into multiple frequency segments; for each frequency segment, excluding spectral values ​​higher than a preset quantile, and performing median statistics on the remaining spectral values ​​to obtain the corresponding normal spectral anchor point; and interpolating based on the multiple normal spectral anchor points located on the left and right sides of the current logical repair group to obtain the local reference baseline within the current logical repair group.

5. The method according to claim 1, characterized in that, In step S3, the difference between the sound pressure level of the frequency band to be processed and the sound pressure level of the equivalent non-voiced reference frequency band formed by the currently selected mode is used as the spectral residual; in the reference mode, the target correction area is determined based on the spectral residual of the spectrum to be processed relative to the first equivalent non-voiced reference spectrum. In the no-reference mode, the spectral residual of the spectrum to be processed relative to the second equivalent no-voice reference spectrum is decomposed into a broad trend residual and a local peak detail residual based on local robust statistics. The target correction region is determined based on the broad trend allowance, the local peak detail allowance, and the member peak structure. A continuously attenuating transition region is set on both sides of the target correction region. The target correction region and the transition region together constitute the correction support region. Frequency bands in the candidate search region that do not meet the corresponding correction conditions are not included in the target correction region.

6. The method according to claim 5, characterized in that, In step S4, under reference mode, the target power spectrum P t (f) Determined by the following formula: ， Among them, P x (f) represents the bandwidth power of the spectrum to be processed, P ref w(f) is the band power of the first equivalent non-vocal reference spectrum, and w(f) is the frequency weight; w(f) = 1 in the target correction region, w(f) = 0 outside the correction support region, and changes continuously in the transition region.

7. The method according to claim 5, characterized in that, In step S4, in the no-reference mode, the sound pressure level of the frequency band to be processed is higher than the positive residual of the second equivalent non-voiced reference spectrum. for: ， The positive residuals are decomposed into broad-trend residuals. and local peak detail residuals : ， Broad trend target residual and local peak detail target residual They are respectively: ， ， Target frequency band sound pressure level and target power spectrum They are respectively: ， ， Among them, L x (f) represents the sound pressure level of the frequency band to be processed, L b (f) represents the band sound pressure level of the second equivalent non-vocal reference spectrum, A board (f) represents the allowable range for a broad trend, A peak (f) represents the allowable amount of local peak detail. As the core area of ​​the member summit, p0 is the reference sound pressure level, which is a texture protection zone.

8. The method according to claim 5, characterized in that, In step S5, P is taken within the target correction region. c (f)=P t (f) Take P outside the texture protection area and the correction support area. c (f)=P x (f) Correcting the power spectrum in the transition region. Determine using the following formula: ， Where g(f) is a transition weight that varies continuously between 0 and 1. .

9. A propeller-driven peak elimination system, characterized in that, A method for performing any one of claims 1-8, comprising: a spectrum analysis module for obtaining power spectral density and band sound pressure level from time-domain data to be processed; a mode selection and equivalent reference spectrum construction module for independently executing a reference processing path or a non-reference processing path according to user selection and forming a corresponding equivalent non-vocal reference spectrum; a peak cluster identification module for determining candidate vocal peak clusters, member peaks, and corrected support regions; and a spectrum correction module for constructing a target power spectrum in the linear power domain and forming a continuous transition.

Images