Self-adaptive noise reduction filtering method, system, medium and equipment

Abstract

The invention discloses a self-adaptive noise reduction filtering method and system, a medium and equipment. The method comprises the following steps: carrying out direct current offset removal preprocessing on an input dynamic signal; generating a time-frequency representation matrix through time-frequency transformation, and calculating an amplitude spectrum according to the time-frequency representation matrix; analyzing the amplitude spectrum along a time axis, extracting short-time and long-time noise amplitudes, and fusing the short-time and long-time noise amplitudes to generate a basic noise amplitude; calculating a signal-to-noise amplitude ratio based on the basic noise amplitude, and determining a final noise amplitude; calculating a noise reduction spectrum amplitude according to the final noise amplitude, retaining original phase information, and reconstructing a noise reduction time domain signal through inverse transformation; the system comprises a preprocessing module, an amplitude spectrum generation module, a dual-time-scale noise estimation module, a denoising module and a noise reduction and signal reconstruction module. According to the method, through dual-time scale noise estimation, dynamic distribution of noise is accurately described, and in combination with self-adaptive noise reduction adjustment based on a signal-to-noise amplitude ratio, complex and non-stationary noise is effectively suppressed, and meanwhile, the relative strength of a characteristic frequency band is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing, in particular to an adaptive noise reduction filtering method, system, medium and device. BACKGROUND

[0002] Dynamic signals have wide applications in industrial, energy, medical and communication fields, such as mechanical equipment running sound, biomedical signals (such as electrocardiogram or electroencephalogram), wireless communication signals, etc. These signals usually contain rich state information, which can reflect the health condition of equipment, the running characteristics of system or biological activities. However, dynamic signals are easily disturbed by complex environmental noise, which has non-stationary, wide-spectrum and time-varying characteristics, often mixed with the spectrum of target signals, significantly reducing the accuracy of signal analysis. For example, in industrial environment, mechanical vibration, turbulent noise, transient interference and background noise are highly overlapped with target signals, making it difficult to extract features and limiting the application of signals in condition monitoring and fault diagnosis.

[0003] Traditional noise reduction methods, such as spectral subtraction, Wiener filtering and wavelet denoising, rely on static noise models or fixed parameters, which are difficult to adapt to the non-stationary characteristics of dynamic noise, resulting in signal quality degradation or distortion of key features. Spectral subtraction suppresses stationary noise by estimating the noise power spectrum, but it is easy to introduce artifacts in non-stationary scenarios; Wiener filtering is based on the least mean square error criterion, and its performance is limited when dynamic noise changes rapidly; although wavelet denoising provides time-frequency resolution through multi-scale decomposition, improper threshold selection can lead to feature distortion in complex noise environment, especially when noise and target signal spectrum overlap, the separation efficiency is low.

[0004] Signal decomposition technology is an important direction of dynamic signal processing, which supports feature extraction and noise suppression by decomposing the signal into IMF components. However, empirical mode decomposition (EMD) is affected by mode mixing and end effect, generating false components under strong noise; variational mode decomposition (VMD) is more robust, but the parameter adjustment is complex and the computational cost is high, which conflicts with the real-time application requirements; non-negative matrix factorization (NMF) and wavelet packet denoising can extract periodic features in specific scenarios, but face strong noise interference, the accuracy decreases and the computational overhead is large; local mean decomposition (LMD) and singular spectrum analysis (SSA) are effective, but they are highly sensitive to noise and difficult to handle large-scale dynamic signals.

[0005] In recent years, with the increasing demand for non-stationary signal processing, dynamic noise reduction technology has made progress, such as adaptive weighted mean filter (AWMF) combined with gradient descent enhancement, sparse attention denoising and deep learning-based noise estimation perform well in signal enhancement or semi-stationary noise scenarios, but these methods are not suitable for complex industrial noise environments. Deep learning methods are particularly difficult to meet the low latency requirements of real-time monitoring due to their reliance on large amounts of training data and high computational resources, and the time-frequency changes of dynamic noise pose higher challenges to the dynamic response and robustness of the algorithm.

[0006] In addition, traditional noise reduction methods are prone to introduce false signals or overfitting noise, reducing reliability, while new methods such as neural Kalman filter, recursive denoising learning and wavelet denoising based on Mahalanobis distance have made progress in stationary or semi-stationary scenarios, but they are not suitable for non-stationary noise, and the signal feature protection capability is limited, making it difficult to suppress noise while avoiding interference with key features, and the versatility needs to be improved.

[0007] Therefore, there is an urgent need for an adaptive noise reduction method that can effectively suppress dynamic noise while strengthening the relative strength of the feature band to meet the needs of diverse signal processing scenarios. SUMMARY

[0008] Therefore, the present application provides an adaptive noise reduction filtering method, system, medium and device to at least solve the problem of existing noise reduction methods interfering with key signal features while suppressing noise and multi-scene adaptability.

[0009] To achieve the above purpose, the present application adopts the following technical solutions: An adaptive noise reduction filtering method, comprising the following steps: S1. obtaining a dynamic signal from a signal receiving device , obtaining a corresponding time-domain signal through preprocessing, wherein is the signal length; S2. obtaining a corresponding time-frequency representation matrix through time-frequency transformation, used to represent the complex amplitude of the th frame and the th frequency point, and calculating the amplitude spectrum according to ; S3. using a double-time-scale noise estimation method to analyze the amplitude spectrum along the time axis, extracting short-time noise amplitude and long-time noise amplitude , introducing dynamic weight to fuse and to generate the basic noise amplitude ; S4. Based on the signal noise amplitude ratio , according to the minimum amplitude threshold value , and based on the dynamic noise scaling factor and the estimated noise amplitude ; S5. According to and the denoised spectrum amplitude , the original phase information is retained, and is reconstructed into a denoised time domain signal by inverse transformation.

[0010] Preferably, the specific content of S1 preprocessing includes: acquiring a dynamic signal from a signal receiving device , performing DC offset preprocessing on , removing low-frequency interference, and obtaining a preprocessed time domain signal .

[0011] Preferably, the specific content of S2 includes: obtaining the corresponding time-frequency representation matrix by time-frequency transformation, wherein the method of time-frequency transformation specifically includes: short-time Fourier transform STFT, fast Fourier transform FFT, wavelet transform or wavelet packet transform; Based on the time-frequency representation matrix , the amplitude spectrum is calculated: .

[0012] Preferably, the specific content of time-frequency transformation by short-time Fourier transform STFT method includes: ; ; ; In the formula, is the window length, indicating the number of sampling points of each time window, is the frame shift, represents the time domain sample of the input dynamic signal in the frame; represents the window function, represents the total number of frames, represents the floor function, is the frequency index of the time-frequency representation matrix, , , is the number of points of the time-frequency transform, is the imaginary unit; is the time frame index, ; is the number of overlapping samples between adjacent frames.

[0013] Preferably, the amplitude spectrum is analyzed along the time axis in S3 to extract the short-time noise amplitude and the long-time noise amplitude The specific content includes: A double-time-scale noise estimation is adopted to calculate the short-time noise amplitude and the long-time noise amplitude by analyzing the statistical distribution of the amplitude spectrum along the time axis: ; ; wherein and denote the first quartile and the first quartile, respectively; and are the quartile thresholds for the short-time and long-time noise estimation, respectively; reflects the local background level of the noise, and captures the overall trend and possible abrupt changes of the noise.

[0014] Preferably, a dynamic weight is introduced in S3 to fuse and generate the basic noise amplitude The specific content includes: A dynamic weight is introduced to balance the contributions of the two estimates of the short-time and long-time: ; ; wherein is the initial weight, is the time frame index, , is the frequency index of the time-frequency representation matrix, , , is the number of points of the time-frequency transform, is the imaginary unit, and are upper and lower bounds of the range, satisfying 0≤ ≤1; The base noise amplitude is: .

[0015] Preferably, in S4, the signal noise amplitude ratio is calculated based on The specific content of calculating the minimum amplitude threshold includes: The signal noise amplitude ratio is calculated based on the base noise amplitude , which is used to quantify the relative strength of the signal and noise: ; The minimum amplitude threshold is calculated based on : ; ; In the formula, is a scaling factor that controls the range of the minimum amplitude threshold, ensuring that the key features of the signal are not weakened.

[0016] Preferably, in S4, the estimated noise amplitude is calculated based on the dynamic noise scaling factor and The specific content includes: The dynamic noise scaling factor is introduced to dynamically adjust the noise amplitude: ; In the formula, is a scaling factor that controls the sensitivity of the adjustment factor, is the initial weight; The estimated noise amplitude is calculated based on the dynamic noise scaling factor : .

[0017] Preferably, the specific content of S5 includes: The denoised spectral amplitude is calculated according to : ; The original phase is preserved, and the complex spectrum after denoising is recovered: ; ; ​​The inverse transform is used to reconstruct the time-domain signal from the denoised spectrum. : ; In the formula, A windowing function for overlapping functions.

[0018] Preferably, it also includes: S6. Real-time monitoring of the noise reduction time-domain signal. Quality, generating performance evaluation reports.

[0019] An adaptive noise reduction filtering system, comprising: The preprocessing module is used to acquire dynamic signals from the signal receiving device. The corresponding time-domain signal is obtained after preprocessing. ,in The signal length; The amplitude spectrum generation module is used to obtain the amplitude spectrum through time-frequency transformation. The corresponding time-frequency representation matrix , used to indicate the first Frame, First The complex amplitude of the frequency point, and according to Calculate the amplitude spectrum ; The dual-timescale noise estimation module is used to analyze the amplitude spectrum along the time axis using a dual-timescale noise estimation method. Extracting short-time noise amplitude and long-term noise amplitude Introducing dynamic weights Will and Perform fusion to generate basic noise amplitude ; Denoising module, used for based Calculate the signal-to-noise ratio ,according to Calculate the minimum amplitude threshold And based on dynamic noise scaling factor as well as Calculate and estimate noise amplitude ; The noise reduction and signal reconstruction module is used to determine the signal based on the noise reduction and signal reconstruction module. and Calculate the noise reduction spectrum amplitude Preserving the original phase information, the inverse transform is used to... Reconstructed into a denoised time-domain signal .

[0020] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the adaptive noise reduction filtering methods described above.

[0021] An electronic device comprising: a processor and a memory storing one or more programs; when the one or more programs are executed by the processor, implementing any one of the adaptive noise reduction filtering methods above.

[0022] Via the technical solutions described above, compared with the prior art, the present disclosure provides an adaptive noise reduction filtering method, system, medium and device, which has the following beneficial effects: 1. Excellent noise suppression capability: by introducing a double-time-scale noise estimation, combined with dynamic fusion of short-time noise amplitude and long-time noise amplitude , the present disclosure accurately captures the dynamic characteristics of non-stationary noise, and further realizes efficient noise suppression through adaptive noise reduction adjustment of signal-to-noise amplitude ratio SNMR and dynamic noise scaling factor .

[0023] 2. Key feature protection: minimum amplitude threshold and final noise amplitude are also introduced in the present disclosure, which dynamically controls the noise reduction strength through signal-to-noise amplitude ratio SNMR, effectively avoids excessive signal suppression, and also enhances the relative strength of the feature band.

[0024] 3. Wide adaptability: a variety of time-frequency transform techniques (including but not limited to STFT, FFT, wavelet transform, etc.) and flexible window function and frame shift selection (including but not limited to Hanning window, fixed frame shift or adaptive frame shift) are also introduced in the present disclosure, combined with optimized multi-scale noise estimation and SNMR adaptive noise reduction adjustment algorithm with low computational complexity, which significantly improves the robustness to complex non-stationary noise environment. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can also be obtained by those skilled in the art without creative labor.

[0026] Figure 1 A flowchart of an adaptive noise reduction filtering method provided by the present disclosure; Figure 2 A time-frequency graph of an original urban environment sound signal provided by an embodiment of the present disclosure; Figure 3 A spectrum graph of an urban environment sound signal after noise reduction by the present method provided by an embodiment of the present disclosure; Figure 4The urban environment sound signal spectrum diagram after noise reduction by applying spectrum subtraction provided by the embodiment of the present application; Figure 5 The urban environment sound signal spectrum diagram after noise reduction by applying Wiener filtering provided by the embodiment of the present application; Figure 6 The urban environment sound signal spectrum diagram after noise reduction by applying wavelet provided by the embodiment of the present application; Figure 7 The urban environment sound signal spectrum diagram after noise reduction by applying wavelet packet provided by the embodiment of the present application; Figure 8 The hardware system structure schematic diagram of a hydroelectric generating set provided by the embodiment of the present application; Figure 9 The time-frequency diagram of the abnormal working condition sound of the hydroelectric generating set collected by the embodiment of the present application; Figure 10 The noise reduction result of the abnormal working condition sound provided by the embodiment of the present application; Figure 11 The sub-band energy proportion feature provided by the embodiment of the present application; wherein Figure 11 (a) Figure 11 (h) are the energy proportion change curves of the first to eighth sub-bands with time, respectively; Figure 12 The verification result schematic diagram of the CNN model provided by the embodiment of the present application; Figure 13 The hardware system schematic diagram of a wind turbine provided by the embodiment of the present application; Figure 14 The abnormal sound time-frequency diagram of the wind turbine collected by the embodiment of the present application; Figure 15 The noise reduction result schematic diagram of the method provided by the embodiment of the present application; Figure 16 The extracted MFCC feature schematic diagram provided by the embodiment of the present application; Figure 16 (a) Figure 16 (l) are the change curves of 12 MFCC features, respectively; Figure 17 The verification result schematic diagram of the time delay neural network TDNN model provided by the embodiment of the present application; Figure 18 The hardware system schematic diagram of a bridge monitoring provided by the embodiment of the present application; Figure 19 The bridge vibration signal collected by the embodiment of the present application; Figure 20 The noise reduction result schematic diagram of the method provided by the embodiment of the present application; Figure 21A schematic diagram of the extracted time-domain features provided in an embodiment of the present invention; Figure 21 (a)- Figure 21 (j) The curves showing the changes of 10 time-domain features are shown in turn; Figure 22 This is a schematic diagram illustrating the verification results of the ResNet residual network model provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This invention provides an adaptive noise reduction filtering method, such as... Figure 1 As shown, it includes the following steps: S1. Acquire dynamic signals from signal receiving equipment. The corresponding time-domain signal is obtained after preprocessing. ,in The signal length; S2. Obtained through time-frequency transformation The corresponding time-frequency representation matrix , used to indicate the first Frame, First The complex amplitude of the frequency point, and according to Calculate the amplitude spectrum ; S3. Employ a dual-time-scale noise estimation method to analyze the amplitude spectrum along the time axis. Extracting short-time noise amplitude and long-term noise amplitude Introducing dynamic weights Will and Perform fusion to generate basic noise amplitude ; S4. Based on Calculate the signal-to-noise ratio According to the signal-to-noise amplitude ratio Calculate the minimum amplitude threshold And based on dynamic noise scaling factor as well as Calculate and estimate noise amplitude ; S5. Based on and Calculate the noise reduction spectrum amplitude , the original phase information is reserved, and the inverse transform is used to reconstruct a denoised time domain signal . .

[0029] It should be noted that: The signal is preprocessed in the initial stage, and the DC offset is removed to effectively eliminate the DC component interference, so as to obtain a preprocessed time domain signal The DC offset removal can be achieved by various technical means, including but not limited to traditional mean subtraction, high-pass filtering method to suppress low-frequency DC component, and adaptive filtering technology.

[0030] In order to further implement the above technical solutions, the specific content of the preprocessing in S1 includes: Obtaining a dynamic signal from a signal receiving device , removing the DC offset preprocessing of , removing the low-frequency interference, and obtaining a preprocessed time domain signal .

[0031] In order to further implement the above technical solutions, the specific content of S2 includes: Obtaining a corresponding time-frequency representation matrix by time-frequency transformation , wherein the time-frequency transformation method specifically includes: short-time Fourier transform STFT, fast Fourier transform FFT, wavelet transform or wavelet packet transform; On the basis of the time-frequency representation matrix , the amplitude spectrum is calculated: .

[0032] In order to further implement the above technical solutions, the specific content of the time-frequency transformation by the short-time Fourier transform STFT method includes: ; ; ; In the formula, is the window length, indicating the number of sampling points of each time window, is the frame shift, indicates the time domain sample of the input dynamic signal in the frame; indicates the window function, indicates the total number of frames, indicates the floor function, is the frequency index of the time-frequency representation matrix, , , the number of points for time-frequency transform, imaginary unit; the index of time frame, the number of overlapping sampling points between adjacent frames, used to ensure the smooth reconstruction of the signal and avoid distortion caused by windowing.

[0033] It should be noted that: The power spectrum can be implemented by time domain to time-frequency domain transform technology, including but not limited to short-time Fourier transform STFT, fast Fourier transform FFT, wavelet transform and wavelet packet transform. In this embodiment, short-time Fourier transform STFT is taken as an example.

[0034] In order to further implement the above technical solutions, the amplitude spectrum along the time axis in S3 is analyzed , and the specific content of extracting short-time noise amplitude and long-time noise amplitude includes: Using double-time-scale noise estimation, by analyzing the statistical distribution of the amplitude spectrum along the time axis, the short-time noise amplitude and the long-time noise amplitude are calculated respectively: ; ; In the formula, and respectively represent the first quantile and the first quantile; and are the quantile thresholds of short-time and long-time noise estimation, which can be dynamically adjusted in value range to adapt to different noise environments; reflects the local background level of noise, captures the overall trend and possible sudden changes of noise.

[0035] In order to further implement the above technical solutions, a dynamic weight is introduced in S3 and to generate the basic noise amplitude . The specific content includes: introducing a dynamic weight to balance the contribution of short-time and long-time estimation: ; ; In the formula, is the initial weight, is the index of time frame, ,​ frequency index of the time-frequency representation matrix, , , number of points of the time-frequency transform, imaginary unit, and is upper and lower bounds of the range, satisfying 0≤ < ≤1; then the basic noise amplitude is: .

[0036] It should be noted that: The noise of the dynamic signal has the characteristics of short-time burstiness and long-time stationarity, and single-scale noise estimation is difficult to accurately capture its dynamic distribution. To solve this problem, the present application uses double-time scale noise estimation. and are short-time and long-time noise estimation thresholds, respectively, which can be calculated based on statistical methods, or other methods, including but not limited to mean plus multiple of standard deviation or K-means clustering method, to optimize the performance of noise estimation.

[0037] reflects the local background level of the noise, captures the overall trend and possible burstiness of the noise. To fuse the short-time and long-time noise characteristics, a dynamic weight is introduced to balance the contributions of the two estimates. The weight reflects the average intensity of the signal relative to the long-time noise, and is used to adjust the contribution ratio of the short-time and long-time estimates.

[0038] In order to further implement the above technical solutions, the signal noise amplitude ratio is calculated based on , and the minimum amplitude threshold is calculated according to the signal noise amplitude ratio The specific content includes: The signal noise amplitude ratio is calculated based on the basic noise amplitude , which is used to quantify the relative intensity of the signal and the noise: ; The minimum amplitude threshold is calculated based on the signal noise amplitude ratio : ; In the formula, is a scaling factor that controls the range of the minimum amplitude threshold, ensuring that the key features of the signal are not weakened.

[0039] To further implement the above technical solutions, S4 is based on the dynamic noise scaling factor And The specific content of calculating the estimated noise amplitude includes: Introducing a dynamic noise scaling factor , to dynamically adjust the noise amplitude: ; Among them, is a scaling factor, which controls the sensitivity of the adjustment factor, is the initial weight, calculated according to the formula above ; Based on the dynamic noise scaling factor , the estimated noise amplitude is calculated: .

[0040] It should be noted that: In order to effectively denoise, the present application calculates the signal-to-noise amplitude ratio SNMR based on the basic noise amplitude, which is used to quantify the relative strength of the signal and the noise. In order to avoid excessive suppression of the signal during denoising, a minimum amplitude threshold is further calculated based on SNMR.

[0041] To further implement the above technical solutions, the specific content of S5 includes: According to and , the noise reduction spectrum amplitude is calculated: ; Preserve the original phase and restore the complex spectrum after noise reduction : ; Use inverse transform to reconstruct the time domain signal from the noise reduced spectrum : ; In the formula, is an overlap window function.

[0042] To further implement the above technical solutions, it also includes: S6. By monitoring the quality of the noise reduction time domain signal in real time, a performance evaluation report is generated.

[0043] An adaptive noise reduction filtering system includes: A preprocessing module for obtaining a dynamic signal from a signal receiving device, and obtaining a corresponding time domain signal after preprocessing wherein is the signal length; an amplitude spectrum generation module configured to obtain an amplitude spectrum by time-frequency transform corresponding time-frequency representation matrix for representing complex amplitudes of the i-th frame, the j-th frequency bin, and calculating an amplitude spectrum ; ; ; ; a dual-time-scale noise estimation module configured to analyze the amplitude spectrum along the time axis by using a dual-time-scale noise estimation method, extract short-time noise amplitudes and long-time noise amplitudes , introduce a dynamic weight , and fuse and to generate a basic noise amplitude ; ; a denoising module configured to calculate a signal-to-noise amplitude ratio based on , calculate a minimum amplitude threshold according to the signal-to-noise amplitude ratio , and calculate an estimated noise amplitude based on a dynamic noise scaling factor and ; a noise reduction and signal reconstruction module configured to calculate a noise reduction spectrum amplitude based on and , retain original phase information, and reconstruct the noise reduction time-domain signal from by inverse transform.

[0044] A computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements any of the above adaptive noise reduction filtering methods.

[0045] An electronic device, comprising: a processor and a memory for storing one or more programs; when the one or more programs are executed by the processor, any of the above adaptive noise reduction filtering methods is implemented.

[0046] The application will be further described below through specific experiments: The experiment is based on the actual collection of urban environmental sound signal data (only environmental sound signals are used as examples in this embodiment, and in actual application, adaptive filtering can be performed on dynamic modulation signals, such as original acoustic and vibration signals during operation of a hydroelectric generator set, original sound signals during operation of a wind turbine generator set, and vibration displacement original signals during operation of a bridge), the sampling rate is 32 kHz, and the signal time domain range is 0 to 1.8 seconds. In this experiment, the main parameter configuration is: the window length is set to 256 sampling points, the overlap length is set to 128 sampling points, the FFT point number is set to 512, the short-time noise amplitude and the long-time noise amplitude are set to the 25th and 75th percentiles, respectively, 0.5 is set as the lower limit of weight calculation, 0.9 is set as the upper limit of weight calculation, 0.005 is set as the base value of the minimum amplitude threshold, 1.5 is set as the dynamic noise scaling factor parameter.

[0047] As shown in Figures 2-7 , the first 1.1 seconds of the signal is the characteristic frequency band, and 1.1 seconds to 1.8 seconds is the noise frequency band, wherein Figure 2 represents the original spectrum diagram of the signal, Figure 3 is the signal spectrum diagram after applying the noise reduction method, Figure 4 is the signal spectrum diagram after applying the spectral subtraction noise reduction method, the spectral subtraction uses a reduction factor =1.0, a floor factor =0.1, and noise estimation is based on the 20th percentile of the amplitude spectrum to moderately suppress noise and reduce artifacts; Figure 5 is the signal spectrum diagram after applying the Wiener filter noise reduction, the filter kernel size of the Wiener filter is 5, and the balance between noise suppression and signal fidelity is achieved; Figure 6 is the signal spectrum diagram after applying the wavelet noise reduction, the wavelet noise reduction uses db4 wavelet, the decomposition level is 4, the soft threshold method is used, and the threshold is calculated based on the noise variance to achieve adaptive noise reduction and preserve signal details, Figure 7 is the signal spectrum diagram after applying the wavelet packet noise reduction, the wavelet packet noise reduction uses db4 wavelet, the decomposition level is 4, the soft threshold method is used to reduce artifacts, and the effectiveness of the noise reduction method is evaluated.

[0048] The sampling rate of the dynamic signal is 32000 Hz, and the total duration is about 1.8 seconds (57498 sampling points in total). After normalization and removal of DC offset, the time-frequency representation matrix is generated, and the amplitude spectrum is further calculated. The amplitude spectrum size is (257, 450), corresponding to 257 frequency components and 450 time frames, reflecting the time-frequency characteristics of about 1.8 seconds of signal. The amplitude spectrum statistical values are shown in Table 1.

[0049]

[0050] By analyzing the 1.8-second amplitude spectrum, the short-time and long-time noise amplitudes are calculated using a multi-scale method. The short-time noise and long-time noise matrices are generated based on the 25% and 75% quantiles, respectively, with a matrix size of (257, 1). Then, based on the formula , the is calculated as 0.1480, and the output is shown in Tables 2 and 3.

[0051]

[0052]

[0053] Using the basic noise matrix and the amplitude spectrum , the SNMR is calculated, the minimum amplitude threshold is calculated, the dynamic noise scaling factor is calculated, and the final noise amplitude is calculated. The output is shown in Tables 4 and 5.

[0054]

[0055]

[0056] Using the final noise amplitude, the final amplitude spectrum matrix is calculated, the complex spectrum matrix is reconstructed, and the denoised signal is reconstructed. The output is shown in Table 6.

[0057]

[0058] Based on the spectral statistical values and the generalized index values, a comprehensive comparative analysis of the six methods (no denoising, this method, spectral subtraction, Wiener filtering, wavelet denoising, and wavelet packet denoising) is performed.

[0059] 1. Comparison of time-frequency matrix statistical indicators after denoising As Figures 2-7As shown, these spectrograms collectively demonstrate the signal's distribution characteristics in time domain: the first 1.1 seconds correspond to the characteristic band, mainly containing the periodic or modulated components of the target signal, usually exhibiting higher energy concentration, spectral line clarity and structural features (mainly in green to yellow high-amplitude areas); 1.1 seconds to about 1.8 seconds is the noise band, presenting as randomly distributed broadband noise, with dispersed energy and lack of obvious patterns (mainly in blue to purple low-amplitude areas).

[0060] Specifically, Figure 2 The original spectrogram of the signal directly reflects the time-frequency characteristics of the unprocessed signal, serving as a benchmark for subsequent method evaluation, with the characteristic band having prominent energy, while the noise band, although lower, still has obvious interference. Figure 3 The signal spectrogram after applying the method, with the noise band amplitude significantly reduced (color tending towards deep purple, close to -100 dB), while the characteristic band maintains high fidelity and energy concentration, with sharper spectral lines. Figures 4-7 The signal spectrograms after applying different comparison methods (such as spectral subtraction, Wiener filtering, wavelet denoising, and wavelet packet denoising): as shown, these comparison methods can partially suppress noise (noise area color darkens), but have more residual interference, lower characteristic band fidelity, and overall amplitude decline is not as complete as the method; in contrast, the method demonstrates stronger noise suppression ability and better feature protection and enhancement, achieving more uniform low-amplitude distribution in the noise band, while enhancing the relative strength and clarity of the characteristic band.

[0061] Table 7 visually compares the final output spectrum statistics, the first 1.1 seconds spectrum statistics, and the spectrum statistics beyond 1.1 seconds of the six methods, focusing on the mean and maximum.

[0062]

[0063] 1) The method compared with no denoising Excellent noise suppression ability: the method significantly outperforms no denoising in noise suppression. The final output spectrum statistics mean value decreases from -20.0055 dB of no denoising to -75.4310 dB, with a decrease of about 55 dB, indicating strong noise suppression. The first 1.1 seconds mean value decreases from -9.4297 dB to -59.8047 dB (decrease of about 50 dB), and the 1.1 seconds beyond mean value decreases from -36.6246 dB to -99.9867 dB (decrease of about 63 dB). This significant noise reduction is due to the double-time-scale noise estimation and the adaptive adjustment of SNMR and dynamic noise scaling factor in S3, accurately capturing and eliminating non-stationary noise.

[0064] Relative intensity enhancement of key features: Despite the strong noise suppression, the maximum value of the first 1.1 seconds of the method drops slightly from 19.4851 dB without noise reduction to 18.9729 dB (only about 0.5 dB lower), retaining the peak intensity close to the original signal. The maximum value beyond the last 1.1 seconds drops from -5.4981 dB to -39.5729 dB, but this is consistent with the target frequency band focusing on the first 1.1 seconds, and the minimum amplitude threshold and dynamic adjustment in S3 ensure the relative intensity of the key features is enhanced.

[0065] 2) Comparison of the method with other noise reduction methods Stronger noise suppression capability: Compared with other noise reduction methods (spectral subtraction -22.2795 dB, Wiener filtering -24.4751 dB, wavelet denoising -27.6328 dB, wavelet packet denoising -43.2331 dB), the final output spectrum statistical mean of the method is -75.4310 dB, which is the lowest, and the suppression intensity is about 32 dB -53 dB higher. The mean of the first 1.1 seconds -59.8047 dB is about 47 dB -50 dB lower than other methods (-9.6612 dB to -12.4225 dB), and the mean beyond the last 1.1 seconds -99.9867 dB is about 46 dB -58 dB lower than other methods (-42.1082 dB to -53.8296 dB), showing that the method has more complete suppression of noise in the entire 1.8 second signal.

[0066] Better feature protection and enhancement: The maximum value of the first 1.1 seconds of the method is 18.9729 dB, close to spectral subtraction (19.4788 dB), Wiener filtering (19.5175 dB), wavelet denoising (19.5015 dB), and wavelet packet denoising (19.5027 dB), but its mean -59.8047 dB is much lower than these methods (-9.6612 dB to -12.4225 dB), indicating that a higher relative peak intensity is retained under strong noise reduction. Combined with the concentration of features in the first 1.1 seconds, the dynamic adjustment of S3 effectively enhances the relative intensity of the feature frequency band.

[0067] 2. Comparison and analysis combined with generalized indicators Table 8 provides index data (SNR after noise reduction, spectral contrast, HNR, SDR, FSIM, GMSD) for comparative analysis of the six methods. Among them, SNR, Signal-to-Noise Ratio: signal-to-noise ratio, estimated by MCRA2 algorithm, the higher the value, the better the signal quality; Spectral Contrast: spectral contrast, based on the average of the peak and valley difference of the frequency band, the higher the value, the more prominent the separation degree of harmonics and noise; HNR, Harmonics-to-Noise Ratio: harmonic-to-noise power ratio, the higher the value, the more dominant the harmonic component; SDR, Source-to-Distortion Ratio: source-to-distortion ratio, the higher the value, the smaller the distortion; FSIM, Feature Similarity Index: feature similarity, range [0, 1], the closer to 1, the more similar to the original signal characteristics; GMSD, Gradient Magnitude Similarity Deviation: gradient similarity deviation, the smaller the value, the better the quality of the image matrix.

[0068]

[0069] 1) The method compared with no noise reduction Excellent noise suppression: The method improves SNR from 10.61 without noise reduction to 11.53 (about 0.92 dB increase) through Step 2's double-time-scale noise estimation and S3's SNMR dynamic adjustment, indicating improved signal-to-noise ratio (noise reduction estimated by MCRA2). Spectral Contrast increases significantly from 12.96 to 36.56 (about 23.6 increase), reflecting enhanced spectral feature separation (peak-valley difference increase). SDR increases from 0 to 9.83, showing significant reduction in noise power, verifying the superiority of the method in noise suppression.

[0070] Relative intensity enhancement of key features: After noise reduction, the method increases Spectral Contrast to 36.56 (vs 12.96), indicating prominent target frequency band features (segmented contrast optimization). GMSD decreases from 1 to 1.05, close to 1, indicating stable quality. SDR 9.83 indicates better signal reconstruction quality than no noise reduction, and the relative intensity of features is enhanced.

[0071] 2) The method compared with other noise reduction methods Better noise suppression: The SNR (11.53) of the phase-based method is higher than that of spectral subtraction (10.59), wavelet denoising (10.73), and wavelet packet denoising (10.72), and is only lower than that of Wiener filtering (11.16), but the Spectral Contrast (36.56) is much higher than that of other methods (8.73-19.54), indicating stronger spectral feature separation (maximum-minimum difference). The SDR (9.83) is lower than that of spectral subtraction (21.4), Wiener filtering (23.24), wavelet denoising (35.36), and wavelet packet denoising (34.19), but reflects reasonable distortion control, proving the unique advantage of the method in noise suppression.

[0072] Better feature protection and enhancement: The Spectral Contrast (36.56) of the method is much higher than that of other methods, indicating significant target band enhancement (segment contrast optimization). The FSIM (0.17) is lower than that of other methods (0.35-0.56), but the GMSD (1.05) is close to that of other methods (1.01-1.03), indicating controllable feature distortion. Although the SDR (9.83) is lower than that of other methods, the peak preservation (18.9729 dB) in Table 7 proves that the relative intensity of the feature is enhanced under strong noise suppression, which is better than the balance of other methods.

[0073] The application will be further described below through specific application examples.

[0074] Example 1 As shown in Figure 8 , the multi-channel distributed high-resolution sound and vibration signal acquisition scheme of the hydroelectric generator set mainly includes audio sensors, vibration sensors, signal acquisition cards, and host computers. Analog signals are collected through sensors, and then digital signals are output through conditioning circuits and acquisition cards, and transmitted to a server for processing and analysis using Ethernet.

[0075] In this embodiment, the main parameter configuration is as follows: the window length is set to 256 sampling points, the overlap length is set to 128 sampling points, the FFT point number is set to 512, the short-time noise amplitude and the long-time noise amplitude are set to 25% and 75%, respectively, 0.5 is set as the lower limit of weight calculation, 0.9 is set as the upper limit of weight calculation, 0.005 is set as the base value of the minimum amplitude threshold, 1.5 is set as the dynamic noise scaling factor parameter.

[0076] Figure 9The frequency spectrum analysis of the collected abnormal working condition sound of the hydroelectric generating set is shown. The horizontal axis represents time (unit: seconds), ranging from 0 to 10 seconds; the vertical axis represents frequency (unit: Hz), ranging from 0 to 16000 Hz. The amplitude (unit: dB) gradually changes from a higher amplitude (about -10 dB) to a lower amplitude (about -40 dB) to show the intensity distribution of the sound signal. This abnormal sound is caused by the entry of foreign matter into the interior of the hydraulic turbine, resulting in the appearance of abnormal frequency bands and sound characteristics, among which the abnormal frequency bands are marked in the figure, and are in sharp contrast with the background noise.

[0077] The sampling rate of the acoustic and vibration signal is 32000 Hz, and the total duration is about 10 seconds (320000 sampling points in total). After normalization and removal of the direct current offset, the time-frequency representation matrix is generated, and the amplitude spectrum is further calculated, and the amplitude spectrum size is (257, 2501), corresponding to 257 frequency components and 2501 time frames, reflecting the time-frequency characteristics of about 10 seconds of signal, wherein the amplitude spectrum statistical value is shown in Table 9.

[0078]

[0079] By analyzing the 10-second amplitude spectrum, the short-time and long-time noise amplitudes are calculated by using the multi-scale method, and the short-time noise and long-time noise matrices are generated based on the 25% and 75% quantiles respectively, and the matrix size is (257, 1). Then, based on the formula , the is calculated as 0.1374, and the output is shown in Tables 10 and 11.

[0080]

[0081]

[0082] The basic noise matrix and the amplitude spectrum are used to calculate the signal noise amplitude ratio SNMR, calculate the minimum amplitude threshold, calculate the dynamic noise scaling factor, and calculate the final noise amplitude, and the output is shown in Tables 12 and 13.

[0083]

[0084]

[0085] Using the final noise amplitude, the final amplitude spectrum matrix is calculated, the complex spectrum matrix is reconstructed, and the denoised signal is reconstructed, and the output is shown in Table 14.

[0086]

[0087] Figure 10The audio spectrum analysis after applying this noise reduction method is shown. The horizontal axis represents time (in seconds), ranging from 0 to 10 seconds; the vertical axis represents frequency (in Hz), ranging from 0 to 16000 Hz. Amplitude range and Figure 9 Consistent. Through with Figure 9 As can be seen from the comparison, the method of the present invention significantly suppresses the background noise frequency band (manifested as an increase in large areas of low amplitude in the figure), while effectively preserving the abnormal characteristic frequency band caused by foreign objects inside the turbine (manifested as a more prominent high amplitude area). This indicates that the noise reduction method effectively maintains the relative intensity of key abnormal features while removing interference noise.

[0088] like Figure 11 As shown, by performing a three-level decomposition of the reconstructed signal using the db3 wavelet, the energy proportion characteristics of each sub-band were extracted. Figure 11 (a) ~ Figure 11 (h) are the energy percentage curves of the first to eighth sub-bands over time, respectively. These time-series plots show the energy percentage curves of the eight sub-bands over time. Among them, the high-frequency sub-bands show obvious peak fluctuations, reflecting the periodic characteristics of abnormal operating condition sound signals, such as transient energy concentration caused by foreign object impact.

[0089] To detect abnormal operating conditions in the acoustic and vibration signals of hydropower units, a two-dimensional convolutional neural network (CNN) model is constructed. The energy features extracted by wavelet packet decomposition are used as input for abnormal operating condition identification. The model architecture is as follows: Input layer: Receives wavelet packet energy features, and the data is formatted as a two-dimensional tensor (with 1 channel). The feature extraction module includes: a first convolutional layer with 16 kernels, stride 1, padding 1, batch normalization, and ReLU activation; a first max pooling layer with a window and stride 2; a second convolutional layer with 32 kernels, stride 1, padding 1, batch normalization, and ReLU activation; a second max pooling layer with a window and stride 2; a third convolutional layer with 64 kernels, stride 1, padding 1, batch normalization, and ReLU activation; and an adaptive average pooling layer with a fixed output size of 2.

[0090] The classifier module includes: Fully connected layer 1: input 64×2×2, output 64, ReLU activation, Dropout (probability 0.5); Fully connected layer 2: output 2 neurons (normal / abnormal), no explicit Softmax (handled by loss function); Loss function: Cross-entropy loss, combined with the Softmax function to calculate classification probabilities. The model is implemented in the PyTorch framework, runs on the CPU, and is compact and efficient, suitable for real-time operational monitoring.

[0091] In this embodiment, to verify the effectiveness of the CNN model in the abnormal condition detection of hydroelectric generating units, 1730 data samples are used. Among them, 1300 are normal samples and 430 are abnormal sound samples. These samples are based on the extracted wavelet packet energy features as input. The dataset is divided into training set (1384 samples), validation set (173 samples) and test set (173 samples) in the ratio of 8:1:1 to ensure the model generalization ability. The optimizer is Adam and the initial learning rate is 0.001.

[0092] The model test set accuracy reaches 98.84%, showing high efficiency in distinguishing normal and abnormal conditions. Lower false positive and false negative further verify the robustness of the model. Batch normalization and Dropout effectively prevent overfitting, and learning rate scheduling optimizes the convergence speed.

[0093] The model realizes accurate abnormal detection of hydroelectric generating unit sound and vibration signals through a two-dimensional CNN architecture combined with wavelet packet energy features. Its high accuracy (98.84%) and stable classification performance prove its reliability in complex condition monitoring, providing an efficient solution for intelligent diagnosis of hydroelectric generating units.

[0094] Example Two: As shown in Figure 13 , the multi-channel distributed high-resolution acoustic signal acquisition and analysis scheme of the wind turbine mainly includes audio sensors, signal acquisition cards, and host computers. Analog signals are collected by sensors, then digital signals are output through conditioning circuits and acquisition cards, and are transmitted to the host computer in the booster station for processing and analysis. In this embodiment, the main parameter configurations are: the window length is set to 256 sampling points, the overlap length is set to 128 sampling points, the FFT point number is set to 512, the short-time noise amplitude and the long-time noise amplitude are set to 25% and 75% respectively, 0.5 is set as the lower limit of weight calculation, 0.9 is set as the upper limit of weight calculation, 0.005 is set as the base value of the minimum amplitude threshold, 1.5 is set as the dynamic noise scaling factor parameter.

[0095] Figure 14The frequency spectrum analysis of the collected abnormal sound of the wind turbine is shown. The horizontal axis represents time (unit: second), ranging from 0 to 10 seconds; the vertical axis represents frequency (unit: Hz), ranging from 0 to 16000 Hz. The amplitude (unit: dB) gradually changes from a higher amplitude (about -10 dB) to a lower amplitude (about -40 dB) to show the intensity distribution of the sound signal. This abnormal sound is caused by the abnormality of the blade bearing of the unit, resulting in abnormal frequency bands and sound characteristics, in which the abnormal frequency bands are marked in the figure, and are in sharp contrast with the background noise.

[0096] The sampling rate of the dynamic signal is 32000 Hz, and the total duration is about 10 seconds (320000 sampling points in total). After the signal is normalized and the DC offset is removed, a time-frequency representation matrix is generated, and the amplitude spectrum is further calculated, resulting in an amplitude spectrum size of (257, 2501), corresponding to 257 frequency components and 2501 time frames, reflecting the time-frequency characteristics of about 10 seconds of signal, wherein the amplitude spectrum statistical values are shown in Table 15.

[0097]

[0098] By analyzing the 10-second amplitude spectrum, the short-time and long-time noise amplitudes are calculated using a multi-scale method, and the short-time noise and long-time noise matrices are generated based on the 25% and 75% quantiles, respectively, with a matrix size of (257, 1). Then, based on the formula , the is calculated as 0.1620, and the output is shown in Tables 16 and 17.

[0099]

[0100]

[0101] The basic noise matrix and the amplitude spectrum are used to calculate the signal noise amplitude ratio SNMR, calculate the minimum amplitude threshold, calculate the dynamic noise scaling factor, and calculate the final noise amplitude, and the output is shown in Tables 18 and 19.

[0102]

[0103]

[0104] Using the final noise amplitude, the final amplitude spectrum matrix is calculated, the complex spectrum matrix is reconstructed, and the denoised signal is reconstructed, and the output is shown in Table 20.

[0105]

[0106] Figure 15The audio spectrum analysis after applying this noise reduction method is shown. The horizontal axis represents time (in seconds), ranging from 0 to 10 seconds; the vertical axis represents frequency (in Hz), ranging from 0 to 16000 Hz. The amplitude (in dB) range is... Figure 14 Consistent. Through with Figure 14 As can be seen from the comparison, the method of the present invention significantly suppresses the background noise frequency band (manifested as an increase in large areas of low amplitude in the figure), while effectively preserving the abnormal characteristic frequency band caused by the blade bearing (manifested as a more prominent high amplitude area). This indicates that the noise reduction method effectively maintains the relative intensity of key abnormal features while removing interference noise.

[0107] like Figure 16 As shown, MFCC features are extracted from the denoised signal. These time-series plots illustrate the changes in 12 MFCC features.

[0108] To achieve abnormal state detection of acoustic signals from wind turbine generators, this method constructs a Time Delay Neural Network (TDNN) model, using MFCC features as input, which is suitable for pattern recognition of non-stationary time-series signals. This model introduces a time delay mechanism to efficiently capture dynamic context information, and its architecture is as follows: Input layer: Receives MFCC feature sequences and transposes them to adapt to one-dimensional convolution operations.

[0109] The temporal feature extraction module includes: TDNN layer 1: a one-dimensional convolution (kernel size 5, padding 2), with input channels equal to the feature dimension and output channels of 128; followed by batch normalization, ReLU activation, and Dropout (probability 0.5). TDNN layer 2: a one-dimensional convolution (kernel size 5, padding 2), with input channels of 128 and output channels of 128; followed by batch normalization and ReLU activation. Adaptive pooling layer: compresses the temporal dimension to a fixed length of 1.

[0110] The classifier module includes: a fully connected layer with 128-dimensional input and two-class output (normal / abnormal), where the classification probability is implicitly handled by the cross-entropy loss function.

[0111] The Time Delay Neural Network (TDNN) model is implemented in Python and runs on the CPU, making it suitable for real-time anomaly monitoring. In this embodiment, to verify the effectiveness of the TDNN model, acoustic data samples from 2450 wind turbine generators (1800 normal samples and 650 abnormal samples) were used as input, based on MFCC features. The dataset was divided into a training set (1960 samples), a validation set (245 samples), and a test set (245 samples) in an 8:1:1 ratio to ensure the model's generalization ability. The optimizer used was Adam with an initial learning rate of 0.001, and weight decay and learning rate step scheduling were introduced (decaying to 0.9 times every 5 cycles).

[0112] As Figure 17 shown, the model achieved an accuracy of 97.96% on the test set, demonstrating high discrimination ability between normal and abnormal states. The low false positive and false negative rates further verify the robustness of the model. Batch normalization and Dropout effectively prevent overfitting, and learning rate scheduling optimizes the convergence speed.

[0113] This model, through a one-dimensional TDNN architecture combined with MFCC features, achieves accurate anomaly detection of wind turbine acoustic signals. Its high accuracy (97.96%) and stable classification performance prove its reliability in complex working condition monitoring, providing an efficient solution for intelligent operation and maintenance of wind turbines.

[0114] Example Three: The multi-channel distributed vibration displacement signal acquisition scheme for bridge monitoring mainly includes displacement sensors, vibration sensors, data acquisition and processing terminals, etc. The analog signal is collected by the sensor, then the digital signal is output through the conditioning circuit and acquisition card, and transmitted to the data acquisition and processing terminal for processing and analysis. After that, wireless communication is used to alarm the abnormality.

[0115] In this embodiment, the main parameter configuration is: window length is set to 256 sampling points, the overlap length is set to 128 sampling points, the FFT point number is set to 512, the short-time noise amplitude and the long-time noise amplitude are set to 25% and 75% respectively, 0.5 is set as the lower limit of weight calculation, 0.9 is set as the upper limit of weight calculation, 0.005 is set as the base value of the minimum amplitude threshold, 1.5 is set as the dynamic noise scaling factor parameter.

[0116] Figure 19 The bridge vibration signal is shown, with the time axis from 0 to 5 seconds and the amplitude range from about -4 to 4. It is installed on an actual bridge structure to record the vibration response induced by vehicle passing or wind load. The signal as a whole shows high-frequency random oscillation, forming dense band-shaped fluctuations, and noise components dominate. This noise causes the signal details to be blurred, masking the inherent dynamic characteristics of the bridge and making it difficult to identify clear periodic patterns. Upon closer observation, the signal is relatively stable in the first 2 seconds, with small amplitude fluctuations. At about 2 seconds, the amplitude increases, with alternating peaks and valleys, indicating the starting time of external impact or load, such as vehicle passing or wind load-induced. Corresponding to the free vibration response of the bridge, potential damping oscillation patterns are hidden therein.

[0117] The sampling rate of the dynamic signal is 32000 Hz, and the total duration is about 5 seconds (total 160000 sampling points). After normalization and removal of DC offset, the time-frequency representation matrix is generated, and the amplitude spectrum is further calculated, and the amplitude spectrum size is (257, 1251), corresponding to 257 frequency components and 1251 time frames, reflecting the time-frequency characteristics of about 5 seconds of signal, wherein the amplitude spectrum statistical value is shown in Table 21.

[0118]

[0119] By analyzing the 1.8 second amplitude spectrum, the short-time and long-time noise amplitudes are calculated by using the multi-scale method, and the short-time noise and long-time noise matrices are generated based on the 25% and 75% quantiles respectively, and the matrix size is (257, 1). Then, based on the formula , the is 0.1504, and the output is shown in Table 22 and Table 23.

[0120]

[0121]

[0122] The basic noise matrix and the amplitude spectrum are used to calculate the signal noise amplitude ratio SNMR, calculate the minimum amplitude threshold, calculate the dynamic noise scaling factor, and calculate the final noise amplitude, and the output is shown in Table 24 and Table 25.

[0123]

[0124]

[0125] Using the final noise amplitude, the final amplitude spectrum matrix is calculated, the complex spectrum matrix is reconstructed, and the denoised signal is reconstructed, and the output is shown in Table 26.

[0126]

[0127] Figure 20The denoised bridge vibration signal is shown, with a time axis from 0 to 5 seconds and an amplitude range of about -1 to 1, significantly better than the -4 to 4 range of the pre-denoising signal, indicating that the denoising effectively reduces the noise amplitude. The signal still shows high-frequency random oscillation, but the band fluctuation is less sparse than before denoising, the noise influence is significantly weakened, and a relatively clear periodic pattern is revealed, reflecting the inherent dynamic characteristics of the bridge. Observing the details, the signal is relatively stable in the first 2 seconds, with small amplitude fluctuations and low noise baseline, reflecting the suppression of environmental interference after denoising. At about 2 seconds, the amplitude increases significantly, with regular peak-valley alternation, indicating the start of external impact (such as vehicle load), which is consistent with the pre-denoising, but the waveform is smoother and the damped oscillation pattern is more easily identified. This shows that the denoising method not only removes the interference noise, but also better maintains the key feature information of the original signal.

[0128] As shown in Figure 21 , 10 kinds of time domain features are extracted from the denoised bridge vibration signal, and these time series diagrams show the change trend of mean, root mean square, standard deviation, skewness, kurtosis, peak value, wave crest factor, shape factor, pulse factor and gap factor. Extract a set of features every 0.1 seconds and draw it as a 2x5 subgraph to intuitively present the change of features over time. In the feature curve, the mean and root mean square values show obvious peak fluctuations, reflecting the periodic response caused by impact load (such as vehicle passing) in the bridge vibration signal.

[0129] To verify the effectiveness of the ResNet model, a total of 2160 bridge vibration displacement data samples (1600 normal samples and 560 abnormal samples) are used as input (feature dimension 10) based on 10 kinds of time domain features. The dataset is divided into training set (1728 samples), validation set (216 samples) and test set (216 samples) in the ratio of 8:1:1 to ensure the model's generalization ability. The model is implemented under the PyTorch framework and runs on CPU.

[0130] The ResNet model uses a one-dimensional convolutional architecture to process the time domain feature sequence of the bridge vibration displacement signal. The input layer receives a 10-dimensional feature vector and expands it to 64 channels through an initial convolutional layer. The model contains multiple residual blocks (layer1 and layer2), each composed of two convolutional layers, batch normalization and ReLU activation. The residual connection ensures smooth information transmission and avoids deep network degradation. The adaptive average pooling layer compresses the feature dimension, and the fully connected layer outputs a 2-class probability (normal / abnormal). The Dropout layer (probability 0.5) prevents overfitting. This design fully utilizes residual learning to improve the generalization ability of non-stationary signals and supports accurate identification of bridge structure anomalies.

[0131] As shown in Figure 21As shown, the accuracy of the model on the test set reached 96.76%, showing high efficiency in distinguishing normal and abnormal structural states. The low false positive and false negative further verified the robustness of the model. The residual connection and batch normalization effectively prevent overfitting, and the learning rate scheduling effectively optimizes the convergence speed. The model realizes precise abnormal monitoring of bridge building vibration displacement signals through the ResNet architecture combined with time domain features. Its high accuracy (96.76%) and stable classification performance prove its reliability in structural health monitoring, providing an efficient solution for intelligent state evaluation of bridge buildings.

[0132] The above examples are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method of adaptive noise reduction filtering, characterized by, The method comprises the following steps: S1. obtaining a dynamic signal from a signal receiving device , obtaining a corresponding time domain signal through preprocessing , wherein is the signal length; S2. Obtain by time-frequency transform Corresponding time-frequency representation matrix , for representing the complex amplitude of the first frame, the first frequency point, and calculating the amplitude spectrum according to ; S3. Adopting double-time-scale noise estimation method, analyzing amplitude spectrum along time axis , extracting short-time noise amplitude and long-time noise amplitude , introducing dynamic weight , fusing and to generate basic noise amplitude ; S4. based on computing a signal-to-noise amplitude ratio , according to computing a minimum amplitude threshold , and based on a dynamic noise scaling factor and computing an estimated noise amplitude ; S5. In accordance with and computing the denoised spectral amplitudes , preserving the original phase information, by inverse transforming to a denoised time-domain signal .

2. The adaptive noise reduction filtering method of claim 1, wherein, The specific content of the preprocessing in S1 comprises: Acquire dynamic signals from signal receiving devices. ,right DC offset removal preprocessing is performed to remove low-frequency interference, resulting in the preprocessed time-domain signal. .

3. The adaptive noise reduction filtering method of claim 1, wherein, The specific content of S2 comprises: obtained by time-frequency transformation a corresponding time-frequency representation matrix wherein the method by time-frequency transformation specifically comprises a short-time Fourier transform (STFT), a fast Fourier transform (FFT), a wavelet transform or a wavelet packet transform. On the basis of the time-frequency representation matrix the amplitude spectrum is calculated: 。 4. The adaptive noise reduction filtering method of claim 2, wherein, The specific content of the time-frequency transformation by the short-time Fourier transform (STFT) method comprises: ； ； ； wherein is the window length, indicating the number of sample points per time window, is the frame shift, indicates the time domain samples of the input dynamic signal in the frame; indicates the window function, indicates the total number of frames, indicates the floor function, is the frequency index of the time-frequency representation matrix, , , is the number of points of the time-frequency transform, is the imaginary unit; is the time frame index, ; is the number of overlapping sample points between adjacent frames.

5. The adaptive noise reduction filtering method of claim 1, wherein, S3 analyzing the magnitude spectrum along the time axis extracting short-time noise magnitudes and long-time noise magnitudes The details include: Using a double-time-scale noise estimation, the short-time noise amplitude and the long-time noise amplitude are calculated by analyzing the amplitude spectrum along the time axis, respectively. ； ； wherein and denote the first decile and the first decile, respectively; and are the decile thresholds for the short-time and long-time noise estimates, respectively; reflects the local background level of the noise, captures the overall trend and possible abrupt changes in the noise.

6. The method of adaptive noise reduction filtering according to claim 5, wherein, Introducing dynamic weights in S3 Fusing and to generate the base noise amplitude The specific content includes: Introducing dynamic weights to balance the contribution of both short and long estimates: ； ； wherein is an initial weight, is a time frame index, , is a frequency index of the time-frequency representation matrix, , , is a number of points of the time-frequency transform, is an imaginary unit, and is upper and lower bounds of a range, satisfying 0≤ < 1 ≤ 1; the base noise amplitude is: 。 7. The method of claim 1, wherein, In S4, based on Computing signal noise amplitude ratio , according to Computing minimum amplitude threshold The specific content includes: Based on the base noise amplitude Computing the signal noise amplitude ratio for quantifying the relative strength of signal and noise: ； Based on Computing a minimum amplitude threshold : ； wherein is a scaling factor that controls the range of the minimum amplitude threshold, ensuring that the signal's key features are not attenuated.

8. The method of adaptive noise reduction filtering according to claim 7, wherein, S4 based on dynamic noise scaling factor and calculating an estimated noise amplitude The details include: Introducing dynamic noise scaling factor to dynamically adjust the noise amplitude: ； wherein, is a scaling factor, controlling the sensitivity of the adjustment factor, is an initial weight; Dynamic noise scaling factor Computing an estimated noise amplitude : 。 9. The method of claim 1, wherein, The specific content of S5 comprises: In accordance with and Computing the noise reduction spectrum magnitude : ； Preserving original phase, restoring complex spectrum after noise reduction : ； reconstructing the noise-reduced spectrum into a time-domain signal using an inverse transform : ； In the formula, is a superimposed window function.

10. The method of adaptive noise reduction filtering according to claim 1, wherein, Further comprising: S6. Monitoring the noise-reduced time-domain signal in real time Quality, generating a performance evaluation report.

11. An adaptive noise reduction filter system based on the adaptive noise reduction filter method of any one of claims 1-10, characterized in that, Comprising: a pre-processing module for obtaining a dynamic signal from a signal receiving device , a corresponding time domain signal is obtained after pre-processing , wherein is the signal length; An amplitude spectrum generation module is configured to obtain an amplitude spectrum of the time-frequency representation by a time-frequency transform The corresponding time-frequency representation matrix , for representing the complex amplitude of the first frame, the first frequency point, and calculating the amplitude spectrum according to ;​ A double-time-scale noise estimation module is configured to analyze the amplitude spectrum along the time axis using a double-time-scale noise estimation method , extract short-time noise amplitudes , and long-time noise amplitudes , introduce dynamic weights , fuse and to generate a basic noise amplitude ; a denoising module configured to compute an estimated noise amplitude based on a signal noise amplitude ratio , according to a minimum amplitude threshold and based on a dynamic noise scaling factor and a signal noise amplitude ratio ; a noise reduction and signal reconstruction module for computing a noise reduced spectrum amplitude and from the original spectrum amplitude , preserving the original phase information, by inverse transform to a noise reduced time domain signal .

12. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the adaptive noise reduction filtering method in any one of claims 1-10.

13. An electronic device comprising: A processor and a memory, the memory is used to store one or more programs; characterized in that, when the one or more programs are executed by the processor, a kind of adaptive noise reduction filtering method as claimed in any one of claims 1-10 is realized.

Images