A mixed-flow water turbine vibration signal processing method based on adaptive enhanced EEMD
The adaptive enhanced EEMD method solves the problem of insufficient decomposition of vibration signals in mixed-flow turbines, achieving more accurate fault feature extraction and signal decomposition, and improving the high resolution and robustness of signal processing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- KUNMING UNIV OF SCI & TECH
- Filing Date
- 2025-06-25
- Publication Date
- 2026-07-24
AI Technical Summary
Existing EEMD systems suffer from insufficient signal decomposition and inadequate adaptability when processing vibration signals from mixed-flow turbines, especially in fault conditions where they struggle to effectively separate fault characteristics.
An adaptive enhanced EEMD method is adopted, which improves the signal decomposition process by screening IMF components through wavelet denoising, adaptive frequency band weighted noise injection, dual-constraint mode stopping criteria and fault sensitivity index, and by introducing a compensation term into the residual.
It achieves more accurate fault feature separation, improves the signal-to-noise ratio, enhances the high resolution and robustness of signal decomposition, and can effectively suppress noise interference and mode mixing in the time and frequency domain.
Smart Images
Figure CN120593890B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of vibration signal analysis and processing technology, specifically to a vibration signal processing method for mixed-flow turbines based on adaptive enhanced EEMD. Background Technology
[0002] In a mixed-flow turbine, water enters the runner radially and exits axially. This design results in a compact structure and stable operation, while also exhibiting high conversion efficiency. It maintains high power generation efficiency under varying head and flow conditions, making mixed-flow turbines widely applicable to hydropower stations of all sizes. Vibration signals from mixed-flow turbines often contain a large amount of abnormal data; therefore, preprocessing is typically required when assessing and predicting the unit's health status based on vibration signals.
[0003] Ensemble Empirical Mode Decomposition (EEMD) builds upon EMD decomposition by repeatedly adding white noise to improve decomposition stability and reduce mode aliasing. It utilizes the flat spectral characteristics of white noise and cancels out the influence of noise on IMF components through repetition and averaging, thereby highlighting the true characteristics of the signal itself. Due to its excellent properties, it is frequently used for processing nonlinear and non-stationary signals.
[0004] However, when processing vibration signals from mixed-flow turbines, especially those under fault conditions, there are still problems such as insufficient signal decomposition and inadequate adaptability. Summary of the Invention
[0005] The purpose of this invention is to provide a vibration signal processing method for mixed-flow turbines based on adaptive enhanced EEMD, which solves the problems of insufficient signal decomposition and inadequate adaptability in existing EEMD decomposition processing.
[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a vibration signal processing method for mixed-flow turbines based on adaptive enhanced EEMD, characterized by comprising the following steps: S1. Acquire vibration signals from the mixed-flow turbine. And perform wavelet denoising to obtain ; S2. Inject adaptive frequency band weighted noise into the denoised signal to obtain the denoised signal; S3. Perform EMD decomposition on the noisy signal. During the EMD decomposition process, modify the stopping condition to a dual-constraint modal stopping criterion. Use the dual constraints of the number of extreme points and the rate of change of residual energy to obtain the IMF component set and residual. S4. Effectively screen the IMF components based on the fault sensitivity index, and introduce a compensation term into the residual to obtain the final decomposition result, which is used as the vibration signal processing result.
[0007] A further technical solution is that the specific steps of step S2 are as follows: S2-1. Noise is generated according to the following formula. : in, Sub-bands after the signal frequency band is divided; frequency band The noise gain coefficient, ; The frequency band weighting function controls the distribution of noise in different frequency bands and is dynamically calculated from the power spectral density. For the first Gaussian white noise in each frequency band, with a mean of 0 and a standard deviation of [value missing]. ; The calculation is as follows: in, For the signal in the frequency band The power spectral density is obtained through short-time Fourier transform; This is a power threshold used to distinguish between high and low power frequency bands, and is taken as 10% to 20% of the total signal power. ; is the kurtosis parameter of the Sigmoid function. ;in in, , Indicates vibration signal Standard deviation; Indicates the total signal power; S2-2. Generate the noisy signal: .
[0008] Indicates the number of decompositions.
[0009] A further technical solution is that the dual-constraint modal stopping criterion includes: in, This is the control coefficient for the number of extreme points. ; The signal length; The threshold value for the rate of change of residual energy is used to determine whether the decomposition converges. Suitable for non-stationary signals; For the first In the decomposition of the first The residual signal after the next iteration.
[0010] A further technical solution is that the dual-constraint modal stopping criterion is applicable under the following conditions: When the signal frequency band is >500Hz , ; When the signal frequency band is <10Hz , ; When the signal frequency band is between 10Hz and 500Hz , .
[0011] A further technical solution is that the specific steps of step S4 are as follows: S4-1. Define Fault Sensitivity Indicators : in, As a percentage of energy, For frequency band enhancement factor, For the first The energy of the IMF ; The total energy of all IMFs; This is a typical frequency set of common faults in hydropower units. For IMF components at frequency The amplitude at that point; The average spectral amplitude of the IMF component; S4-2. Filter valid IMFs using the following formula. satisfy The retention condition that does not meet the requirements is considered a residual. Indicates the fault sensitivity threshold; S4-3. A compensation term is introduced into the residual to enhance the sparsity of the high-frequency impulse components and suppress low-frequency modulation interference through the sign function. The final residual formula is as follows: in, For the first The compensation weight coefficients for each IMF are obtained by normalizing the kurtosis values; For the sign function, take the IMF derivative; S4-4. The final decomposition result is obtained as the vibration signal processing result, and the formula is as follows: .
[0012] Compared with the prior art, the beneficial effects of the present invention are: effective suppression of background noise through frequency band weighted noise injection, effective suppression of mode aliasing through dual-constraint stopping criteria and fault sensitivity screening, enabling more accurate separation of fault features and improvement of signal-to-noise ratio; its adaptive noise strategy and fault sensitivity screening mechanism can dynamically balance noise suppression and feature preservation, so that the decomposed signal has both high resolution and strong robustness in the time and frequency domain. Attached Figure Description
[0013] Figure 1 This is a schematic diagram illustrating the definition of the high-speed acquisition system in the embodiment.
[0014] Figure 2 The above is a sampled waveform diagram from an example.
[0015] Figure 3 The waveform diagram is the decomposed waveform in the example.
[0016] Figure 4 This is a comparison chart of time-frequency analysis in the examples.
[0017] Figure 5 This is a comparison diagram of the envelope spectrum in the examples.
[0018] Figure 6 This is a comparison chart of impact energy retention rates in the examples. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0020] Example The traditional EEMD decomposition process can be summarized as follows: Noise generation: in, It is the first The white noise added later, To fix the noise standard deviation, M is the number of decompositions.
[0021] Mode decomposition For each Perform EMD decomposition to obtain the IMF set. and residual ensemble average The final IMF is the mean of all decomposition results: To address the problem that the above methods are insufficient for decomposing vibration signals from mixed-flow turbines, this invention proposes an adaptive enhanced EEMD (AE-FS-EEMD) decomposition method with fault sensitivity for processing vibration signals from mixed-flow turbines, comprising the following steps: S1. Acquire vibration signals from the mixed-flow turbine. And perform wavelet denoising to obtain .
[0022] S2. Inject adaptive frequency band weighted noise into the denoised signal to obtain the denoised signal. Reduce noise injection in high signal power bands (such as fault characteristic frequencies) to avoid masking features; enhance noise in low signal-to-noise ratio bands to improve mode separation stability.
[0023] The specific steps are as follows: S2-1. Noise is generated according to the following formula. : in, The sub-bands after the signal frequency band is divided are specifically taken as multiples of the rotational frequency. For example, the rotational frequency harmonics caused by shaft misalignment, rotor imbalance, etc. are generally 2 times the frequency, so k is 2. frequency band The noise gain coefficient, ; The frequency band weighting function controls the distribution of noise in different frequency bands and is dynamically calculated from the power spectral density. For the first Gaussian white noise in each frequency band, with a mean of 0 and a standard deviation of [value missing]. ; The calculation is as follows: in, For the signal in the frequency band The power spectral density is obtained through short-time Fourier transform; This is a power threshold used to distinguish between high and low power frequency bands, and is taken as 10% to 20% of the total signal power (10% when the signal is strongly periodic; 20% in a strong noise environment, such as air conditioning; and 15% in other cases). ; is the kurtosis parameter of the Sigmoid function. ;in in, , Indicates vibration signal Standard deviation; This indicates the total power of the signal.
[0024] S2-2. Generate the noisy signal: .
[0025] S3. Perform EMD decomposition on the noisy signal. During the EMD decomposition process, modify the stopping condition to a dual-constraint modal stopping criterion, using dual constraints of the number of extreme points and the rate of change of residual energy to obtain the IMF component set and residuals. The dual-constraint modal stopping criterion can avoid premature termination when decomposing the high-frequency impact components of hydropower units, ensuring that weak fault characteristics are not missed.
[0026] The dual-constraint modal stopping criterion includes: in, This is the control coefficient for the number of extreme points. ; The signal length; The threshold value for the rate of change of residual energy is used to determine whether the decomposition converges. Suitable for non-stationary signals; For the first In the decomposition of the first The residual signal after the next iteration.
[0027] The above-mentioned dual-constraint modal stopping criteria are applicable under the following conditions: When the signal frequency band is >500Hz , When decomposing high-frequency components (such as bearing impact and gear meshing noise), the signal changes drastically instantaneously, and the number of extreme points may decrease rapidly, but weak impact characteristics are easily masked by noise.
[0028] When the signal frequency band is <10Hz , When decomposing low-frequency components (such as hydraulic imbalance and impeller oscillation), the signal changes gradually, and the residual energy change rate may remain below the threshold for a long time, but the number of extreme points reflects the physical meaning of the mode.
[0029] When the signal frequency band is between 10Hz and 500Hz , When decomposing intermediate frequency components (such as mechanical loosening or misalignment), the signal exhibits both non-stationarity and periodicity. See Table 1 for details.
[0030] Table 1 Parameter Setting Reference Table The specific Python code (partial) is as follows: def stop_criterion(r_prev, r_current, extrema_count, imf_index): N = len(r_prev) gamma = 1.2 if imf_index <= 3 else (0.6 if imf_index >= 4 else0.8) epsilon = 0.01 if imf_index <= 3 else (0.05 if imf_index >= 4else 0.02) # Calculate the rate of change of residual energy energy_ratio = np.linalg.norm(r_current - r_prev) / np.linalg.norm(r_prev) # Judgment conditions if imf_index <= 3: # High-frequency stage return energy_ratio <= epsilon # Depends only on rule 2 elif imf_index >= 4: # Low-frequency phase return extrema_count <= gamma * np.log(N) # Depends only on rule 1 else: # Mid-frequency stage return (extrema_count <= gamma * np.log(N)) and (energy_ratio<= epsilon) S4. Effectively screen the IMF components based on the fault sensitivity index, and introduce a compensation term into the residual to obtain the final decomposition result, which is used as the vibration signal processing result.
[0031] The specific steps of step S4 are as follows: S4-1. Define Fault Sensitivity Indicators : in, As a percentage of energy, For frequency band enhancement factor, For the first The energy of the IMF ; The total energy of all IMFs; This is a typical frequency set of common faults in hydropower units. For IMF components at frequency The amplitude at that point; The average spectral amplitude of the IMF component; S4-2. Filter valid IMFs using the following formula. satisfy The retention condition that does not meet the requirements is considered a residual. Indicates the fault sensitivity threshold; S4-3. A compensation term is introduced into the residual to enhance the sparsity of the high-frequency impulse components and suppress low-frequency modulation interference through the sign function. The final residual formula is as follows: in, For the first The compensation weight coefficients for each IMF are obtained by normalizing the kurtosis values; For the sign function, take the IMF derivative; S4-4. The final decomposition result is obtained as the vibration signal processing result, and the formula is as follows: .
[0032] To verify the above method, The data acquisition system hardware mainly consists of three parts: sensors, acquisition cards, and edge computers. The vibration sensor is MLS-9, the sound sensor is AWA14423, the eddy current sensor is RSW3308, and the pressure pulsation sensor is AK-4C. The acquisition cards are Advantech's IDAQ-801 and IDAQ-817 series. Similarly, the edge computer is Advantech's UNO-348 series. Specific equipment parameters are shown in Table 2. Table 2 Hardware Configuration The high-speed acquisition module utilizes Advantech's USB high-speed acquisition series boards paired with corresponding data acquisition drivers. All vibration signals acquired by this system are AC-coupled voltage signals. The basic sensor information definitions for the high-speed acquisition system are as follows: Figure 1 As shown in (a), this block diagram mainly determines the channel location, amplitude range, coupling method, etc. of the sensors used; the vibration signals acquired by this system are all AC-coupled voltage signals. The acquisition task is defined as follows: Figure 1As shown in (b), in the middle of the figure, the number of sampling points, sampling rate, and sampling mode need to be configured according to the sensor type connected to different acquisition cards. In this system, the sampling mode is selected as multi-channel multi-sampling (multi-channel means that all channels used by the acquisition card can work at the same time, and multi-sampling means that multiple points are collected at one time. Generally, the number of multi-sampling points is the same as the sampling rate, which means the amount of data collected at one time per second. Multi-sampling facilitates fast Fourier transform). The sampling rate and number of sampling points of the high-speed acquisition card with the sound sensor are both 20k, and the sampling rate and number of sampling points of other acquisition cards are both 2048.
[0033] To align with the research theme of this paper, but limited by the availability of real fault data, a destructive experiment was conducted using a discarded turbine runner. A notch was manually cut into the runner blades, and the runner was then installed on the original unit to simulate wear failures during operation. Considering safety factors, this experiment only retained data from the first 10 seconds of stable operation. The experiment used a mixed-flow turbine unit manufactured by Yunnan Yuxi Hydropower Equipment Co., Ltd., specifically model HLA551-LJ-43, manufactured in the Hydro-Electric Coupling Laboratory of Kunming University of Science and Technology. The runner diameter is 43 cm, the design head is 12 m, the rated flow rate is 0.7 m³ / s, and the turbine output is 65 kW. The generator is an SF55-10 / 740 with a rated capacity of 55 kW and a speed of 600 rpm. This unit maintains its complete shaft system and support structure, making it suitable for studying the vibration characteristics of the shaft system. The test unit was designed according to the actual power plant production requirements and can achieve grid-connected operation and operation with isolated loads. Specific experimental waveforms are shown below. Figure 2 As shown.
[0034] An adaptive enhanced EEMD operation was performed on the vibration fault sequence to obtain IMFs such as Figure 3 As shown in the figure. Regarding the selection of the number of sets and the decomposition noise intensity, as the number of sets increases, the influence of the Gaussian white noise added in the adaptive enhancement EEMD process on the decomposition effect gradually decreases and tends to stabilize. Moreover, different studies have shown that the influence of noise intensity on the result error is also relatively slight. Therefore, when using adaptive enhancement EEMD to process vibration fault time series, the number of sets is set to 200 and the auxiliary noise intensity is 0.4.
[0035] To further verify the effectiveness and practicality of the method proposed in this invention, a comparison of various approaches was conducted. The details are as follows: Time-frequency analysis comparison The comparison results are shown below. Figure 4The time-frequency analysis results show that the adaptive enhanced EEMD method exhibits significant advantages in decomposing unit vibration signals: its IMF3 component presents a high-energy concentrated region at the target fault frequency, with a compact energy distribution and clear boundaries, indicating that the method can accurately extract high-frequency impact features. In contrast, the IMF3 of traditional EEMD shows dispersed energy and mixed low-frequency components in the same frequency band, leading to mode aliasing. Simultaneously, the adaptive enhanced EEMD effectively suppresses high-frequency noise through wavelet pre-denoising and frequency-weighted noise injection, resulting in a significantly lower energy floor than the traditional method (non-characteristic frequency bands are uniformly blue). Traditional EEMD, due to its fixed noise strategy, shows sporadic high-energy points in the same region, indicating severe noise residue. This improvement stems from the dual-constraint stopping criterion's preservation of high-frequency impact components and dynamic suppression of low-frequency interference, enabling the adaptive enhanced EEMD to possess both high resolution and strong robustness in the time-frequency domain.
[0036] Envelope spectrum comparison chart The comparison results are shown below. Figure 5 The envelope spectrum comparison shows that the adaptive enhanced EEMD method has significant advantages in feature extraction and noise suppression: within the target frequency band of 100-200Hz, the amplitude of the adaptive enhanced EEMD is significantly higher than that of VMD and traditional EEMD. For example, at 150Hz, the amplitude of the adaptive enhanced EEMD is approximately -20dB, while that of traditional EEMD is below -40dB, indicating that it preserves the energy of relevant frequencies more completely. Simultaneously, in the non-fault frequency band below 100Hz, the amplitude of the adaptive enhanced EEMD is generally lower, with a peak value close to -80dB, verifying its effective suppression of background noise through frequency band weighted noise injection. This demonstrates the suppressive effect of the dual-constraint stopping criterion and fault sensitivity screening on mode mixing. This result proves that the adaptive enhanced EEMD can more accurately separate fault features and improve the signal-to-noise ratio.
[0037] Comparison of impact energy retention rates (different SNRs) The comparison results are shown below. Figure 6The comparison of impact energy retention rates clearly shows that the adaptive enhanced EEMD method exhibits significantly better robustness and signal preservation capabilities than the traditional EEMD in low signal-to-noise ratio (SNR) environments. Under strong noise interference with an SNR of -5dB, the adaptive enhanced EEMD retains approximately 90% of the impact energy, while the traditional EEMD retains less than 50%. This demonstrates that its wavelet pre-denoising and frequency-weighted noise injection techniques effectively suppress noise overwhelming key features. As the SNR increases to 10dB, the retention rate of the adaptive enhanced EEMD reaches 95%, while the traditional EEMD only achieves 70%. This verifies the accurate capture capability of its dual-constraint stopping criterion for high-frequency weak impact components. Furthermore, the overall trend of the curves shows that the retention rate of the adaptive enhanced EEMD is consistently higher than that of the traditional method at different SNRs, especially at extremely low SNRs (SNR < 0dB), where the retention rate increases by more than 45%. This proves that its adaptive noise strategy and fault sensitivity screening mechanism can dynamically balance noise suppression and feature preservation.
[0038] Although the invention has been described herein with reference to several illustrative embodiments, it should be understood that those skilled in the art can devise many other modifications and implementations that will fall within the scope of this application.
Claims
1. A vibration signal processing method for mixed-flow turbines based on adaptive enhanced EEMD, characterized in that... Includes the following steps: S1. Acquire vibration signals from the mixed-flow turbine. And perform wavelet denoising to obtain ; S2. Inject adaptive frequency band weighted noise into the denoised signal to obtain the denoised signal; The specific steps of step S2 are as follows: S2-1. Noise is generated according to the following formula. : ； in, Sub-bands after the signal frequency band is divided; frequency band The noise gain coefficient, ; The frequency band weighting function controls the distribution of noise in different frequency bands and is dynamically calculated from the power spectral density. For the first Gaussian white noise in each frequency band, with a mean of 0 and a standard deviation of [value missing]. ; The calculation is as follows: ； in, For the signal in the frequency band The power spectral density is obtained through short-time Fourier transform; This is a power threshold used to distinguish between high and low power frequency bands, and is taken as 10% to 20% of the total signal power. ; is the kurtosis parameter of the Sigmoid function. ;in ； in, , Indicates vibration signal standard deviation Indicates the total power of the vibration signal; S2-2. Generate the noisy signal: ； Indicates the number of decompositions S3. Perform EMD decomposition on the noisy signal. During the EMD decomposition process, modify the stopping condition to a dual-constraint modal stopping criterion. Use the dual constraints of the number of extreme points and the rate of change of residual energy to obtain the IMF component set and residual. S4. Effectively screen the IMF components based on the fault sensitivity index, and introduce a compensation term into the residual to obtain the final decomposition result, which is used as the vibration signal processing result; The specific steps of step S4 are as follows: S4-1. Define Fault Sensitivity Indicators : ； in, As a percentage of energy, For frequency band enhancement factor, For the first The energy of the IMF ; The total energy of all IMFs; This is a typical frequency set of common faults in hydropower units. For IMF components at frequency The amplitude at that point; The average spectral amplitude of the IMF component; S4-2. Filter valid IMFs using the following formula. ； satisfy The retention condition that does not meet the requirements is considered a residual. Indicates the fault sensitivity threshold; S4-3. A compensation term is introduced into the residual to enhance the sparsity of the high-frequency impulse components and suppress low-frequency modulation interference through the sign function. The final residual formula is as follows: ； in, For the first The compensation weight coefficients for each IMF are obtained by normalizing the kurtosis values; For the sign function, take the IMF derivative; S4-4. The final decomposition result is obtained as the vibration signal processing result, and the formula is as follows: 。 2. The method according to claim 1, characterized in that: The dual-constraint modal stopping criterion includes: ； in, This is the control coefficient for the number of extreme points. ; The signal length; The threshold value for the rate of change of residual energy is used to determine whether the decomposition converges. Suitable for non-stationary signals; For the first In the decomposition of the first The residual signal after the next iteration.
3. The method according to claim 2, characterized in that: The dual-constraint modal stopping criterion is applicable under the following conditions: When the signal frequency band is >500Hz , ; When the signal frequency band is <10Hz , ; When the signal frequency band is between 10Hz and 500Hz , .