Zero-interventional leaf tip timing signal extraction method based on multi-feature fusion
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2026-03-12
- Publication Date
- 2026-06-09
AI Technical Summary
In zero-intervention leaf tip timing measurement, signal amplitude attenuation, waveform broadening and phase nonlinear delay lead to a decrease in the accuracy of traditional methods and a low signal-to-noise ratio. Existing methods lack effective processing for asymmetric distortion signals.
A multi-feature fusion method is adopted, including features such as waveform peak value, zero crossing point, energy centroid and cross-correlation peak value, and the fusion weight is dynamically adjusted according to the signal quality to adaptively extract the arrival time.
It significantly improves the accuracy and robustness of arrival time extraction under complex operating conditions, reduces computational complexity, and enhances the reliability and noise resistance of signal extraction.
Smart Images
Figure CN122171013A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of non-contact vibration testing and signal processing technology for rotating machinery, specifically relating to a zero-intervention blade tip timing signal extraction method based on multi-feature fusion. Background Technology
[0002] Blade Tip Timing (BTT) technology uses precise measurements of the time difference between the blade passing over a sensor to infer blade vibration parameters, making it an important means of monitoring the health of aero-engine blades.
[0003] Traditional BTT (Block Transmission Theory) technology typically requires drilling holes in the casing to mount optical or capacitive sensors in order to obtain steep pulse signals with a high signal-to-noise ratio. However, in the monitoring of active-duty aircraft engines and certain special tests, drilling holes is often strictly prohibited in order to maintain the integrity and airtightness of the casing structure.
[0004] Therefore, "zero-intervention" measurement technology based on magnetic fields or electromagnetic waves penetrating the metal casing has become a research hotspot. In zero-intervention measurement mode, the sensor is placed on the outer wall of the casing, and the induced magnetic field needs to penetrate the metal casing, which has conductive and magnetic properties. Due to the electromagnetic shielding effect of the casing (eddy current loss and skin effect), the blade signal acquired by the sensor has the following problems:
[0005] First, it exhibits typical "weak coupling" characteristics: the signal amplitude is significantly attenuated, and the signal-to-noise ratio (SNR) is low;
[0006] Second, the waveform is severely broadened, changing from a sharp pulse to a smooth "bell-shaped" wave;
[0007] Third, the phase nonlinear delay and asymmetric distortion cause a serious decrease in the accuracy of traditional peak detection or fixed threshold zero-crossing detection, and drift with changes in rotation speed.
[0008] Existing time-of-arrival extraction methods are mostly designed for ideal high signal-to-noise ratio signals, and lack specific processing strategies for asymmetric distortion signals after penetrating the casing.
[0009] Therefore, there is an urgent need for an adaptive time-of-arrival extraction method that can adapt to waveform distortion and resist strong noise interference. Summary of the Invention
[0010] To address the aforementioned technical problems, this invention proposes a zero-intervention leaf-end timing signal extraction method based on multi-feature fusion. This method comprehensively utilizes multiple features such as waveform peak values, zero-crossing points, energy centroids, and cross-correlation peak values, and dynamically adjusts the fusion weights according to signal quality, achieving high-precision arrival time acquisition.
[0011] To achieve the above objectives, the present invention provides the following technical solution:
[0012] A zero-intervention leaf tip timing signal extraction method based on multi-feature fusion includes the following steps:
[0013] (1) Determine the rotation frequency and rotation period based on the key phase signal, and construct a candidate time window for each blade per rotation according to the key phase trigger time, the number of blades and the theoretical time interval between adjacent blades, so as to intercept the candidate pulse segment signal within the candidate time window;
[0014] (2) Within the candidate pulse segment signal, four types of arrival time features are extracted in parallel: peak time, energy centroid time, cross-correlation peak time, and fitted zero-crossing time.
[0015] (3) Based on the signal quality index of the candidate pulse segment signal, adaptively calculate the fusion weight of the four types of arrival time features, and perform weighted fusion of the four types of arrival time features based on the fusion weight to obtain the fused arrival time;
[0016] (4) Remove outliers from the fusion arrival time and verify the consistency of the time interval between adjacent blades on the arrival time sequence after removing outliers, and output the final arrival time result.
[0017] Further preferred, the construction of candidate time windows in step (1) specifically involves:
[0018] The first step is determined based on the key phase triggering time, the number of blades, and the theoretical time interval between adjacent blades. Circle Theoretical arrival time of each blade Candidate time windows are constructed centered on the theoretical arrival time:
[0019] ,
[0020] in This is a fixed phase offset between the bond phase and blade number 0. The width of the candidate time window; This represents the theoretical time interval between adjacent blades.
[0021] Extracting peak moments in step (2) Specifically:
[0022] Time range of candidate pulse segments Internal satisfaction The moment is taken as the peak moment, where This is the candidate pulse segment signal.
[0023] More preferably, the moment of energy centroid extraction is performed in step (2). Specifically:
[0024] Determined by the centroid of the integral of the signal energy distribution within the candidate pulse segment:
[0025] ,
[0026] in The candidate pulse segment signal has a time range of [missing information]. , The sampling time.
[0027] Extracting the cross-correlation peak time in step (2) Specifically:
[0028] Candidate pulse segment With pre-built standard blade waveform template Perform cross-correlation calculation; the cross-correlation function is:
[0029] ,
[0030] in, Sampling time, To delay time,
[0031] The peak position of the cross-correlation function is taken as the arrival time feature:
[0032] ,
[0033] in, This is the cross-correlation function.
[0034] More preferably, the standard blade template Pre-build by following these steps:
[0035] Under low blade vibration conditions, K blade pulse samples were collected, and the average of the samples was calculated after peak alignment.
[0036] ,
[0037] And on Normalization was performed, where To determine the sample size, This is a sample of the aligned blade pulses.
[0038] In step (2), the zero-crossing time of the fitting is extracted. Specifically:
[0039] Linear fitting is performed on the rising and falling edges of the candidate pulse segment, assuming the signal baseline is... The rising and falling edges are fitted as follows: and ;
[0040] The time of intersection of the two fitted lines and the baseline is: , ;
[0041] Fitting zero-crossing time ,
[0042] in , These represent the slope and intercept of the fitted line along the rising edge, respectively. , These are the slope and intercept of the fitted line along the falling edge, respectively. Sampling time.
[0043] More preferably, in step (3), the adaptive calculation of the fusion weights and the obtaining of the fusion arrival time are specifically as follows:
[0044] Based on the signal quality indices of the candidate pulse segment signals, the fusion weights for the peak time, energy centroid time, cross-correlation peak time, and fitted zero-crossing time are dynamically calculated. , , , The weights satisfy and
[0045] ;
[0046] Fusion arrival time It is obtained by adaptive weighting of four types of arrival time features.
[0047] .
[0048] More preferably, the outlier removal in step (4) specifically involves:
[0049] Outlier detection is performed on the fused arrival time series using median absolute bias (MAD).
[0050] ,
[0051] when When an outlier is detected, it is identified and removed. This is a preset threshold.
[0052] More preferably, the arrival time series consistency check in step (4) specifically includes:
[0053] The arrival time series after outlier removal is validated by checking the time interval between adjacent leaves:
[0054] ,
[0055] when Relocation or removal is triggered at certain times, among which This represents the theoretical time interval between adjacent blades. The consistency threshold is the theoretical time interval between adjacent blades. 5%-10%.
[0056] Compared with the prior art, the present invention has at least the following beneficial effects:
[0057] First, adaptive truncation of candidate time windows reduces computational complexity. Candidate time windows are constructed based on the bond phase signal and the number of blades, precisely limiting the extraction range to local pulse segments, significantly reducing computational load and avoiding false triggering caused by noise spikes and interference from adjacent blades.
[0058] Second, parallel extraction of multiple features comprehensively mines waveform information. Addressing the characteristics of zero-intervention signal amplitude attenuation, waveform broadening, and asymmetric distortion, four types of features are extracted in parallel: peak value, energy centroid, cross-correlation peak value, and fitted zero-crossing point. This characterizes the blade arrival event from different dimensions, overcoming the poor adaptability of single features under complex operating conditions.
[0059] Third, adaptive weighted fusion achieves a dynamic balance between accuracy and robustness. The fusion weights of the four types of features are dynamically adjusted according to the signal quality indicators: when the signal-to-noise ratio is low, the focus is on anti-noise features such as cross-correlation and energy centroid; when the signal-to-noise ratio is high, the focus is on accuracy features such as zero crossing and peak value, so that the algorithm can intelligently adapt to the full envelope operating conditions of the engine.
[0060] Fourth, dual post-processing verification ensures the reliability of the output sequence. Outliers are removed using median absolute deviation, and the consistency of time intervals between adjacent blades is verified. This ensures the reliability of the output sequence from both statistical distribution and physical law perspectives, providing a high-quality time reference for subsequent vibration inversion.
[0061] Fifth, it significantly improves extraction performance under complex operating conditions. Experiments show that compared with the single peak extraction method, the standard deviation of arrival time extraction is reduced by more than 40%; when the signal-to-noise ratio is as low as 3dB, the effective extraction rate is still above 90%, which is better than the less than 60% of the traditional method. It effectively solves the technical problems of weak signal, waveform distortion and operating condition fluctuation in zero-intervention measurement.
[0062] In summary, this invention effectively solves the problems of weak signal, waveform distortion, and operating condition fluctuation in zero-intervention blade tip timing measurement by using parallel extraction of multiple features and adaptive weighted fusion. It significantly improves the accuracy and robustness of arrival time extraction and provides a reliable technical means for blade health monitoring in enclosed casing environments such as aero-engines and gas turbines. Attached Figure Description
[0063] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.
[0064] Figure 2 A schematic diagram illustrating the principle of zero-intervention penetrating blade tip timing measurement;
[0065] Figure 3 A schematic diagram illustrating the differences in the positions of multidimensional features within a candidate pulse segment;
[0066] Figure 4 This is a schematic diagram of an adaptive feature weight allocation strategy based on signal-to-noise ratio. Detailed Implementation
[0067] Specific embodiments will be described in further detail for the present invention.
[0068] Figure 1 The diagram shows the overall process flow of the method of this invention, illustrating the complete processing steps from signal acquisition to the final time series output, highlighting the parallel feature extraction and the adaptive weighted fusion strategy based on signal quality.
[0069] Figure 2 This is a schematic diagram of the zero-intervention penetrating blade tip timing measurement principle, showing a non-contact sensor installed on the outside of a metal casing. The magnetic field or electromagnetic wave generated by the sensor penetrates the casing wall to sense the passage of the internal rotating blade.
[0070] Figure 3 This diagram illustrates the multidimensional feature positional differences within a candidate pulse segment, showcasing a typical pulse waveform distorted and broadened by the casing shielding effect. The diagram clearly marks the peak extraction times within the candidate pulse segment. Energy center of gravity moment Cross-correlation peak time and the fitting time crossing zero The four types of arrival time features demonstrate the inaccuracy of extracting a single feature.
[0071] Figure 4 This is a schematic diagram of a feature weight adaptive allocation strategy based on signal-to-noise ratio (SNR). The curves show how, as the SNR increases, the system automatically reduces its reliance on anti-noise features (such as cross-correlation and energy centroid) and instead increases the weight of high-precision features (such as zero-crossing points and peak values), thereby achieving optimal extraction results under different operating conditions.
[0072] Example 1: Vibration monitoring of compressor blades of a certain type of aero-engine
[0073] This embodiment uses a compressor blade of a certain type of aero-engine as the test object. The casing material is titanium alloy with a wall thickness of 3mm, and the number of blades is 62. The sensor adopts a strong magnetic bias low-frequency eddy current sensor, with 3 sensors installed along the circumference of the casing. The sensor output is connected to a 16-bit high-speed data acquisition card, and the sampling rate is set to 2MHz.
[0074] Step 1: Signal Acquisition and Frequency Conversion Determination
[0075] Real-time acquisition of zero-intervention blade tip timing observation signals and key phase signals. The key phase sensor outputs one pulse per revolution. The current revolution period T=12.5ms is obtained by calculating the time interval between adjacent key phase pulses, corresponding to a revolution frequency f=80Hz.
[0076] Step 2: Constructing Candidate Time Windows
[0077] Determine the first Circle Theoretical arrival time of each blade With candidate time windows ,in This is a fixed phase offset between the bond phase and blade number 0. The width of the candidate time window.
[0078] Step 3: Construction of Standard Blade Template
[0079] During the engine's low-speed (low-vibration) and stable rotation phase, 100 blade signals are continuously collected and synchronously averaged to obtain a noise-free and vibration-free "standard blade waveform template" for subsequent cross-correlation calculations.
[0080] Step 4: Extraction of multiple arrival time features
[0081] (1) Peak time Search within the candidate window for the time corresponding to the point where the absolute value of the signal is maximum. For example... Figure 3 As shown, the peak position of the distorted waveform is offset from the theoretical time, and the error is large when used alone.
[0082] (2) Moment of energy center of gravity Calculated according to the energy center of gravity formula:
[0083] ,
[0084] This feature is not sensitive to noise, but there is a systematic bias when the waveform is severely asymmetrical.
[0085] (3) Peak time of cross-correlation Compare the candidate pulse segment signal with the standard blade waveform template of this blade. Perform cross-correlation calculations:
[0086] ;
[0087] The peak position of the cross-correlation function is taken as the arrival time feature. This feature makes full use of the full waveform information and can maintain high stability even under low signal-to-noise ratio conditions.
[0088] (4) Fitting the zero-crossing time By performing linear fitting on the rising and falling edges of the candidate pulse segment, and assuming the signal baseline is... ,
[0089] The rising and falling edges are fitted as follows: and ;
[0090] The time of intersection is , ;
[0091] Fitting zero-crossing time ,
[0092] in , These represent the slope and intercept of the fitted line along the rising edge, respectively. , These are the slope and intercept of the fitted line along the falling edge, respectively. Sampling time.
[0093] Step 5: Signal quality assessment and adaptive weighted fusion
[0094] Calculate the local signal-to-noise ratio (SNR) of the candidate pulse segment:
[0095] The fusion weights of the four features are dynamically adjusted based on the SNR. This embodiment adopts a piecewise linear weight allocation strategy:
[0096] If an SNR < 5dB (strong noise environment) is detected, the system automatically adjusts the weights as follows:
[0097] , , , At this point, we mainly rely on cross-correlation and energy centroid to "resist interference".
[0098] If an SNR > 20dB (clear signal) is detected, the system adjusts the weights as follows: , , , At this point, the focus is mainly on improving accuracy by using zero-crossing points.
[0099] The fusion arrival time is calculated as follows:
[0100] ,
[0101] The weight It is dynamically calculated based on the signal quality index of the candidate pulse segment, and the weights satisfy... and Based on the local signal-to-noise ratio (a measure of the ratio of signal power to background noise power, typically calculated as the ratio of signal energy to the root mean square of noise within a candidate pulse segment), and through a piecewise linear weighting strategy.
[0102] 1. In noisy environments: Signals may be overwhelmed by noise, and the system automatically increases the weight of anti-interference features. (For example:) =0.5 (cross-correlation) =0.3 (energy center of gravity) =0.1, =0.1);
[0103] 2. In a clear signal environment: The waveform is clear, and the system then increases the weights of features with higher positioning accuracy. (For example:) =0.6 (fitting crosses zero), =0.2, =0.1, =0.1).
[0104] Step 6: Outlier Removal and Consistency Verification
[0105] Outlier detection is performed on the fused arrival time series of consecutive multiple cycles. The median absolute bias (MAD) method is used.
[0106] ,
[0107] when Values identified as outliers are removed at certain times. The preset threshold is initially set to 3. Under low signal-to-noise ratio conditions, to prevent the time shift caused by severe blade vibration from being mistakenly rejected, it can be appropriately adjusted to 3.5.
[0108] Perform a consistency check on the time intervals between adjacent blades for the remaining valid arrival times:
[0109] ,
[0110] when Relocation or removal is triggered at certain times, among which As a consistency threshold, This represents the theoretical time interval between adjacent blades.
[0111] In this embodiment, the consistency threshold Take the theoretical interval 5%-10%.
[0112] Step 7: Results Output and Performance Comparison:
[0113] The final output is a stable and reliable arrival time series, which is used for subsequent blade vibration parameter inversion. Comparative tests were conducted under the conditions of 8000 rpm rotation speed and 200 μm vibration amplitude. Compared with the single peak extraction method, the standard deviation of arrival time extraction by this method is reduced by more than 40%.
[0114] Even under weak signal conditions with a signal-to-noise ratio as low as 3dB, this method can still maintain an effective extraction rate of over 90%, while the traditional threshold method has an extraction rate of less than 60%.
[0115] In summary, this invention effectively solves the problems of weak signal, waveform distortion, and operating condition fluctuation in zero-intervention blade tip timing measurement by using parallel extraction of multiple features and adaptive weighted fusion. It significantly improves the accuracy and robustness of arrival time extraction and provides a reliable technical means for blade health monitoring in enclosed casing environments such as aero-engines and gas turbines.
[0116] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for extracting zero-intervention leaf tip timing signals based on multi-feature fusion, characterized in that, Includes the following steps: (1) Determine the rotation frequency and rotation period based on the key phase signal, and construct a candidate time window for each blade per rotation according to the key phase trigger time, the number of blades and the theoretical time interval between adjacent blades, so as to intercept the candidate pulse segment signal within the candidate time window; (2) Within the candidate pulse segment signal, four types of arrival time features are extracted in parallel: peak time, energy centroid time, cross-correlation peak time, and fitted zero-crossing time. (3) Based on the signal quality index of the candidate pulse segment signal, adaptively calculate the fusion weight of the four types of arrival time features, and perform weighted fusion of the four types of arrival time features based on the fusion weight to obtain the fused arrival time; (4) Remove outliers from the fusion arrival time and verify the consistency of the time interval between adjacent blades on the arrival time sequence after removing outliers, and output the final arrival time result.
2. The method for extracting zero-intervention leaf tip timing signals based on multi-feature fusion according to claim 1, characterized in that, The specific steps for constructing candidate time windows in step (1) are as follows: The first step is determined based on the key phase triggering time, the number of blades, and the theoretical time interval between adjacent blades. Circle Theoretical arrival time of each blade Candidate time windows are constructed centered on the theoretical arrival time: ， in This is a fixed phase offset between the bond phase and blade number 0. The width of the candidate time window; This represents the theoretical time interval between adjacent blades.
3. The method for extracting zero-intervention leaf tip timing signals based on multi-feature fusion according to claim 1, characterized in that, Extracting peak moments in step (2) Specifically: Time range of candidate pulse segments Internal satisfaction The moment is taken as the peak moment, where This is the candidate pulse segment signal.
4. The method for extracting zero-intervention leaf tip timing signals based on multi-feature fusion according to claim 1, characterized in that, Extracting the energy centroid moment in step (2) Specifically: Determined by the centroid of the integral of the signal energy distribution within the candidate pulse segment: ， in The candidate pulse segment signal has a time range of [missing information]. , The sampling time.
5. The method for extracting zero-intervention leaf tip timing signals based on multi-feature fusion according to claim 1, characterized in that, Extracting the cross-correlation peak time in step (2) Specifically: Candidate pulse segment With pre-built standard blade waveform template Perform cross-correlation calculation; the cross-correlation function is: ， in, Sampling time, To delay time, The peak position of the cross-correlation function is taken as the arrival time feature: ， in, This is the cross-correlation function.
6. The method for extracting zero-intervention leaf tip timing signals based on multi-feature fusion according to claim 5, characterized in that, The standard blade template Pre-build by following these steps: Under low blade vibration conditions, K blade pulse samples were collected, and the average of the samples was calculated after peak alignment. ， And on Normalization was performed, where To determine the sample size, This is a sample of the aligned blade pulses.
7. The method for extracting zero-intervention leaf tip timing signals based on multi-feature fusion according to claim 1, characterized in that, In step (2), the zero-crossing time of the fitting is extracted. Specifically: Linear fitting is performed on the rising and falling edges of the candidate pulse segment, assuming the signal baseline is... The rising and falling edges are fitted as follows: and ; The time of intersection of the two fitted lines and the baseline is: , ; Fitting zero-crossing time , in , These represent the slope and intercept of the fitted line along the rising edge, respectively. , These are the slope and intercept of the fitted line along the falling edge, respectively. Sampling time.
8. The method for extracting zero-intervention leaf tip timing signals based on multi-feature fusion according to claim 1, characterized in that, The adaptive calculation of fusion weights and the fusion arrival time in step (3) are specifically as follows: Based on the signal quality indices of the candidate pulse segment signals, the fusion weights for the peak time, energy centroid time, cross-correlation peak time, and fitted zero-crossing time are dynamically calculated. , , , The weights satisfy and ； Fusion arrival time It is obtained by adaptive weighting of four types of arrival time features. 。 9. The method for extracting zero-intervention leaf tip timing signals based on multi-feature fusion according to claim 1, characterized in that, The outlier removal in step (4) specifically involves: Outlier detection is performed on the fused arrival time series using median absolute bias (MAD). ， when When an outlier is detected, it is identified and removed. This is a preset threshold.
10. The method for extracting zero-intervention leaf tip timing signals based on multi-feature fusion according to claim 1, characterized in that, The arrival time series consistency check in step (4) specifically involves: The arrival time series after outlier removal is validated by checking the time interval between adjacent leaves: ， when Relocation or removal is triggered at certain times, among which This represents the theoretical time interval between adjacent blades. The consistency threshold is the theoretical time interval between adjacent blades. 5%-10%.