Multimodal Monitoring System for Compressor Blade Cracks in Turbofan Engines

Abstract

The present invention discloses a multimodal monitoring system for cracks in turbofan engine compressor blades, which belongs to the field of multimodal monitoring and includes: a multimodal acquisition module that obtains multimodal data of each compressor blade, performs dynamic filtering and denoising, and temperature compensation preprocessing, completes time axis alignment, performs data reconstruction and feature extraction, and constructs a multimodal feature vector. Based on this, the trigger module conducts preliminary abnormality diagnosis, determines whether the blade is abnormal through key feature thresholds, and determines whether to trigger depth assessment in combination with periodic thresholds, identifies potential cracked blades, and generates excitation decisions. When executing the excitation decision, the intelligent diagnosis module monitors the strain gradient changes for excitation monitoring, compares the vibration characteristics of the potential cracked blade with the control blade, determines whether there is a crack, and implements multi-level early warning, aiming to solve the problem that the existing technology is difficult to quickly and real-time detect cracks in turbofan engine compressor blades and ensure the safe and stable operation of the engine.

Description

Technical Field

[0001] The present invention relates to the field of multimodal monitoring, and more particularly to a multimodal monitoring system for cracks in turbofan engine compressor blades. Background Art

[0002] The existing turbofan engine compressor blade crack monitoring technology system is primarily based on offline detection methods. Its core limitation lies in the temporal decoupling of detection behavior from the equipment's operating status. Traditional non-destructive testing technologies such as ultrasonic phased arrays and array eddy currents must be implemented in a shutdown state. The detection process is accompanied by complex physical contact and mechanical disassembly, resulting in significant time delays in monitoring behavior. This detection mode is difficult to adapt to the high-load and long-endurance operating characteristics of the engine, and cannot effectively capture the dynamic evolution of crack initiation to propagation. In particular, it lacks real-time early warning capabilities for sudden crack propagation events that may occur during maintenance intervals.

[0003] The current technical system in the field of online monitoring has technical bottlenecks in multiple dimensions. Although acoustic emission detection can sense the stress wave signal of crack expansion, it is limited by noise interference and signal attenuation in complex flow field environments, and its spatial positioning accuracy is difficult to meet engineering requirements; although fiber grating sensing technology can realize real-time monitoring of strain fields, the implantable installation method of its sensor array significantly increases the structural complexity and maintenance costs. In addition, there is still the problem of being unable to provide early warning of blade cracks and locate the position of the cracked blade. Therefore, in order to overcome these limitations, the present invention proposes a multimodal monitoring system for turbofan engine compressor blade cracks. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a multimodal monitoring system for cracks in turbofan engine compressor blades, which breaks through the technical bottlenecks of traditional turbofan engine compressor blade crack monitoring methods in non-stationary signal processing, dynamic operating condition adaptation and early crack identification, and realizes real-time dynamic monitoring and weak feature capture of turbofan engine compressor blade cracks.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The multi-modal monitoring system for turbofan engine compressor blade cracks includes:

[0007] Acquire multimodal data of compressor blades and perform time parameter alignment to construct multimodal feature vectors;

[0008] Based on the multi-modal feature vectors, conduct preliminary abnormal diagnosis on the compressor blades to identify the abnormal blades and mark them. Then, construct a joint trigger rule to trigger in-depth evaluation of the abnormal blades, construct a sequence of trend eigenvalue, and identify potential cracked blades by measuring the difference degree of the trend eigenvalues of the compressor blades, and generate an excitation decision for the potential cracked blades.

[0009] During the execution of the excitation decision, monitor the change of the strain gradient characteristics of the potential cracked blades to conduct excitation monitoring, and select a control blade for the potential cracked blade. Through the comparative analysis of the vibration data characteristics between the control blade and the potential cracked blade, determine whether there is a crack in the potential cracked blade.

[0010] It also includes multi-level early warning according to the results of the comparative analysis.

[0011] Specifically, the specific steps of the preliminary abnormal diagnosis include:

[0012] Receive the multi-modal feature vectors of the compressor blades, select key features through recursive feature elimination and combined with domain prior knowledge; at the same time, obtain the operating condition parameters of the compressor of the turbofan engine and construct an operating condition feature vector.

[0013] Based on Gaussian process regression, establish a non-linear mapping relationship between the operating condition parameters and the key feature thresholds, and dynamically adjust the range of each key feature threshold in real time according to the operating condition feature vector.

[0014] Compare the real-time key feature values with the corresponding key feature threshold ranges, and judge whether all the real-time key feature values of the compressor blades are within the corresponding key feature threshold ranges. If so, it is determined that the compressor blades are in a normal state; otherwise, it is determined that the compressor blades are in an abnormal state and mark them as abnormal blades.

[0015] Specifically, the specific steps of the preliminary abnormal diagnosis also include:

[0016] If it is determined that the compressor blades are in an abnormal state, configure a monitoring threshold to measure the time length for monitoring the compressor blades in an abnormal state. Adopt a joint trigger rule to judge whether to trigger in-depth evaluation according to the number of key features of the abnormal blades exceeding the key feature threshold range and the degree of exceeding of the key feature values within the monitoring threshold.

[0017] Configure a periodic threshold to measure the periodic interval for each compressor blade to trigger in-depth evaluation. If it is determined that the compressor blades are in a normal state, judge whether the time interval from the last in-depth evaluation of the compressor blades is equal to the periodic threshold. If so, trigger in-depth evaluation.

[0018] Specifically, the specific steps of the in-depth evaluation include:

[0019] Based on a periodic threshold, an evaluation period is set according to a sliding time window management strategy, which is used to dynamically adjust the data acquisition window. Taking the current moment as a reference, a multi-modal feature vector of each compressor blade within the evaluation period is obtained;

[0020] For each compressor blade, the values of key features within the evaluation period are arranged in chronological order to construct a key feature value sequence of the compressor blade;

[0021] Trend features are respectively extracted from the key feature value sequences of the compressor blades. The trend features include: monotonic features, non-linear trend features, and periodic features. Each trend feature value is calculated by a non-parametric statistical method to construct a trend feature value sequence for each compressor blade.

[0022] Specifically, the specific steps of in-depth evaluation further include:

[0023] A clustering algorithm is used to cluster the trend feature value sequences of the compressor blades to identify groups of compressor blades with trend deviations and mark them as deviated blades;

[0024] The trend feature value sequences marked as abnormal blades and deviated blades are respectively removed, and the mean values of the trend feature value sequences of the remaining compressor blades are respectively calculated to construct a baseline trend feature value sequence;

[0025] A distance metric is used to calculate the degree of difference between the trend feature value sequences of each deviated blade and abnormal blade and the baseline trend feature value sequence respectively;

[0026] A difference threshold is configured to determine whether the degree of difference between the trend feature value sequence of a deviated blade or an abnormal blade and the baseline trend feature value sequence is greater than the difference threshold. If it is greater, it is marked as a potentially cracked blade.

[0027] Specifically, the specific steps of generating an excitation decision for potentially cracked blades include:

[0028] The circumferential position coordinates and the degree of difference of potentially cracked blades are obtained in real time. Combining the degree of difference with the basic energy of laser pulses, the laser pulse energy is dynamically adjusted;

[0029] The natural frequency of potentially cracked blades is obtained according to the real-time rotational speed, and a laser excitation frequency band is set with the natural frequency as the center;

[0030] According to the position of the potentially cracked blade in the compressor blade group, a transmitter is selected, and the azimuth angle and elevation angle of the transmitter are set according to the relative position between the transmitter coordinates and the circumferential position coordinates of the potentially cracked blade;

[0031] The laser beam is emitted to the surface of the potential crack blade through the transmitter. When there are multiple potential crack blades in a group of compressor blades, parallel excitation is implemented, and the laser beam is emitted in time-sharing order according to the circumferential order of the potential crack blades.

[0032] Specifically, the specific steps for determining whether a potential crack blade has cracks include:

[0033] Extract stress data from the multimodal data of the blade with potential cracks, and perform stress diagnosis on the blade group where the blade with potential cracks is located, including:

[0034] Extract strain gradient eigenvalues based on stress data and configure strain thresholds. If the strain gradient eigenvalue of a blade with a potential crack exceeds the strain threshold, laser excitation is stopped and an excitation warning is triggered.

[0035] Based on the principle of spatial symmetry, the position coordinates of the potential cracked blade in the compressor blade group are determined, and the control blade is selected according to the spatial layout structure of the compressor blade;

[0036] Extracting vibration data features from multimodal data of blades with potential cracks and control blades, the vibration data features include vibration order amplitude and time-frequency spectrum features;

[0037] Configure the order amplitude threshold. For the vibration order amplitude, calculate the relative deviation between the amplitude of the blade with potential crack and the vibration order of the control blade at the same order. If the relative deviation between the amplitude of the blade with potential crack and the vibration order of the control blade at the same order exceeds the order amplitude threshold, the blade with potential crack is judged to have a crack.

[0038] Specifically, the specific steps of determining whether a potential crack blade has cracks also include:

[0039] Configuring principal component thresholds , perform empirical mode decomposition on the vibration data of the potential crack blade and the control blade, decompose the vibration data into multiple eigenmode functions, and select the first mode function with the highest energy proportion. Perform Hilbert transform on the intrinsic mode functions to obtain the time-frequency spectrum;

[0040] The characteristic frequency of the time-frequency spectrum is identified by peak detection. On the time-frequency spectrum, the frequency bandwidth is set around the characteristic frequency, and the energy density of the characteristic frequency is calculated by integration.

[0041] An energy deviation threshold is set to calculate the relative deviation of the energy density of the characteristic frequency of the blade with potential cracks and the control blade. If the deviation is greater than the energy deviation threshold, it is determined that the blade with potential cracks has cracks.

[0042] When it is determined that a potential crack blade has cracks, the stimulation of the potential crack blade is stopped and a first-level blade warning is issued;

[0043] Configure the excitation duration, monitor the duration of exciting the potentially cracked blade, and if it is equal to the excitation duration and no crack is determined in the potentially cracked blade, issue a secondary blade warning.

[0044] Specifically, the specific steps for constructing the multi-modal feature vector include:

[0045] Obtain the real-time rotational speed of the compressor, which is used to resample the vibration data in the multi-modal data to generate vibration data with equal angular intervals; and define the target order according to the number of compressor blades in each group.

[0046] Define the resonance frequency band, filter the resampled vibration data to obtain the filtered vibration data and perform Hilbert transform to generate the orthogonal data of the vibration data, calculate the envelope data, and perform Fourier transform on the envelope data to obtain the envelope spectrum.

[0047] Perform windowing on the resampled vibration data and then perform Fourier transform to obtain the spectrum of the resampled vibration data.

[0048] According to the target order, map the frequency axes of the spectrum of the resampled vibration data and the envelope spectrum to the order axis, and use the interpolation algorithm to generate the corresponding order spectrum.

[0049] Extract features from the order spectrum respectively, and combine with the time-domain features of the multi-modal data to splice and form the multi-modal feature vector of each compressor blade.

[0050] Specifically, the joint trigger rules include:

[0051] Within the monitoring threshold time range, count the number of key features that exceed the corresponding key feature threshold range. When the number of key features that exceed the key feature threshold range exceeds the preset number threshold, the first rule is satisfied.

[0052] Calculate the relative deviation degree of each key feature value that exceeds the key feature threshold range with respect to its corresponding key feature threshold range. When there is a relative deviation degree that exceeds the preset deviation degree threshold, the second rule is satisfied.

[0053] When both the first rule and the second rule are satisfied, that is, the number of key features that exceed the key feature threshold range reaches the preset number threshold and the relative deviation degree of at least one key feature reaches the preset deviation degree threshold, it is determined to trigger the depth evaluation.

[0054] The beneficial effects of the present invention:

[0055] The present invention realizes the accurate extraction and spatio-temporal alignment of multi-dimensional features such as blade vibration and stress through multi-modal data fusion technology, significantly improving the accuracy of crack feature recognition; the dynamic threshold adaptation mechanism combines Gaussian process regression and joint trigger rules to effectively reduce the false alarm rate under complex working conditions; the trend feature analysis and baseline comparison algorithm based on a sliding time window successfully capture the weak change trend of early cracks; the dynamic optimization strategy of laser excitation parameters and parallel time-sharing emission technology greatly improves the crack verification efficiency while ensuring excitation safety; the multi-level early warning mechanism realizes fault classification response and closed-loop management through strain gradient monitoring and vibration data comparison analysis. A full-process monitoring system covering real-time monitoring, early crack warning, precise positioning and verification is constructed, which has real-time dynamic adaptation ability, weak feature recognition ability and multi-target collaborative processing ability, and finally significantly improves the safety and reliability of the operation of the compressor of the turbofan engine. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a functional schematic diagram of the multi-modal monitoring system for cracks in the compressor blades of the turbofan engine of the present invention;

[0057] Figure 2 is a flowchart of the specific steps for the preliminary abnormal diagnosis of the present invention;

[0058] Figure 3 is a flowchart of the specific steps for the in-depth evaluation of the present invention;

[0059] Figure 4 is a flowchart of the specific steps for stress diagnosis of the present invention;

[0060] Figure 5 is a flowchart of the specific steps for determining whether there are cracks in the potentially cracked blades of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0061] Please refer to Figure 1 , this embodiment introduces a multi-modal monitoring system for cracks in the compressor blades of a turbofan engine, including: a multi-modal acquisition module, a trigger module and an intelligent diagnosis module;

[0062] The multi-modal acquisition module is used to obtain multi-modal data of each compressor blade through distributed sensors, including vibration data and stress data, and perform data preprocessing on the acquired multi-modal data, including: dynamically adjusting the filter parameters according to the engine speed, denoising the vibration data to filter out non-blade vibration interference, performing temperature compensation on the stress data, and after time-axis alignment of the preprocessed multi-modal data, performing feature extraction through multi-modal data reconstruction to construct a multi-modal feature vector of each compressor blade.

[0063] In this embodiment, based on the stress distribution, aerodynamic load characteristics, and crack propagation mechanism of the compressor blade, a distributed design is adopted. Two orthogonal strain sensors are arranged at the root tenon of each compressor blade to monitor the circumferential and radial strain data of the compressor blade under centrifugal load. MEMS accelerometers are evenly and symmetrically installed on the inner wall circumference of the compressor casing to monitor the full-circumference vibration data. According to the operating speed and vibration characteristics of the compressor blade, the acquisition frequency of the vibration data is set. For example, the acquisition frequency is more than twice the highest vibration frequency of the blade to avoid signal aliasing. The acquisition frequency of the stress data is determined according to the rate of strain change. Since the strain change is relatively slow, the acquisition frequency can be reduced, but it should be ensured that the dynamic changes of the stress can be captured. A band-pass filter is used to filter out the low-frequency and high-frequency non-blade vibration interferences. The mapping relationship between the engine speed and the filter parameters is established. Through experiments and theoretical analysis, the variation law of the vibration frequency of the blade at different engine speeds is determined. For example, when the engine speed increases, the vibration frequency of the blade also increases accordingly. At this time, the passband range of the filter is adjusted so that it can better capture the vibration signal of the blade. During the data acquisition process, the engine speed is monitored in real time, and the filter parameters are dynamically adjusted according to the mapping relationship. A temperature sensor is arranged near the strain sensor to monitor the temperature change at the position where the strain sensor is located in real time. The temperature sensor should be in close contact with the strain sensor to ensure that the temperature at the same position can be accurately measured. The output data of the strain sensor at different temperatures are collected through experiments, and a relationship model between the temperature and the strain measurement value is established. Methods such as polynomial fitting and neural networks are used to establish the model. During the data acquisition process, the measurement value of the temperature sensor is obtained in real time, and the collected stress data are compensated in real time according to the temperature compensation model. After aligning the vibration data and the stress data on the time axis using the interpolation algorithm, the features of the vibration data and the stress data are extracted, including time-domain features, frequency-domain features, joint features, and change rate features, and a high-dimensional multi-modal feature vector is formed by splicing.

[0064] Preferably, the specific steps for constructing the multi-modal feature vector of each compressor blade include:

[0065] Obtain the real-time speed of the compressor and resample the vibration data based on the real-time speed to generate vibration data with equal angular intervals, that is: according to the real-time speed, the vibration signal is resampled into a sequence with equal angular intervals through the interpolation algorithm as the resampled vibration data to eliminate the spectral ambiguity caused by speed fluctuations; the phase information of the resampled vibration data is completely retained, providing an accurate angular domain signal basis for subsequent order analysis and ensuring the correspondence between the frequency components and the mechanical structure characteristics.

[0066] Define the target order according to the number of compressor blades in each group. For example, if the number of blades in a certain stage of the compressor is 50, then define the target order as 50 to focus on analyzing the 50th order and its harmonics; directly associate the vibration energy distribution with the characteristics of the blade rotating machinery to improve the recognition of crack-related features.

[0067] Combine historical vibration data to delimit the high-frequency resonance frequency band. Use a band-pass filter to filter the resampled vibration data, obtain the filtered vibration data and perform Hilbert transform to generate its orthogonal data to calculate the envelope data, and perform Fourier transform on the envelope data to obtain the envelope spectrum; decouple the low-frequency modulation information in the high-frequency carrier signal, where the low-frequency modulation information includes the crack impact period, generate the envelope spectrum, and highlight the modulation characteristics related to the blade rotation frequency.

[0068] Perform windowing on the resampled vibration data to reduce spectral leakage, perform discrete Fourier transform to obtain the spectrum of the resampled vibration data; provide an accurate spectral basis for order mapping.

[0069] According to the target order, map the frequency axes of the spectrum of the resampled vibration data and the envelope spectrum to the order axis, and use the interpolation algorithm to generate the corresponding order spectrum to focus on the blade resonance characteristics;

[0070] Extract features from the order spectrum of the resampled vibration data and the order spectrum corresponding to the envelope spectrum respectively, including amplitude features and phase features; combine the time-domain features of the vibration data including peak value and kurtosis with the time-domain features of the stress data, that is, the strain gradient feature, and splice them to form a multi-modal feature vector for each compressor blade.

[0071] The trigger module is used to perform a preliminary anomaly diagnosis on each compressor blade during the operation of the compressor of the turbofan engine according to the obtained multi-modal feature vector, quickly judge whether there is an abnormal phenomenon, and combine the periodic threshold to judge whether to trigger a depth evaluation to identify potential cracked blades and generate a decision to excite potential cracked blades;

[0072] Please refer to Figure 2 , preferably, the specific steps of the preliminary anomaly diagnosis include:

[0073] Receive the multimodal data and multimodal feature vectors of each compressor blade in real time. Select the key features of the multimodal feature vectors by recursive feature elimination combined with domain prior knowledge, including: vibration order amplitude, envelope spectrum sideband energy ratio, stress gradient, strain rate, etc., for rapid diagnosis of each compressor blade; Based on the historical multimodal feature vectors of each compressor blade, construct a linear model, with the multimodal feature vectors as the input and whether the blade is abnormal as the output. Then, calculate the weight of each feature, gradually remove the feature with the smallest weight, retrain the model, and repeat this process until the number of remaining features reaches the preset value to determine the features closely related to the crack propagation of the blade, including vibration order amplitude, envelope spectrum sideband energy ratio, stress gradient, etc. Combine and adjust these features with the features selected by recursive feature elimination to finally determine the key features.

[0074] At the same time, obtain the operating condition parameters of the current compressor of the turbofan engine, including real-time speed, load, and temperature, and construct an operating condition feature vector; collect the historical operating condition parameters and the corresponding key feature values, divide them into a training set and a test set, and establish a nonlinear mapping relationship between the operating condition parameters and the key feature thresholds based on Gaussian process regression to dynamically adjust the range of each key feature threshold;

[0075] Compare the real-time key feature values with the corresponding key feature threshold ranges to determine whether all real-time key feature values are within the corresponding key feature threshold ranges. If so, it is determined that the compressor blade is in a normal state; otherwise, it is determined that the compressor blade is in an abnormal state, and this blade is marked as an abnormal blade;

[0076] If it is determined that the compressor blade is in an abnormal state, configure a monitoring threshold to measure the length of time for monitoring the compressor blade in an abnormal state. Adopt a joint trigger rule to determine whether to trigger a deep evaluation based on the number of key features of the abnormal blade exceeding the key feature threshold range and the degree of exceeding of the key feature values within the monitoring threshold;

[0077] Configure a periodic threshold to measure the periodic interval for each compressor blade to trigger a deep evaluation. If it is determined that all blades in the current blade group are in a normal state, determine whether the time interval since the last deep evaluation of the compressor blade is equal to the periodic threshold. If it is equal, trigger a deep evaluation; otherwise, do nothing.

[0078] The joint trigger rule includes:

[0079] Within the monitoring threshold time range, count the number of key features that exceed the corresponding key feature threshold range. When the number of key features exceeding the key feature threshold range exceeds the preset quantity threshold, the first rule is satisfied. This condition is set to ensure that multiple key features are abnormal simultaneously because the abnormality of a single feature may be caused by accidental factors, while the simultaneous abnormality of multiple features can better reflect substantial problems in the blade group.

[0080] Meanwhile, calculate the relative deviation degree of each key feature value that exceeds the key feature threshold range with respect to its corresponding key feature threshold range. When there is a relative deviation degree exceeding the preset deviation degree threshold, the second rule is satisfied. This condition emphasizes that at least one key feature has shown a significant abnormal change, and such a significant change may be an important signal indicating that the blade group failure has developed to a certain extent.

[0081] When both the first rule and the second rule are satisfied, that is, when the number of key features exceeding the key feature threshold range reaches the preset quantity threshold and the relative deviation degree of at least one key feature reaches the preset deviation degree threshold, it is determined that a depth assessment is triggered. This combined triggering method comprehensively considers the quantity and degree of feature abnormalities, can more accurately capture the abnormal state of the blade group, avoid false triggering caused by accidental fluctuations of individual features, and can also timely detect potential serious fault hazards, providing a reliable basis for further diagnosis and treatment. If the combined triggering rule is satisfied, a depth assessment is triggered; otherwise, the multi-modal feature vectors of each compressor blade are evaluated in real time.

[0082] Please refer to Figure 3 , preferably, the specific steps of the depth assessment include:

[0083] Based on the periodic threshold, set the evaluation period according to the sliding time window management strategy. The sliding time window management strategy is used to comprehensively consider the data of multiple historical periods before the current moment by dynamically adjusting the data window, so as to more comprehensively capture the change trend of the compressor blade performance. The periodic threshold represents a basic time unit, and the evaluation period will be set based on this. That is, the time length covered by the evaluation period is a positive integer multiple of the periodic threshold. Taking the current moment as the reference, trace back the time range covered by the evaluation period forward to form a dynamic data window;

[0084] Obtain the multi-modal feature vectors of each compressor blade within the evaluation period. For each compressor blade, arrange the values of the key features within the evaluation period in chronological order to construct a key feature value sequence of the compressor blade;

[0085] Trend features are extracted from the key eigenvalue sequences of compressor blades respectively. The trend features include monotonic features, nonlinear trend features and periodic features. The trend eigenvalues are calculated using non-parametric statistical methods, including the Mann-Kendall test statistic Z value, Sen's slope, Cox-Stuart trend strength, and autocorrelation analysis. The monotonic eigenvalues calculated by non-parametric statistical methods for each compressor blade, such as the Z value and Sen's slope, the nonlinear trend eigenvalues including the p-value of the Cox-Stuart test and the polynomial fitting coefficient, and the periodic eigenvalues including the lag order corresponding to the peak of the autocorrelation function, are combined together to construct a trend eigenvalue sequence for each compressor blade.

[0086] A trend characteristic value sequence for each compressor blade is constructed, and a clustering algorithm is used to cluster these trend characteristic value sequences. Clusters of compressor blades with deviating trends are identified and marked as deviating blades. A density-based spatial clustering algorithm can be used to cluster the trend characteristic value sequence of compressor blades, dividing the blades into core points, boundary points, and noise points. By analyzing the clustering results, blades in different clusters are identified as having different trend characteristics, and blade groups that deviate from the normal clusters are marked as deviating blades. For example, if a cluster contains the majority of blades, while a few other blades form small clusters or are noise points, these few blades can be marked as deviating blades.

[0087] Based on the clustering results, the trend feature value sequences marked as abnormal and deviated leaves are identified and removed. The trend feature values of the remaining normal leaves are statistically analyzed, and the mean of each trend feature value is calculated to construct a baseline trend feature value sequence. During the mean calculation process, a weighted average method can be used to assign different weights to different leaves based on their distance from the cluster center to improve the accuracy of the baseline.

[0088] The distance metric is used as a method to measure the degree of difference between the trend characteristic value sequence of each deviated leaf and abnormal leaf and the baseline trend characteristic value sequence, and the degree of difference between the trend characteristic value sequence of each deviated leaf and abnormal leaf and the baseline trend characteristic value sequence is calculated respectively;

[0089] By analyzing historical data and combining it with actual fault cases, a difference threshold is configured to determine whether the difference between the trend characteristic value sequence of the deviated blade or abnormal blade and the baseline trend characteristic value sequence is greater than the difference threshold. If so, it is marked as a blade with potential cracks. Otherwise, no processing is performed.

[0090] Preferably, the specific steps of generating an incentive decision include:

[0091] Obtain the circumferential position coordinates of the potentially cracked blade and its degree of difference in real time, and dynamically adjust the laser pulse energy according to the degree of difference to enhance the crack response signal, that is:

[0092]

[0093] Among them, is the laser pulse energy after dynamic adjustment, is the basic laser pulse energy, which is determined through material damage threshold experiments, is the energy gain coefficient, and the value range is 0.5 ≤ ≤ 1.5 to avoid overload, is the allowable maximum difference threshold, is the degree of difference of the potentially cracked blade; when the degree of difference is large, it indicates that the blade may have relatively serious cracks. Appropriately increase the energy gain coefficient and increase the laser pulse energy to enhance the crack response signal; when the degree of difference is small, reduce the energy gain coefficient and reduce the laser pulse energy. While ensuring the detection effect, reduce the unnecessary impact on the compressor blade. At the same time, set the allowable maximum difference threshold. When the degree of difference exceeds this threshold, adopt a special energy adjustment strategy or send out a warning signal.

[0094] Monitor the engine speed in real time. According to the structural parameters of the blade, including the number of blades, mass distribution, stiffness, and real-time speed information, calculate the natural frequency of the potentially cracked blade. Centered on the natural frequency, set the laser excitation frequency band; the range of this frequency band is determined according to the dynamic characteristics of the blade and the detection requirements, and can be set to 1.2 - 2.0 times the natural frequency. Emitting laser pulses within the laser excitation frequency band can better stimulate the resonance effect of the blade, enhance the crack response signal, and improve the sensitivity of crack detection.

[0095] According to the position of the potentially cracked blade in the compressor blade group, select the transmitter. The transmitters are evenly arranged on the inner wall of the compressor casing in a circumferentially uniform distribution. Each group of transmitters contains 3 orthogonal direction units. This layout design can cover all areas inside the compressor to ensure effective detection of all blades. At the same time, a certain gap is maintained between the transmitter and the blade tip. This gap is accurately calculated and experimentally verified, which can not only ensure that the laser beam can accurately irradiate the blade surface, but also avoid collision or interference between the transmitter and the blade. According to the coordinates of the transmitter and the circumferential position coordinates of the potentially cracked blade, calculate the relative position relationship between the two, and adjust the azimuth angle and pitch angle of the transmitter in real time to ensure that the laser beam accurately aligns with the surface of the potentially cracked blade and improve the excitation effect.

[0096] The laser beam is emitted to the surface of the blade with potential cracks through a transmitter. When there are multiple blades with potential cracks in a group of compressor blades, parallel excitation is implemented. According to the circumferential order of the blades with potential cracks, the laser beam is emitted at different times, and a certain time interval is set between laser pulses to avoid signal interference between different blades and ensure the accuracy of the detection results.

[0097] The intelligent diagnosis module is used to monitor the excitation process through stress diagnosis and select a reference blade when making excitation decisions. By comparing the characteristics of the vibration data of the blade with potential cracks and the reference blade, it is determined whether there are cracks in the blade with potential cracks, and multi-level early warnings are carried out.

[0098] Preferably, the specific steps for determining whether there are cracks in the blade with potential cracks include:

[0099] Please refer to Figure 4 , obtain the stress data in the multi-modal data of the blade with potential cracks, and perform stress diagnosis on the blade group where the blade with potential cracks is located, including:

[0100] Extract the strain gradient eigenvalue. The strain gradient eigenvalue reflects the change of strain in the blade under stress. Its calculation method is based on the strain data collected by the orthogonal strain sensors at the blade root tenon. By calculating the ratio of the difference in strain at adjacent positions to the distance, the strain gradient is obtained.

[0101] Configure the strain threshold. The determination of the strain threshold is based on the statistical analysis of the strain gradient data of a large number of normal blades under the same excitation conditions, and comprehensively considers factors such as the material properties of the compressor blades, the designed stress-bearing range, and the actual operating conditions. Perform stress diagnosis and judgment. If the strain gradient eigenvalue of the blade with potential cracks is greater than the strain threshold, immediately stop the laser excitation. When the strain gradient eigenvalue is greater than the strain threshold, it indicates that the mechanical response of the blade during the laser excitation process is abnormal, and there may be situations such as an accelerated crack propagation rate and excessive damage to the blade structure. At this time, stopping the excitation in time can effectively avoid further irreparable damage to the compressor blade. At the same time, trigger an excitation warning and send an alarm signal to the operator to remind him to conduct a more in-depth inspection and evaluation of the compressor blade. The alarm signal can be presented in various ways such as an audible and visual alarm device and an information pop-up window of the engine monitoring system to ensure the safe and stable operation of the engine.

[0102] Please refer to Figure 5 , based on the principle of spatial symmetry, determine the position coordinates of the blade with potential cracks in the compressor blade group, and clarify the blade corresponding to its spatially symmetric position according to the spatial layout structure of the compressor blades, and use this blade as the initially selected reference blade.

[0103] If the control blade at the spatially symmetric position of the potentially cracked blade is still a potentially cracked blade, then the next blade at the spatially symmetric position of the potentially cracked blade is reselected as the control blade in the priority order from top to bottom and from left to right. The "up", "down", "left", and "right" directions here are determined based on the fixed coordinate system of the compressor blade group. For example, the compressor axial direction is the up-down direction, and the blade circumferential direction is the left-right direction. When selecting, adjacent blades are checked in turn until a non-potentially cracked blade is found as the control blade.

[0104] Extract the vibration data features in the multi-modal data of the potentially cracked blade and the control blade. The vibration data features include vibration order amplitude and time-frequency spectrum diagram features; different vibration data features have different sensitivities to blade faults. For example, changes in vibration order amplitude may be related to blade resonance or imbalance, and the time-frequency spectrum diagram features can reveal the variation law of the vibration signal in time and frequency, providing rich information for subsequent crack judgment.

[0105] Conduct a comparative analysis of the vibration data features of the potentially cracked blade and the control blade, including:

[0106] Configure the order amplitude threshold. For the vibration order amplitude, calculate the relative deviation of the amplitudes of the potentially cracked blade and the control blade at the same order. If the relative deviation of the amplitudes of the potentially cracked blade and the control blade at the same order exceeds the order amplitude threshold, it is determined that the potentially cracked blade has a crack; when the relative deviation exceeds the order amplitude threshold, it indicates that the vibration amplitude of the potentially cracked blade has an abnormal change, and this change may be caused by internal structural damage of the blade, thus providing a strong basis for crack judgment and improving the accuracy of crack diagnosis.

[0107] Configure the principal component threshold , perform empirical mode decomposition on the vibration data of the potentially cracked blade and the control blade, decompose the vibration data into multiple intrinsic mode functions, and select the first intrinsic mode functions with the largest energy proportion for Hilbert transform to obtain the time-frequency spectrum diagram, which reflects the change of data energy over time and frequency, highlights the essential features of the signal, makes the analysis of the blade vibration state more in-depth and detailed, and helps to discover potential fault information;

[0108] Identify the characteristic frequency of the time-frequency spectrum diagram through peak detection. On the time-frequency spectrum diagram, set the frequency bandwidth around the characteristic frequency, and calculate the energy density of the characteristic frequency through integration to further extract useful information in the vibration signal. Changes in the characteristic frequency and its energy density can reflect changes in the dynamic characteristics of the blade. When there is a crack in the blade, the characteristic frequency and energy distribution of its vibration will change, enabling a more accurate judgment of whether there is a crack in the blade and the degree of crack development.

[0109] Set an energy deviation threshold, calculate the relative deviation of the energy density of the characteristic frequencies between the potentially cracked blade and the reference blade. If it is greater than the energy deviation threshold, indicating a significant difference in the energy density of the characteristic frequencies between the potentially cracked blade and the reference blade, then it is determined that there is a crack in the potentially cracked blade.

[0110] When it is determined that there is a crack in the potentially cracked blade, stop exciting the potentially cracked blade, issue a first-level blade warning. The first-level blade warning uses a strong alarm method, including a high-volume alarm sound, a flashing red warning light, etc., and at the same time generate first-level blade warning information, including the location of the cracked blade and relevant characteristic data, to remind the operator to take measures as soon as possible.

[0111] Configure the excitation duration, which should be determined comprehensively according to factors such as the material properties of the blade and the intensity of laser excitation. Monitor the continuous duration of exciting the potentially cracked blade. If it is equal to the excitation duration and it is not determined that there is a crack in the potentially cracked blade, then issue a second-level blade warning. The second-level blade warning is used to prompt the operator that although no crack has been detected in this blade, the excitation duration has reached and there is an abnormality in the blade, and further attention is required.

[0112] Working principle and its effects:

[0113] Full-process monitoring of compressor blade cracks in turbofan engines through multimodal data fusion and intelligent diagnosis technology: First, a distributed sensor network is used to collect multi-dimensional data such as vibration, stress, and temperature in real time. A resampling technology based on rotational speed synchronization is adopted to eliminate the rotational speed fluctuation interference in non-stationary signals. Combining Hilbert transform and envelope analysis, a multimodal feature vector containing time-domain features, frequency-domain features, and stress gradient features is constructed, solving the problem of incomplete feature extraction in traditional methods for non-stationary signal processing. In the abnormal diagnosis stage, a recursive feature elimination algorithm is used to screen key features, and a non-linear mapping model between operating parameters and feature thresholds is established by combining Gaussian process regression to achieve dynamic threshold update. Compared with traditional fixed-threshold methods, the false alarm rate under complex operating conditions is significantly reduced. The joint trigger rule effectively balances detection sensitivity and specificity by counting the number of features exceeding the threshold within the monitoring period and calculating the deviation degree of key features, avoiding misjudgment of single features. An adaptive sliding time window strategy is used to extract the long-term change trend of key features, and non-parametric statistical methods are used to quantify trend features such as monotonically increasing and periodic changes. Combining density clustering algorithm to identify the blade group with trend deviation. By constructing a baseline trend feature value sequence and calculating the dynamic time warping distance difference degree, the feature drift caused by early cracks can be effectively detected. For potentially cracked blades, the laser excitation parameters are adjusted in real time based on modal analysis, and a spatial coordinate positioning technology is used to achieve coordinated control of the azimuth angle and pitch angle of the emitter. The parallel time-sharing excitation strategy ensures the safety and efficiency of multi-target processing. During the excitation process, the sudden change of strain gradient and the deviation of vibration order amplitude are monitored synchronously. Combining empirical mode decomposition and Hilbert transform to extract the time-frequency domain feature differences, and the multi-level warning mechanism realizes the full-process management from excitation anomaly to crack confirmation through a hierarchical response strategy.

[0114] This system breaks through the technical bottlenecks of traditional monitoring technologies in non-stationary signal processing, dynamic operating condition adaptation, and early fault identification, and constructs a closed-loop monitoring system including data acquisition, feature fusion, intelligent diagnosis, precise excitation, and verification warning. The crack recognition accuracy is improved through multi-dimensional feature fusion, the dynamic threshold mechanism significantly enhances the adaptability to complex operating conditions, the early warning algorithm realizes the capture of weak features, the precise excitation technology ensures the safety and efficiency of crack verification, and the multi-level warning mechanism optimizes the maintenance decision response process. These technological innovations together improve the intelligent level of turbofan engine blade health management and provide a solid guarantee for the safe and reliable operation of the engine.

[0115] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A multi-modal monitoring system for cracks in compressor blades of a turbofan engine, characterized in that include: Acquire multimodal data of compressor blades and perform time parameter alignment to construct multimodal feature vectors; Based on the multimodal eigenvectors, a preliminary abnormality diagnosis of the compressor blades is performed to identify abnormal blades and mark them. The trigger depth of the abnormal blades is evaluated by constructing a joint triggering rule. A trend eigenvalue sequence is constructed and the degree of difference in the compressor blade trend eigenvalues is measured to identify potential cracked blades. The incentive decision for the potential cracked blades is generated. The specific steps of the preliminary abnormality diagnosis include: Receive the multimodal feature vector of the compressor blade, select key features through recursive feature elimination and combine with domain prior knowledge; at the same time, obtain the operating condition parameters of the turbofan engine compressor and construct the operating condition feature vector; Based on Gaussian process regression, a nonlinear mapping relationship between operating condition parameters and key feature thresholds is established. According to the operating condition feature vector, the threshold range of each key feature is dynamically adjusted in real time. Compare the real-time key feature values with the corresponding key feature threshold ranges to determine whether all real-time key feature values of the compressor blades are within the corresponding key feature threshold ranges. If so, the compressor blades are determined to be in a normal state; otherwise, the compressor blades are determined to be in an abnormal state and are marked as abnormal blades. If the compressor blades are determined to be in an abnormal state, a monitoring threshold is configured to measure the length of time to monitor the abnormal compressor blades. A joint triggering rule is used to determine whether to trigger a deep assessment based on the number of key features of the abnormal blades that exceed the key feature threshold range within the monitoring threshold and the degree of excess of the key feature values. Configure a periodic threshold to measure the periodic interval for triggering a depth assessment on each compressor blade. If the compressor blade is determined to be in a normal state, determine whether the time interval between the last depth assessment and the compressor blade is equal to the periodic threshold. If so, a depth assessment is triggered. During the execution of the excitation decision, excitation monitoring is performed by monitoring the changes in the strain gradient characteristics of the blade with potential cracks, and a control blade of the blade with potential cracks is selected. By comparing and analyzing the vibration data characteristics of the control blade and the blade with potential cracks, it is determined whether the blade with potential cracks has cracks. It also includes multi-level early warning based on the results of comparative analysis.

2. The multi-modal monitoring system for cracks in the compressor blades of a turbofan engine according to claim 1, characterized in that, The specific steps of the in-depth assessment include: Based on the periodic threshold, an evaluation period is set according to a sliding time window management strategy. The sliding time window management strategy is used to dynamically adjust the data value window and obtain the multimodal feature vector of each compressor blade within the evaluation period based on the current time. For each compressor blade, the values of the key features within the evaluation period are arranged in chronological order to construct the key feature value sequence of the compressor blade; Trend features are extracted from the key eigenvalue sequences of compressor blades respectively. The trend features include monotonic features, nonlinear trend features and periodic features. Each trend eigenvalue is calculated by non-parametric statistical methods to construct the trend eigenvalue sequence of each compressor blade.

3. The turbofan engine compressor blade crack multi-modal monitoring system according to claim 2, characterized in that: The specific steps of the in-depth assessment also include: A clustering algorithm is used to cluster the trend characteristic value sequence of the compressor blades, and the compressor blade groups with trend deviation are identified and marked as deviated blades. The trend characteristic value sequences marked as abnormal blades and deviated blades are removed respectively, and the mean of the trend characteristic value sequences of the remaining compressor blades is calculated respectively to construct the baseline trend characteristic value sequence; The distance metric is used to calculate the difference between the trend characteristic value sequence of each deviated leaf and abnormal leaf and the baseline trend characteristic value sequence; A difference threshold is configured to determine whether the difference between the trend characteristic value sequence of the deviated blade or abnormal blade and the baseline trend characteristic value sequence is greater than the difference threshold. If so, it is marked as a potential crack blade.

4. The turbofan engine compressor blade crack multi-modal monitoring system according to claim 1, characterized in that: The specific steps of generating the incentive decision for the potential cracked blade include: The circumferential position coordinates of the potential crack blade and its difference degree are obtained in real time. The laser pulse energy is dynamically adjusted by combining the difference degree with the basic energy of the laser pulse. The natural frequency of the blade with potential cracks is obtained according to the real-time rotation speed, and the laser excitation frequency band is set with the natural frequency as the center; Select a transmitter according to the position of the compressor blade group where the potential cracked blade is located, and set the azimuth and pitch angles of the transmitter according to the relative position of the transmitter coordinates and the circumferential position coordinates of the potential cracked blade; The laser beam is emitted to the surface of the potential crack blade through the transmitter. When there are multiple potential crack blades in a group of compressor blades, parallel excitation is implemented, and the laser beam is emitted in time-sharing order according to the circumferential order of the potential crack blades.

5. The turbofan engine compressor blade crack multi-modal monitoring system according to claim 1, characterized in that: The specific steps of determining whether a potential crack blade has cracks include: Extract stress data from the multimodal data of the blade with potential cracks, and perform stress diagnosis on the blade group where the blade with potential cracks is located, including: Extract strain gradient eigenvalues based on stress data and configure strain thresholds. If the strain gradient eigenvalue of a blade with a potential crack exceeds the strain threshold, laser excitation is stopped and an excitation warning is triggered. Based on the principle of spatial symmetry, the position coordinates of the potential cracked blade in the compressor blade group are determined, and the control blade is selected according to the spatial layout structure of the compressor blade; Extracting vibration data features from multimodal data of blades with potential cracks and control blades, the vibration data features include vibration order amplitude and time-frequency spectrum features; Configure the order amplitude threshold. For the vibration order amplitude, calculate the relative deviation between the amplitude of the blade with potential crack and the vibration order of the control blade at the same order. If the relative deviation between the amplitude of the blade with potential crack and the vibration order of the control blade at the same order exceeds the order amplitude threshold, the blade with potential crack is judged to have a crack.

6. The crack multi-modal monitoring system for the compressor blade of a turbofan engine according to claim 5, characterized in that, The specific step of determining whether the potential crack blade has cracks also includes: Configure the principal component threshold , perform empirical mode decomposition on the vibration data of the potentially cracked blades and the control blades, decompose the vibration data into multiple intrinsic mode functions, and select the first intrinsic mode functions with the largest energy proportion to perform Hilbert transform to obtain the time-frequency spectrogram; The characteristic frequency of the time-frequency spectrum is identified by peak detection. On the time-frequency spectrum, the frequency bandwidth is set around the characteristic frequency, and the energy density of the characteristic frequency is calculated by integration. An energy deviation threshold is set to calculate the relative deviation of the energy density of the characteristic frequency of the blade with potential cracks and the control blade. If the deviation is greater than the energy deviation threshold, it is determined that the blade with potential cracks has cracks. When it is determined that a potential crack blade has cracks, the stimulation of the potential crack blade is stopped and a first-level blade warning is issued; Configure the excitation duration, monitor the continuous duration of exciting the potentially cracked blade. If it is equal to the excitation duration and no crack is determined in the potentially cracked blade, a secondary blade warning is issued.

7. The multi-modal monitoring system for cracks in the compressor blades of a turbofan engine according to claim 1, characterized in that, The specific steps for constructing the multi-modal feature vector include: Obtain the real-time rotational speed of the compressor, which is used to resample the vibration data in the multi-modal data to generate vibration data with equal angular intervals; and define the target order according to the number of compressor blades in each group. Define the resonance frequency band, filter the resampled vibration data to obtain the filtered vibration data, perform Hilbert transform on it to generate the orthogonal data of the vibration data, calculate the envelope data, and perform Fourier transform on the envelope data to obtain the envelope spectrum. Perform windowing on the resampled vibration data and then perform Fourier transform to obtain the spectrum of the resampled vibration data. According to the target order, map the frequency axes of the spectrum of the resampled vibration data and the envelope spectrum to the order axis, and use the interpolation algorithm to generate the corresponding order spectrum. Extract features from the order spectrum respectively, and combine with the time-domain features of the multi-modal data to splice and form the multi-modal feature vector of each compressor blade.

8. The turbofan engine compressor blade crack multi-modal monitoring system according to claim 1, characterized in that: The joint trigger rule includes: Within the monitoring threshold time range, count the number of key features that exceed the corresponding key feature threshold range. When the number of key features that exceed the key feature threshold range exceeds the preset number threshold, the first rule is satisfied. Calculate the relative deviation degree of each key feature value that exceeds the key feature threshold range with respect to its corresponding key feature threshold range. When there is a relative deviation degree that exceeds the preset deviation degree threshold, the second rule is satisfied. When both the first rule and the second rule are satisfied, that is, the number of key features that exceed the key feature threshold range reaches the preset number threshold and the relative deviation degree of at least one key feature reaches the preset deviation degree threshold, it is determined to trigger the depth assessment.

Images