Blade structure performance detection equipment and detection method

Abstract

The invention relates to the technical field of performance detection, and discloses a blade structure performance detection device and method, and the method comprises the steps: carrying out the laser ultrasonic scanning of a titanium alloy engine blade, and obtaining multi-mode spectrum feature data; performing parameter inversion and iterative optimization calculation according to the multi-modal spectrum characteristic data to obtain material parameter distribution data of a local area of the blade; performing annular scanning analysis on the titanium alloy engine blade based on the material parameter distribution data to obtain local structure feature data; performing modal decomposition on the local structure feature data to obtain fatigue damage accumulation data; and inputting the accumulated fatigue damage data into a dual decoupling diagnosis model for comparative learning and antagonistic feature separation to obtain a target diagnosis result, the method overcomes the limitation of a traditional linear superposition method, accurately predicts the fatigue damage evolution law under a complex load, effectively solves the problem of working condition interference, and improves the diagnosis accuracy. And the robustness and generalization of fault diagnosis are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of performance detection, and in particular to a blade structure performance detection device and a detection method. Background Art

[0002] Titanium alloy engine blades are key components of aircraft engines, and their structural performance testing is of great significance to ensure the safe operation of the engine. Although traditional ultrasonic testing methods have been widely used in non-destructive testing of blades, they face many challenges under actual working conditions. Most existing testing methods use contact measurement, which is difficult to meet the needs of in-service testing of engine blades, and the measurement efficiency is low, and rapid evaluation cannot be achieved.

[0003] Secondly, due to the complex structure of the blade and the spatial inhomogeneity of the material properties, it is difficult for the traditional single-mode ultrasonic detection method to simultaneously obtain the geometric dimensions and material property information of the blade. Especially when the blade morphology and material properties are uncertain, the existing methods are difficult to accurately identify tiny defects and are prone to missed detection and false detection. In addition, the engine blades are subjected to a complex load spectrum during service, resulting in a strong nonlinear characteristic in the accumulation law of fatigue damage, while the existing damage assessment method based on the principle of linear superposition is not accurate enough. Summary of the invention

[0004] The present invention provides a blade structure performance detection device and a detection method, which overcomes the limitations of the traditional linear superposition method, accurately predicts the evolution law of fatigue damage under complex loads, effectively solves the problem of working condition interference, and enhances the robustness and generalization of fault diagnosis.

[0005] In a first aspect, the present invention provides a method for detecting blade structure performance, the method comprising: Laser ultrasonic scanning is performed on titanium alloy engine blades to obtain multi-modal spectrum feature data; Performing parameter inversion and iterative optimization calculations according to the multi-modal spectrum characteristic data to obtain material parameter distribution data of a local area of ​​the blade; Performing annular scanning analysis on the titanium alloy engine blade based on the material parameter distribution data to obtain local structural feature data; Performing modal decomposition on the local structural characteristic data to obtain fatigue damage accumulation data; The fatigue damage accumulation data is input into a dual decoupling diagnosis model for contrast learning and adversarial feature separation to obtain a target diagnosis result.

[0006] In a second aspect, the present invention provides a blade structure performance detection device, the blade structure performance detection device comprising: A scanning module is used to perform laser ultrasonic scanning on titanium alloy engine blades to obtain multi-modal spectrum feature data; An inversion module, used to perform parameter inversion and iterative optimization calculation according to the multi-modal spectrum characteristic data to obtain material parameter distribution data of a local area of ​​the blade; An analysis module, used for performing an annular scanning analysis on the titanium alloy engine blade based on the material parameter distribution data to obtain local structural feature data; A decomposition module, used for performing modal decomposition on the local structural characteristic data to obtain fatigue damage accumulation data; The separation module is used to input the fatigue damage accumulation data into a dual decoupling diagnosis model for comparative learning and adversarial feature separation to obtain a target diagnosis result.

[0007] In the technical solution provided by the present invention, by adopting a non-contact laser ultrasonic detection system and combining a multi-modal spectrum feature analysis method, the purpose of rapid in-service detection is achieved, and the complete structural characteristic information of the blade is obtained at the same time. The signal processing method based on wavelet denoising and Fourier transform effectively extracts the resonance mode information of body waves and surface waves, and significantly reduces the influence of environmental noise. The improved Levenberg-Marquardt algorithm is used for parameter inversion, and the spatial distribution characteristics of material parameters are accurately identified through multi-starting point iterative optimization and regularization processing. The circular scanning strategy is innovatively introduced, and the anisotropic characteristics of the material are accurately characterized by analyzing the directional distribution of the surface wave energy flow density, thereby improving the spatial resolution of defect detection. The fatigue damage assessment method based on modal decomposition and nonlinear accumulation model overcomes the limitations of the traditional linear superposition method and accurately predicts the evolution law of fatigue damage under complex loads. A dual decoupling diagnosis model is designed, which effectively solves the problem of working condition interference through contrast learning and adversarial feature separation, and enhances the robustness and generalization of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0009] Figure 1 A schematic diagram of the steps of a blade structure performance detection method according to an embodiment of the present invention; Figure 2 It is a schematic diagram of the structure of a blade structure performance detection device in an embodiment of the present invention. DETAILED DESCRIPTION

[0010] An embodiment of the present invention provides a blade structure performance detection device and a detection method. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0011] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of a blade structure performance detection method in an embodiment of the present invention includes: Step S1, performing laser ultrasonic scanning on a titanium alloy engine blade to obtain multi-modal spectrum feature data; It is understandable that the execution subject of the present invention may be a blade structure performance detection device, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0012] Specifically, the laser parameters of the pulse laser transmitter are set to ensure the efficiency and accuracy of the laser ultrasonic scanning process. In the process of setting the laser parameters, key factors such as the laser pulse energy, wavelength, repetition frequency and spot size are considered to form a stable and efficient excitation source on the surface of the titanium alloy engine blade. After the setting is completed, the laser parameters are calibrated to eliminate the influence of environmental factors and equipment errors on the laser emission characteristics, and the calibrated laser emission parameters are obtained. Based on the calibrated laser emission parameters, a two-dimensional scanning galvanometer system is used to construct scanning grid data on the surface of the titanium alloy engine blade. The construction of the scanning grid needs to ensure uniform coverage of the entire blade surface, and consider the complex curved surface morphology of the blade. By optimizing the scanning path, the integrity and measurement accuracy of data acquisition are improved. According to the scanning grid data, the Doppler laser detector is controlled to receive the ultrasonic echo signal. The core of laser ultrasonic technology is that the laser pulse excites the ultrasonic wave on the blade surface and propagates through the material, and then the Doppler laser detector detects the echo signal. Due to the characteristics of high sensitivity and non-contact measurement, the Doppler laser detector can capture the internal structure information of the blade and obtain ultrasonic echo data. During the signal receiving process, the incident angle of the detector, the laser power, and the signal amplification processing are considered to ensure the integrity and signal-to-noise ratio of the echo signal. The ultrasonic echo data is collected. During the collection process, the data acquisition system records the time domain waveform data at a high sampling rate to ensure that the time resolution of the ultrasonic signal is high enough to capture the tiny structural changes inside the material. The time domain waveform data is temperature compensated by the temperature sensor. The temperature compensation process involves real-time monitoring of the temperature change on the blade surface, and correcting the time domain waveform data according to the temperature-ultrasonic velocity correction model to eliminate the influence of temperature change on the ultrasonic propagation velocity, and obtaining the temperature compensated time domain waveform data. The laser focal length correction is performed on the temperature compensated time domain waveform data based on the high-precision displacement sensor. The displacement sensor monitors the distance between the laser focus and the blade surface in real time, and adjusts the time domain waveform data through the focal length correction algorithm to ensure the measurement accuracy of the ultrasonic signal and obtain the original ultrasonic signal. The original ultrasonic signal is processed to extract its key features. In signal processing, the wavelet denoising method is used to suppress the noise in the signal. The wavelet denoising method improves the signal-to-noise ratio by decomposing the original signal into different frequency levels and removing the high-frequency noise components using soft threshold or hard threshold methods. After completing the wavelet denoising, the signal is Fourier transformed to convert the time domain signal into the frequency domain to obtain the spectral characteristics of the internal structure of the blade. During the Fourier transform process, different propagation modes of ultrasonic signals (such as longitudinal waves, shear waves, and surface waves) show different spectral characteristics in the frequency domain, so the spectral analysis method is used to extract the characteristic information of different modes. Through the signal processing process, multimodal spectral feature data is obtained to reflect the internal microstructure, stress distribution, and microscopic damage of the blade.

[0013] The original ultrasonic signal is decomposed by wavelet to obtain signal details at different scales. The process of wavelet decomposition is to decompose the original signal into wavelet coefficients of different frequency bands through multi-scale analysis, so as to extract high-frequency noise and useful signal components respectively. After wavelet decomposition, a set of wavelet coefficients is obtained to characterize the changing characteristics of the signal at different scales. The wavelet coefficients are soft-thresholded to eliminate high-frequency noise while retaining the main components of the signal. When determining the threshold, the standard deviation of each layer of wavelet coefficients is calculated, and an appropriate threshold is selected based on statistical methods to effectively distinguish between signals and noise. The key to soft threshold processing is to perform nonlinear compression on the wavelet coefficients so that the coefficients below the threshold tend to zero, while the coefficients above the threshold are scaled to reduce noise interference and maintain the main characteristics of the signal. After soft threshold processing, the denoised wavelet coefficients are obtained. The denoised wavelet coefficients are reconstructed by wavelet to restore the denoised time domain signal. The process of wavelet reconstruction is the inverse process of wavelet transform. Wavelet coefficients of different scales are combined to form a new time domain signal. At the same time, due to the removal of high-frequency noise, the time domain signal is smoother than the original signal and retains the ultrasonic propagation characteristics inside the blade. The denoised time domain signal is input into the fast Fourier transform module to convert it to the frequency domain to obtain spectrum data. The function of fast Fourier transform is to convert the oscillation mode of the time domain signal into frequency components in order to identify the internal resonance characteristics of the blade material. Through Fourier transform, the ultrasonic propagation characteristics of different modes inside the blade are obtained, and key information is extracted from the spectrum. The spectrum data is bandpass filtered to extract specific resonance modes. For the body wave thickness resonance mode, an appropriate frequency range is selected for bandpass filtering to retain the longitudinal and transverse wave resonance information propagating inside the blade material, while suppressing the interference of non-target frequency components. The frequency components of the surface wave resonance mode are extracted using the two-dimensional spatial mapping method of the spectrum to obtain the propagation characteristics of the surface wave. The surface wave resonance mode has different propagation characteristics from the body wave resonance mode. The former reflects the surface state of the material, while the latter provides detailed information on the material characteristics inside the blade. Through bandpass filtering and spatial mapping, multimodal resonance data of body waves and surface waves are obtained respectively. These data contain the material properties, stress distribution and existing defect information inside the blade. Based on the multimodal resonance data, the fundamental frequency, second harmonic and third harmonic frequencies are extracted to analyze the nonlinear characteristics of the internal structure of the blade. The key to harmonic analysis is to calculate the amplitude ratio of each harmonic component to evaluate the nonlinear response characteristics in the ultrasonic signal, and to calculate the quality factor to measure the energy attenuation characteristics of the resonance mode. The amplitude ratio reflects the nonlinear behavior of the material, while the quality factor is used to evaluate the stability and loss characteristics of the resonance mode. Multimodal spectrum feature data is obtained by comprehensive analysis of the multimodal resonance data.

[0014] Step S2, performing parameter inversion and iterative optimization calculations according to the multi-modal spectrum characteristic data to obtain material parameter distribution data of the local area of ​​the blade; Specifically, a theoretical model considering material anisotropy is constructed based on multimodal spectrum feature data, so that the model fully describes the complex material properties inside the blade. When constructing the model, the elastic modulus, density and Poisson's ratio are set as optimization variables to ensure that the model fully reflects the mechanical properties and dynamic response characteristics of the internal materials of the blade. On the basis of the initial parameter setting, an initial parameter model is formed. Since the material properties are affected by temperature changes, the initial parameter model is temperature corrected to improve the accuracy of parameter inversion. The temperature data of different regions of the blade are collected by a temperature sensor, and the correction coefficient is calculated based on the established temperature-material parameter change relationship to adjust the initial parameter model so that it accurately reflects the state of the material under actual service conditions. After temperature correction, the corrected parameter model is obtained. The temperature-corrected parameter model is input into the Levenberg-Marquardt optimizer, which combines the advantages of the gradient descent method and the Newton method to effectively improve the convergence speed and ensure the stability of the inversion calculation. In the optimization process, the objective function is set to measure the error between the theoretical calculation value and the experimental measurement value. The objective function adopts the weighted mean square error calculation method, so that the optimization process fully considers the contribution of different measurement data to the final result and minimizes the impact of measurement noise and uncertainty on the optimization result. Through this optimization strategy, the parameter model is gradually adjusted under different constraints to improve the accuracy of material parameter inversion. In order to improve the stability of the optimization calculation and prevent parameter overfitting, a regularization term is introduced into the optimization objective function to improve the numerical characteristics of parameter estimation. The introduction of the regularization term effectively suppresses the drastic fluctuation of the parameters and avoids the optimization calculation from falling into the local optimal solution. In order to reasonably determine the regularization coefficient, the L curve method is used for analysis. The L curve method is a parameter selection method for regularization problems. It finds a balance between the regularization strength and the fitting error, so that the optimization process can fully utilize the measurement data and maintain the rationality of the results. Through this method, the regularized optimization function is obtained, so that the optimization calculation achieves the best balance between stability and accuracy. Based on the regularized optimization function, multi-starting point iterative calculation is performed, so that the optimization process jumps out of the local extreme value and converges to the global optimal solution. The multi-starting point iterative calculation starts from different initial parameters, performs optimization calculations multiple times, and adjusts the search direction during each iteration, so that the obtained optimization parameter data is more stable and physically reasonable. Through multiple iterative optimizations, the robustness of parameter inversion is effectively improved, the insensitivity of the calculation results to noise and measurement errors is ensured, and the convergence accuracy of the optimization calculation is improved. The optimization parameter data is spatially mapped to reflect the distribution of material parameters in the local area of ​​the blade. The elastic modulus, density, and Poisson's ratio data obtained by optimization are mapped to the three-dimensional geometric model of the blade, and the discrete optimization results are smoothed using an interpolation algorithm to generate a high-resolution material parameter distribution map.Through the mapping process, the elastic modulus distribution map, density distribution map and Poisson's ratio distribution map of the local area of ​​the blade are obtained.

[0015] Step S3, performing annular scanning analysis on the titanium alloy engine blade based on the material parameter distribution data to obtain local structural feature data; Specifically, appropriate scanning parameters are determined based on the material parameter distribution data to ensure the accuracy and effectiveness of the annular scanning analysis. In the process of determining the scanning parameters, the elastic modulus distribution, density change and spatial distribution characteristics of the Poisson's ratio of the material are comprehensively considered to optimize the scanning path so that the scanning points cover the key areas of the blade surface and minimize the measurement error. On this basis, annular scanning path data is constructed around each scanning point so that the ultrasonic signal propagates in different directions to obtain structural response information. The construction of the annular scanning path needs to be combined with the complex geometric shape of the blade and adopt a non-uniform grid distribution to adapt to the curvature change of the blade surface and ensure the stability of signal propagation. Surface wave measurement is performed on the annular scanning path data to analyze the attenuation characteristics of the surface wave amplitude in different propagation directions. When the surface wave propagates on the blade surface, its amplitude will gradually decay with the increase of the propagation distance, and the attenuation rate is closely related to the microstructure, stress state and anisotropic characteristics of the material. In order to obtain accurate directional attenuation data, the amplitude of the surface wave is measured at different azimuth angles, and the measurement noise is removed by signal processing methods to ensure the reliability of the data. By fitting the attenuation curve of the surface wave amplitude, the attenuation coefficients in different propagation directions are extracted to form directional attenuation data. The directional distribution of the surface wave energy flux density is calculated based on the directional attenuation data to reveal the anisotropic characteristics of the material. The directional distribution of the surface wave energy flux density is derived by analyzing the wave velocity, attenuation characteristics and amplitude changes to establish the directional characteristics of the internal structure of the material. In order to extract the main directional information from the complex data, the principal component analysis method is used to extract the most important directional components in a dimensionality reduction manner, and the anisotropic data of the material is characterized by calculating the principal axis direction. The correlation analysis of the material anisotropic data of adjacent annular regions is performed to evaluate the structural uniformity between different regions of the blade. By calculating the correlation coefficient matrix between adjacent scanning regions, the uniformity of the material at different positions is quantified, and the regions with microstructural changes are identified. The calculation of the correlation coefficient matrix is ​​based on multiple physical parameters such as surface wave attenuation characteristics and energy flux density distribution to ensure the comprehensiveness of the uniformity analysis. A higher correlation coefficient indicates that the material is more uniform in the region, while a lower correlation coefficient means changes in the microstructure, such as uneven grain orientation, residual stress concentration or potential fatigue damage areas. Based on the structural uniformity data, a three-dimensional structural model of the local area is established. The construction of the three-dimensional structural model needs to comprehensively consider key parameters such as thickness distribution and grain orientation distribution to fully characterize the microstructure state of the blade. In the process of thickness distribution modeling, the propagation time information of the ultrasonic signal is combined to reconstruct the geometric morphology of the local area of ​​the blade, and the measured data is smoothed by the interpolation algorithm to improve the accuracy of the model. At the same time, in the modeling process of grain orientation distribution, the material anisotropy data is combined, and the direction cosine matrix is ​​used to calculate the grain orientation angle distribution to reflect the microstructural differences of the blade in different areas.By combining these model parameters, 3D structural model data is generated. The abnormal areas detected in the 3D structural model data are finely scanned to improve the resolution of local structural features. During the fine scanning process, a higher spatial resolution is used for the abnormal areas, and the detection sensitivity of abnormal signals is enhanced through improved signal processing algorithms. Fine scanning can identify tiny structural defects and quantify their impact range. Local structural feature data is obtained through annular scanning analysis and data processing.

[0016] Step S4, performing modal decomposition on the local structural characteristic data to obtain fatigue damage accumulation data; Specifically, first. Perform modal decomposition on the local structural feature data to extract the vibration feature information of the blade structure in different modes. In the process of modal decomposition, the empirical mode decomposition or variational mode decomposition method is used to decompose the original data into a series of intrinsic mode functions, each mode corresponds to a specific vibration frequency and energy distribution, thereby forming modal feature data. Reveal the dynamic response characteristics of the blade under different excitation conditions. Perform generalized Fourier series expansion on the modal feature data to construct a frequency response function. Generalized Fourier series expansion converts complex time domain signals to the frequency domain to reveal the contribution of different frequency components to the vibration response of the blade. Through this expansion, the frequency response function under different modes is calculated. Based on the frequency response function, the kurtosis and skewness of the response signal are analyzed to evaluate the non-Gaussian characteristics of the vibration signal, and the nonlinear coefficient is constructed using the piecewise linearization method. Kurtosis is used to measure the sharpness of the signal distribution, while skewness is used to measure the symmetry of the signal. In order to improve the calculation accuracy, the piecewise linearization method is used to model the nonlinear characteristics, and the best fitting curve is solved by the numerical optimization method to obtain the correction coefficient data. The correction coefficient data is used to describe the nonlinear dynamic characteristics of the blade under different modes and to modify the traditional linear frequency response model to improve its applicability. The correction coefficient data is convolved with the frequency response function to calculate the stress spectrum density function under random vibration load. The purpose of the convolution operation is to comprehensively consider the nonlinear dynamic characteristics of the blade under different modes and calculate its stress distribution under complex load conditions. Through this calculation process, the stress spectrum density function corresponding to different vibration modes is obtained to quantify the stress fluctuation characteristics of the blade under random excitation conditions and to predict its stress response during long-term service. The stress spectrum density function is characterized by eight parameters to extract the main statistical characteristics of the stress spectrum, including mean, standard deviation, peak factor, skewness, kurtosis, energy distribution, frequency center and bandwidth factor. These parameters are used to quantify the stress distribution characteristics of the blade under complex load conditions. The rain flow counting method is used to statistically analyze the amplitude distribution of the stress cycle to identify the cyclic stress characteristics of the blade under different load conditions. The rain flow counting method is a statistical method used for fatigue analysis, which can identify and count the amplitude, period and mean of the stress cycle to form stress cycle data. Based on stress cycle data, a nonlinear accumulation model is used to calculate fatigue damage values ​​to evaluate the fatigue life of blades. Based on the traditional linear damage accumulation model (such as the Miner criterion), the nonlinear accumulation model more accurately describes the damage accumulation process of blades under complex stress sequences and considers the mutual influence between different stress levels. In the calculation process, the load sequence effect coefficient is used as a correction parameter to adjust the calculation results of the nonlinear accumulation model and improve the accuracy of fatigue life prediction. Fatigue damage accumulation data is obtained.

[0017] Step S5: input the fatigue damage accumulation data into the dual decoupling diagnosis model for comparative learning and adversarial feature separation to obtain the target diagnosis result.

[0018] Specifically, the fatigue damage accumulation data is standardized to eliminate the influence of different data scales and ensure that the model effectively captures the inherent characteristics of the data. The contrast decoupling module extracts discriminative features from the fatigue damage data by constructing contrast relationships, so that the data under different damage modes form a more obvious distinction in the feature space. Under the action of the contrast decoupling module, the fatigue damage accumulation data undergoes nonlinear transformation to extract contrast feature data, and establishes a preliminary damage mode distinction in the high-dimensional feature space. After obtaining the contrast feature data, in order to improve the generalization ability of the model and enhance the robustness of the features, the data is enhanced. By performing random rotation, translation and scaling operations on the feature data, a series of enhanced samples are constructed to improve the adaptability of the model when facing different working conditions. These enhanced samples expand the distribution range of the data, allowing the model to learn richer damage features while avoiding the occurrence of overfitting problems. In the data enhancement process, the rotation operation simulates the changes in the damage characteristics of the blade at different angles, while the translation operation simulates the influence of the small structural differences on the blade surface on fatigue damage, and the scaling operation is used to simulate the differences in damage scales to ensure that the generalization ability of the model is improved. After this data enhancement process, enhanced feature data is obtained. Based on contrast feature data and enhanced feature data, the model is optimized by contrast learning method, so that it can effectively perceive different working conditions of the blade. By minimizing the feature distance, the features under the same working condition are made as close as possible, while the features under different working conditions maintain a certain degree of distinction, and a working condition perception mask is generated. The function of the working condition perception mask is to eliminate the non-damage features caused by the change of working conditions, so that the focus of the model is on the real fatigue damage features, and the working condition feature data obtained effectively characterizes the fatigue damage mode of the blade under different service conditions, while avoiding the interference of working condition changes on the diagnosis results. The working condition feature data is input into the adversarial mask decoupling module in the dual decoupling diagnosis model, and the generative adversarial network is used for feature separation. The generative adversarial network consists of a generator and a discriminator, in which the goal of the generator is to generate virtual samples close to the real damage features, while the goal of the discriminator is to distinguish between the generated features and the real features. Through continuous adversarial training, the generator captures the key information in the fatigue damage features and generates high-quality separation feature data. In the architecture of the generative adversarial network, a four-layer convolutional neural network is used for feature extraction, so that the model can learn the damage features at different scales layer by layer, and improve the recognition ability of complex damage patterns. After this process, separated feature data is obtained. BatchNormalization and Dropout operations are performed on the separated feature data. The role of BatchNormalization is to normalize each batch of training data to make the data distribution more stable, thereby accelerating the convergence of the model and reducing the problem of internal covariate shift.The Dropout operation is used to randomly discard some neurons to reduce the model's dependence on specific features, thereby improving the model's generalization ability. After BatchNormalization and Dropout processing, the fault feature data is obtained. The fault feature data is input into the support vector machine classifier, and the radial basis kernel function is used in the support vector machine classifier to ensure that the classifier constructs a complex decision boundary in a high-dimensional space and improves the accuracy of classification. During the support vector machine training process, the kernel parameters are optimized through cross-validation, and the characteristics of the support vector are used to enable the classifier to accurately distinguish different types of fatigue damage. The target diagnosis result of the blade is obtained.

[0019] In the embodiment of the present invention, by adopting a non-contact laser ultrasonic detection system and combining a multi-modal spectrum feature analysis method, the purpose of rapid in-service detection is achieved, and the complete structural characteristic information of the blade is obtained at the same time. The signal processing method based on wavelet denoising and Fourier transform effectively extracts the resonance mode information of body waves and surface waves, and significantly reduces the influence of environmental noise. The improved Levenberg-Marquardt algorithm is used for parameter inversion, and the spatial distribution characteristics of material parameters are accurately identified through multi-starting point iterative optimization and regularization processing. The circular scanning strategy is innovatively introduced to accurately characterize the anisotropic characteristics of the material by analyzing the directional distribution of the surface wave energy flow density, thereby improving the spatial resolution of defect detection. The fatigue damage assessment method based on modal decomposition and nonlinear accumulation model overcomes the limitations of the traditional linear superposition method and accurately predicts the evolution law of fatigue damage under complex loads. A dual decoupling diagnosis model is designed, which effectively solves the problem of working condition interference through contrast learning and adversarial feature separation, and enhances the robustness and generalization of fault diagnosis.

[0020] In a specific embodiment, the process of executing step S1 may specifically include the following steps: The laser parameters of the pulse laser emitter are set to obtain calibrated laser emission parameters, and based on the calibrated laser emission parameters, scanning grid data is constructed on the surface of the titanium alloy engine blade through a two-dimensional scanning galvanometer system; Controlling the Doppler laser detector to receive ultrasonic echo signals according to the scanning grid data to obtain ultrasonic echo data; Acquiring ultrasonic echo data to obtain time domain waveform data, and performing temperature compensation on the time domain waveform data through a temperature sensor to obtain temperature compensated time domain waveform data; Based on the high-precision displacement sensor, the laser focal length is corrected on the time-domain waveform data after temperature compensation to obtain the original ultrasonic signal; The original ultrasonic signal is subjected to wavelet denoising and Fourier transform to obtain multimodal spectrum feature data.

[0021] Specifically, the laser parameters of the pulsed laser transmitter are set to ensure the stability of ultrasonic excitation and measurement accuracy. During the setting process, key parameters include laser pulse energy , pulse width , repetition frequency and spot diameter Among them, the laser pulse energy Affects the intensity and pulse width of the excitation ultrasonic wave Determines the frequency characteristics of ultrasound, and the repetition frequency The spot diameter is related to the time resolution of the measurement. The size of the excitation area is determined. After setting these parameters, they are calibrated to compensate for the influence of equipment errors and environmental factors, and the calibrated laser emission parameters are obtained. Based on the calibrated laser emission parameters, a two-dimensional scanning galvanometer system is used to construct scanning grid data on the surface of the titanium alloy engine blade. The establishment of the scanning grid needs to consider the surface morphology of the blade, define the spatial distribution of the scanning points, and ensure that the data covers the entire detection area. Set the scanning step length and scanning range After that, the number of grid points is given by: Calculated, where Indicates the total number of grid points. During the scanning process, each grid point serves as the action position of the laser pulse. The laser-induced ultrasonic wave propagates inside the material and receives the echo signal through the Doppler laser detector. After receiving the ultrasonic echo signal, data acquisition is performed to obtain the time domain waveform data. The propagation time of the ultrasonic wave in the material Depend on: Calculated, where is the propagation depth inside the blade material, is the ultrasonic propagation velocity in the material, which is related to the elastic modulus of the material. and density Related, by: The propagation characteristics of the ultrasonic signal are analyzed to obtain the internal structure information of the blade. Temperature changes will affect the propagation speed of ultrasonic waves, so a temperature sensor is used to perform temperature compensation on the time domain waveform data. Assuming the temperature change is , then the correction amount of propagation speed is pass: Calculated, where is the temperature sensitivity coefficient of the material. The propagation time after temperature compensation is: Through this correction, the time domain waveform data after temperature compensation is obtained. The laser focal length is corrected based on the high-precision displacement sensor. This will cause the propagation path of the ultrasonic signal to change, so the propagation time is corrected according to the focal length error. Assuming the actual focal length error is , then the corrected propagation time is: Through focal length correction, a more accurate original ultrasonic signal is obtained. The original ultrasonic signal is processed to extract key features. The wavelet denoising method is used to smooth the signal and remove high-frequency noise. Assume that the original signal By signal component and noise Composition, namely: Wavelet coefficients of different scales are obtained through wavelet transform. Let the wavelet coefficients be , then after wavelet threshold processing, the denoised signal is obtained: in is the wavelet basis function. Through the inverse wavelet transform, the denoised time domain signal is restored. The denoised signal is subjected to Fourier transform to obtain multimodal spectrum feature data. The mathematical expression of Fourier transform is: in is the spectrum function, The Fourier transform result is used to extract the different modal information of the ultrasonic signal, such as the propagation characteristics of longitudinal waves, shear waves and surface waves, thereby obtaining multi-modal spectrum feature data.

[0022] In a specific embodiment, the process of performing wavelet denoising and Fourier transform on the original ultrasonic signal to obtain multimodal spectrum feature data may specifically include the following steps: Perform wavelet decomposition on the original ultrasonic signal to obtain wavelet coefficients; Perform soft threshold processing on the wavelet coefficients, determine the threshold by calculating the standard deviation of each layer of wavelet coefficients, and obtain the denoised wavelet coefficients; Perform wavelet reconstruction on the denoised wavelet coefficients to obtain the denoised time domain signal, and input the denoised time domain signal into the fast Fourier transform module to obtain spectrum data; Bandpass filtering is performed on the spectrum data to extract the frequency components of the body wave thickness resonance mode, and at the same time, the spectrum is mapped in two-dimensional space to obtain the surface wave resonance mode and obtain multi-modal resonance data; The fundamental frequency, second harmonic and third harmonic frequencies are extracted based on the multimodal resonance data, and the amplitude ratio and quality factor of each harmonic component are calculated to obtain the multimodal spectrum characteristic data.

[0023] Specifically, the original ultrasonic signal is subjected to wavelet decomposition to extract signal components at different scales. Wavelet decomposition uses a set of orthogonal wavelet basis functions to expand the original signal at multiple scales so that the high-frequency and low-frequency components of the signal are represented separately. Suppose the original ultrasonic signal is , then the mathematical expression of wavelet decomposition is: in, is the wavelet coefficient, indicating the signal at scale and location The contribution of is the wavelet basis function, Indicates the maximum number of decomposition layers. Wavelet decomposition splits the original signal into components of different frequency levels, so that high-frequency noise and useful signals can be separated at different scales. Soft threshold processing is performed on the wavelet coefficients to eliminate the interference of high-frequency noise. The key to soft threshold processing is to set a suitable threshold and use the threshold function to correct the wavelet coefficients. Suppose the standard deviation of the wavelet coefficients is , then the threshold of each layer Determined by the following formula: in, is the number of sampling points of the signal. For coefficients less than the threshold, they are set to zero, and for coefficients greater than the threshold, they are adjusted according to the following formula: in, is the wavelet coefficient after denoising, is a sign function used to preserve the original direction of the signal. Through soft threshold processing, high-frequency noise is effectively suppressed, while low-frequency signals remain basically unchanged, thereby improving the signal-to-noise ratio of the signal. The denoised wavelet coefficients are reconstructed by wavelet to restore the denoised time domain signal. Wavelet reconstruction is the inverse operation of wavelet transform, that is: in, is the denoised time domain signal. After wavelet reconstruction, the main structure of the time domain signal is retained, while the high-frequency noise is effectively filtered out to form a smoother ultrasonic waveform. The denoised time domain signal is input into the fast Fourier transform module to obtain the spectrum data of the signal. The mathematical expression of Fourier transform is: in, is the spectrum data, is the angular frequency, is an imaginary unit. Fast Fourier transform converts the time domain signal into the frequency domain signal, so that different vibration modes of the ultrasonic signal can be distinguished. Bandpass filtering is performed on the spectrum data to extract the frequency components of the body wave thickness resonance mode. The frequency corresponding to the body wave thickness resonance mode Calculated by the following formula: in, is the ultrasonic propagation velocity of the material, is the thickness of the blade. The role of bandpass filtering is to retain The frequency components near the surface are selected while suppressing other irrelevant components to ensure that the signal of the thickness resonance mode is clearly visible. At the same time, the spectrum is mapped in two dimensions to extract the surface wave resonance mode. The propagation characteristics of the surface wave are related to the elastic modulus of the material. and Poisson's ratio Related, its resonant frequency It is approximately calculated by the following relationship: The role of two-dimensional spatial mapping is to visualize the surface wave resonance modes at different spatial positions to form multimodal resonance data. Based on the multimodal resonance data, the fundamental frequency, second harmonic and third harmonic frequencies are extracted, and the amplitude ratio and quality factor of each harmonic component are calculated. , Second Harmonic and the third harmonic The following relations are satisfied: Harmonic Amplitude Ratio It is used to measure the energy distribution between different harmonics. The calculation formula is: in, Indicates the frequency The quality factor It is used to measure the energy attenuation characteristics of the resonance mode, and its calculation formula is: in, is the resonant frequency, is the half-power bandwidth of the resonance peak. The higher the quality factor, the more stable the resonance mode of the signal and the smaller the loss.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: A theoretical model considering material anisotropy is constructed based on multimodal spectrum feature data, and the elastic modulus, density, and Poisson's ratio are set as optimization variables to obtain an initial parameter model. Perform temperature correction on the initial parameter model, calculate the correction coefficient according to the temperature data collected by the temperature sensor, and obtain the parameter model after temperature correction; The temperature-corrected parameter model is input into the Levenberg-Marquardt optimizer, and the objective function is set as the weighted mean square error between the theoretical calculation value and the experimental measurement value to obtain the optimization objective function; A regularization term is introduced into the optimization objective function, and the regularization coefficient is determined by the L-curve method to obtain a regularized optimization function. Perform multi-starting point iterative calculation based on the regularized optimization function to obtain optimization parameter data; The optimized parameter data are spatially mapped to generate spatial distribution maps of elastic modulus distribution, density distribution and Poisson's ratio distribution, and the material parameter distribution data of the local area of ​​the blade are obtained.

[0025] Specifically, a theoretical model considering material anisotropy is constructed based on multimodal spectrum feature data to describe the mechanical properties inside the blade. In this model, the elastic modulus is set ,density and Poisson's ratio As optimization variables, to characterize the mechanical response of the material. The anisotropy of the blade material is represented by the stiffness matrix To describe, the relationship satisfies: in, represents the stress component, is the strain component, is the anisotropic stiffness matrix, whose elements are determined by the optimization variables , and Calculated using different material models. According to the physical mechanism of ultrasonic signal propagation, the ultrasonic velocity The relationship with elastic modulus and density is: Through this relationship, the preliminary elastic modulus, density and Poisson's ratio are estimated according to the multimodal spectrum characteristic data to obtain the initial parameter model. Since the physical properties of the material will change under different temperature conditions, the initial parameter model is corrected for temperature. The temperature data of the blade surface is collected through the temperature sensor , calculate the correction factor of material parameters , so that the elastic modulus, density and Poisson's ratio are corrected with temperature. Among them, the temperature correction formula of the elastic modulus is expressed as: in, is the initial elastic modulus, is the temperature correction coefficient of the elastic modulus, represents the temperature change. Similarly, the temperature corrections for density and Poisson's ratio are: in, and are the temperature correction coefficients for density and Poisson's ratio, respectively. After temperature correction, an updated parameter model is obtained, which accurately describes the physical characteristics of the blade under different service environments. The temperature-corrected parameter model is input into the Levenberg-Marquardt optimizer to solve the optimal solution for the material parameters. The Levenberg-Marquardt optimization method is a nonlinear least squares optimization method that combines the gradient descent method and the Newton method. Its optimization objective function is defined as the theoretical calculated value Compared with experimental measurements The weighted mean square error between , that is: in, represents the parameter vector to be optimized, is the weight coefficient used to balance the contribution of different measurement data, Represents the number of measurement data. The Levenberg-Marquardt method dynamically adjusts the step size during the optimization process, so that the optimization process can converge quickly and avoid the problem of local optimal solution. In order to improve the stability of the optimization calculation, a regularization term is introduced into the optimization objective function to prevent parameter overfitting. The objective function of regularized optimization is expressed as: in, is the regularization term, is the regularization coefficient. To determine the optimal regularization coefficient, the L-curve method is used for calculation. The L-curve method is used to draw and The relationship curve between and selects the inflection point corresponding to As the optimal regularization coefficient, it ensures that the optimization process neither overfits nor loses too much information. Based on the regularized optimization function, multi-starting point iterative calculation is performed to ensure that the optimization parameters converge to the global optimal solution. The strategy of multi-starting point iterative calculation is to start the optimization from multiple randomly initialized parameters, calculate the objective function values ​​at different starting points respectively, and select the best converged parameters as the final solution. This method can effectively prevent the optimization from falling into local minima and improve the stability of parameter estimation. The optimization parameter data is spatially distributed to visualize the changes in material parameters in the local area of ​​the blade. Assume that the elastic modulus distribution obtained by optimization is , the density distribution is , Poisson's ratio distribution is , the interpolation method is used to generate the corresponding spatial distribution map. The two-dimensional interpolation method is used to smooth the optimization data to ensure the continuity of the data, and the distribution of different material parameters is visualized through color mapping technology to form the blade material parameter distribution data.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Determine scanning parameters based on material parameter distribution data, and construct annular scanning path data around each scanning point; Perform surface wave measurement on the circular scanning path data, analyze the attenuation characteristics of the surface wave amplitude in different directions, and obtain directional attenuation data; The directional distribution of surface wave energy flux density is calculated based on the directional attenuation data, and the principal axis direction is extracted through principal component analysis to obtain the material anisotropy data; Conduct correlation analysis of adjacent annular regions on the material anisotropy data, calculate the correlation coefficient matrix, and obtain the structural uniformity data; A three-dimensional structural model of a local area is established based on the structural uniformity data, and model parameters including thickness distribution and grain orientation distribution are constructed to obtain three-dimensional structural model data; The abnormal area detected in the three-dimensional structure model data is scanned in detail to obtain local structure feature data.

[0027] Specifically, the scanning parameters are determined based on the material parameter distribution data to ensure the optimization of the scanning path and the efficiency of data acquisition. In the process of determining the scanning parameters, according to the elastic modulus of the blade ,density and Poisson's ratio Make adaptive adjustments to ensure the coverage and accuracy of the scan. Suppose the spacing of the scan points is , the total length of the scan area is , then the total number of scan points can be obtained by: Calculated, where Indicates the total number of scan points. Around each scan point, a set radius Construct a circular scanning path, the number of points in the circular path By angle resolution Determine, that is During the circular scanning process, the laser ultrasonic pulse will excite the surface waves in different directions, and then the echo signal of the ultrasonic wave will be recorded by the Doppler laser detector. The surface wave measurement is performed on the circular scanning path data to analyze the attenuation characteristics of the surface wave amplitude in different directions. Assume that the surface wave amplitude is ,in is the scanning direction angle, the attenuation model of the surface wave is expressed as: in, is the initial amplitude, is the direction-dependent attenuation coefficient, is the propagation distance. By fitting the amplitude changes in different directions, the anisotropic attenuation characteristics are calculated to form directional attenuation data. Based on the directional attenuation data, the directional distribution of the surface wave energy flow density is calculated. The energy flow density of the surface wave With amplitude and wave speed Related, its expression is in, is the material density, is the surface wave propagation velocity in this direction. In order to extract the main direction information, principal component analysis is performed on the energy flux density data. Assume that the energy flux density data is a matrix , its covariance matrix is ​​expressed as: By calculating the eigenvalues ​​and eigenvectors of the covariance matrix, the principal axis direction is obtained, and the anisotropic data of the material is obtained. The correlation analysis of adjacent annular regions of the material anisotropic data is performed to evaluate the uniformity of the structure. For two adjacent scanning areas and , and its correlation coefficient is calculated as in, and are the energy flow density data of the two regions respectively, and are their respective means. Correlation coefficient matrix It is used to quantify the structural uniformity between different regions. A higher correlation indicates that the material is uniform, while a lower correlation means changes in the microstructure. Based on the structural uniformity data, a three-dimensional structural model of the local area is established to characterize the material distribution characteristics inside the blade. In the three-dimensional structure modeling process, a thickness distribution model is constructed. Assume that the thickness distribution of the blade is , the thickness value of any point is calculated by interpolation method. Secondly, the grain orientation distribution model is established, and the grain orientation angle is set to , then the phase delay of the ultrasonic wave is calculated as: in, is the wavelength of ultrasound, is the phase offset. Through 3D grid interpolation, the thickness distribution and grain orientation distribution are combined to form 3D structure model data. The abnormal area detected in the 3D structure model data is scanned finely to improve the resolution and obtain local structural features. The determination of the abnormal area is set by the correlation threshold, that is, when the correlation coefficient of a certain area is Below the set threshold When the area is marked as abnormal, a higher scanning resolution is used for these abnormal areas. Perform a refinement scan, where is the original scan step length half of , that is: Through fine scanning, higher-precision local material parameter distribution can be obtained, and potential structural defects such as microcracks and stress concentration areas can be analyzed.

[0028] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Performing modal decomposition on local structural feature data to obtain modal feature data; Perform generalized Fourier series expansion on the modal characteristic data to obtain the frequency response function; The kurtosis and skewness of the response signal are calculated based on the frequency response function, and the nonlinear coefficient is constructed by the piecewise linearization method to obtain the correction coefficient data; The correction coefficient data is convolved with the frequency response function to obtain the stress spectrum density function under random vibration load; The stress spectrum density function is characterized by eight parameters, and the amplitude distribution of stress cycles is statistically analyzed by rain flow counting method to obtain stress cycle data. Based on the stress cycle data, the nonlinear accumulation model is used to calculate the fatigue damage value, and the load sequence effect coefficient is set as the correction parameter to obtain the fatigue damage accumulation data.

[0029] Specifically, the local structural feature data is modally decomposed to extract the structural vibration characteristics under different modes. In the process of modal decomposition, the original data is transformed into Decomposed into a series of intrinsic mode functions (IMF), each mode corresponds to a specific vibration frequency and energy distribution, forming modal characteristic data. The empirical mode decomposition method is based on the local extreme points of the signal, gradually extracting the vibration components on different time scales, so that the decomposed modal data effectively reflects the dynamic characteristics of the blade structure. Its mathematical expression is: in, Indicates the decomposition of The intrinsic modes, is the total number of modes, is the residual signal. Through modal decomposition, the response data of the blade under different vibration modes are obtained. The modal characteristic data is expanded by generalized Fourier series to construct the frequency response function. The complex time domain signal is converted to the frequency domain to reveal the contribution of different frequency components to the blade vibration response. For the modal characteristic data , its Fourier series expansion is expressed as: in, and are the cosine and sine coefficients of Fourier expansion, is the angular frequency. Based on Fourier expansion, the frequency response function is calculated , which is defined as: in, is the output spectrum of the system, is the input excitation spectrum. The frequency response function describes the vibration characteristics of the blade at different frequencies. After obtaining the frequency response function, the kurtosis and skewness of the response signal are calculated to evaluate the non-Gaussian characteristics of the vibration signal, and the nonlinear coefficient is constructed using the piecewise linearization method. The kurtosis is used to measure the sharpness of the signal distribution, while the skewness is used to measure the symmetry of the signal. The calculation formulas are: in, is the kurtosis, is the skewness, represents the mathematical expectation operation, is the signal mean, is the signal standard deviation. In order to improve the calculation accuracy, the piecewise linearization method is used to model the nonlinear characteristics, and the best fitting curve is solved by the numerical optimization method to obtain the correction coefficient data. The expression of the nonlinear correction coefficient is expressed as: in, and is the empirical fitting coefficient. The correction coefficient data is used to describe the nonlinear dynamic characteristics of the blade under different modes and to modify the traditional linear frequency response model to improve its applicability. The correction coefficient data is convolved with the frequency response function to calculate the stress spectrum density function under random vibration load. Stress spectrum density function The calculation formula is: Where * represents the convolution operation. Through this calculation process, the stress spectrum density function corresponding to different vibration modes is obtained. The stress spectrum density function quantifies the stress fluctuation characteristics of the blade under random excitation conditions and is used to predict its stress response during long-term service. The stress spectrum density function is characterized by eight parameters to more comprehensively describe the characteristics of the stress spectrum. The eight-parameter characterization method extracts the main statistical characteristics of the stress spectrum, including mean, standard deviation, peak factor, skewness, kurtosis, energy distribution, frequency center and bandwidth factor, which are used to quantify the stress distribution characteristics of the blade under complex load conditions. The rain flow counting method is used to count the amplitude distribution of the stress cycle to identify the cyclic stress characteristics of the blade under different load conditions. The rain flow counting method is a statistical method used for fatigue analysis, which can identify and count the amplitude, period and mean of the stress cycle to form stress cycle data. Based on the stress cycle data, the nonlinear accumulation model is used to calculate the fatigue damage value to evaluate the fatigue life of the blade. Based on the traditional linear damage accumulation model (such as the Miner criterion), the nonlinear accumulation model can more accurately describe the damage accumulation process of the blade under complex stress sequences and consider the mutual influence between different stress levels. The calculation formula of the nonlinear cumulative model is: in, is the fatigue damage value, For the Count of stress cycles, is the fatigue life under this stress amplitude, is the nonlinear cumulative index. In order to improve the accuracy of the calculation, the load sequence effect coefficient Set as the correction parameter so that the final fatigue damage accumulation data is corrected to: in, Obtained by fitting the experimental data. Through the above steps, the fatigue damage accumulation data is obtained.

[0030] In a specific embodiment, the process of executing step S5 may specifically include the following steps: Input the fatigue damage accumulation data into the comparative decoupling module of the double decoupling diagnosis model for feature processing to obtain comparative feature data; Perform data enhancement processing on the contrast feature data, construct enhanced samples using random rotation, translation and scaling operations to obtain enhanced feature data; Based on the contrast feature data and the enhanced feature data, contrast learning is performed, and a working condition perception mask is generated by minimizing the feature distance to obtain the working condition feature data; The working condition feature data is input into the generative adversarial network of the adversarial mask decoupling module in the dual decoupling diagnosis model, and feature extraction is performed through a four-layer convolutional neural network to obtain separation feature data; Perform BatchNormalization and Dropout operations on the separated feature data to obtain fault feature data; The fault feature data is input into the support vector machine classifier, and the radial basis kernel function is used to classify the fault to obtain the target diagnosis result.

[0031] Specifically, the fatigue damage accumulation data is input into the contrast decoupling module of the dual decoupling diagnosis model to perform feature processing and extract key contrast feature data. After feature transformation, the model can capture the pattern differences in the data. Assume that the fatigue damage accumulation data is , whose characteristics are expressed as ,in is the feature extraction function. In the contrast decoupling process, the metric learning method is used to calculate the feature similarity between different samples and define the feature distance is the Euclidean distance: in, and are the feature vectors of different samples, represents the Euclidean norm. By constructing the loss function Optimizing feature representation: in, is the label of the sample pair, is the distance threshold. Through this loss function, the feature distances of similar samples are closer, while the feature distances of different samples are farther, so that comparative feature data can be extracted. Data enhancement is performed on the comparative feature data. The data enhancement method includes random rotation, translation and scaling operations, and the rotation matrix is ​​defined as for: in, is the rotation angle, which is randomly selected to simulate the deformation of the data. and scaling transformation This is achieved through the following formula: in, is the translation amount, is the scaling factor. Through these transformation operations, enhanced samples are generated , and obtain enhanced feature data. Comparative learning is performed based on the contrast feature data and the enhanced feature data to minimize the feature distance and generate a working condition-aware mask. The calculation of the working condition-aware mask is based on the similarity measure in the feature space, and the mask is defined for: in, is a scale parameter that controls the sensitivity of the mask. The mask suppresses the impact of working condition changes, allowing the model to focus on the true characteristics of fatigue damage and obtain working condition feature data. The working condition feature data is input into the adversarial mask decoupling module of the dual decoupling diagnosis model, and the generative adversarial network is used for feature separation. The generative adversarial network consists of a generator and the discriminator The goal of the generator is to generate realistic feature representations, while the goal of the discriminator is to distinguish between real features and generated features. The optimization objective function is as follows: in, is a random noise input. In order to improve the feature extraction capability, a four-layer convolutional neural network is used for feature extraction in the architecture of the generative adversarial network. The convolution operation of the layer is: in, and is the convolution kernel parameter, * indicates the convolution operation, is the activation function. Through four layers of calculation, the separated feature data is obtained. BatchNormalization and Dropout operations are performed on the separated feature data. BatchNormalization stabilizes the data distribution through normalization, and its calculation formula is: in, and are the feature mean and variance respectively, To prevent the denominator from being zero, a tiny constant is used. The Dropout operation randomly discards some neurons to reduce the model's dependence on specific features. The calculation method is: in, is a random mask of Bernoulli distribution, is the retention probability. After these processes, the fault feature data is obtained. The fault feature data is input into the support vector machine classifier for fault identification. Since the characteristics of fatigue damage are highly nonlinear, the radial basis kernel function is used in the support vector machine classifier, and its kernel function is defined as: in, Controls the expansion degree of the kernel function. The goal of the support vector machine classifier is to find the optimal hyperplane to maximize the classification interval, and its optimization objective function is: in, is the hyperplane weight, is the regularization parameter, is a slack variable. The target diagnosis result is obtained through support vector machine calculation.

[0032] The above describes the blade structure performance detection method in the embodiment of the present invention. The following describes the blade structure performance detection device in the embodiment of the present invention. Figure 2 , an embodiment of a blade structure performance detection device in an embodiment of the present invention includes: A scanning module is used to perform laser ultrasonic scanning on titanium alloy engine blades to obtain multi-modal spectrum feature data; An inversion module is used to perform parameter inversion and iterative optimization calculations based on multi-modal spectrum feature data to obtain material parameter distribution data of a local area of ​​the blade; An analysis module is used to perform annular scanning analysis on titanium alloy engine blades based on material parameter distribution data to obtain local structural feature data; Decomposition module, used to perform modal decomposition on local structural characteristic data to obtain fatigue damage accumulation data; The separation module is used to input the fatigue damage accumulation data into the dual decoupling diagnosis model for contrast learning and adversarial feature separation to obtain the target diagnosis result.

[0033] Through the cooperation of the above components, the non-contact laser ultrasonic testing system is adopted, combined with the multi-modal spectrum feature analysis method, to achieve the purpose of rapid in-service detection, and at the same time obtain the complete structural characteristics of the blade. The signal processing method based on wavelet denoising and Fourier transform effectively extracts the resonance mode information of body waves and surface waves, significantly reducing the influence of environmental noise. The improved Levenberg-Marquardt algorithm is used for parameter inversion, and the spatial distribution characteristics of material parameters are accurately identified through multi-starting point iterative optimization and regularization processing. The circular scanning strategy is innovatively introduced to accurately characterize the anisotropic characteristics of the material by analyzing the directional distribution of the surface wave energy flow density, thereby improving the spatial resolution of defect detection. The fatigue damage assessment method based on modal decomposition and nonlinear accumulation model overcomes the limitations of the traditional linear superposition method and accurately predicts the evolution law of fatigue damage under complex loads. A dual decoupling diagnosis model is designed, which effectively solves the problem of working condition interference through contrastive learning and adversarial feature separation, and enhances the robustness and generalization of fault diagnosis.

[0034] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0035] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0036] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting blade structure performance, characterized in that: The method comprises: Laser ultrasonic scanning is performed on titanium alloy engine blades to obtain multi-modal spectrum feature data; Performing parameter inversion and iterative optimization calculations according to the multi-modal spectrum characteristic data to obtain material parameter distribution data of a local area of ​​the blade; Performing annular scanning analysis on the titanium alloy engine blade based on the material parameter distribution data to obtain local structural feature data; Performing modal decomposition on the local structural characteristic data to obtain fatigue damage accumulation data; The fatigue damage accumulation data is input into a dual decoupling diagnosis model for contrast learning and adversarial feature separation to obtain a target diagnosis result.

2. The blade structure performance detection method according to claim 1, characterized in that: The laser ultrasonic scanning of the titanium alloy engine blade is performed to obtain multi-modal spectrum characteristic data, including: The laser parameters of the pulse laser emitter are set to obtain calibrated laser emission parameters, and based on the calibrated laser emission parameters, scanning grid data is constructed on the surface of the titanium alloy engine blade through a two-dimensional scanning galvanometer system; Controlling the Doppler laser detector to receive ultrasonic echo signals according to the scanning grid data to obtain ultrasonic echo data; Acquiring the ultrasonic echo data to obtain time domain waveform data, and performing temperature compensation on the time domain waveform data through a temperature sensor to obtain temperature compensated time domain waveform data; Performing laser focal length correction on the temperature-compensated time-domain waveform data based on a high-precision displacement sensor to obtain an original ultrasonic signal; The original ultrasonic signal is subjected to wavelet denoising and Fourier transform to obtain multimodal spectrum feature data.

3. The blade structure performance detection method according to claim 2, characterized in that: The original ultrasonic signal is subjected to wavelet denoising and Fourier transform to obtain multimodal spectrum feature data, including: Performing wavelet decomposition on the original ultrasonic signal to obtain wavelet coefficients; Performing soft threshold processing on the wavelet coefficients, determining the threshold by calculating the standard deviation of each layer of wavelet coefficients, and obtaining the denoised wavelet coefficients; Performing wavelet reconstruction on the denoised wavelet coefficients to obtain a denoised time domain signal, and inputting the denoised time domain signal into a fast Fourier transform module to obtain spectrum data; Band-pass filtering is performed on the spectrum data to extract the frequency components of the body wave thickness resonance mode, and simultaneously two-dimensionally mapping the spectrum to obtain the surface wave resonance mode to obtain multi-modal resonance data; The fundamental frequency, second harmonic and third harmonic frequencies are extracted based on the multimodal resonance data, and the amplitude ratio and quality factor of each harmonic component are calculated to obtain multimodal spectrum characteristic data.

4. The blade structure performance detection method according to claim 1, characterized in that: The performing of parameter inversion and iterative optimization calculation according to the multi-modal spectrum characteristic data to obtain material parameter distribution data of a local area of ​​the blade includes: Based on the multimodal spectrum characteristic data, a theoretical model considering material anisotropy is constructed, and elastic modulus, density, and Poisson's ratio are set as optimization variables to obtain an initial parameter model; Performing temperature correction on the initial parameter model, calculating a correction coefficient according to temperature data collected by a temperature sensor, and obtaining a temperature-corrected parameter model; Input the temperature-corrected parameter model into the Levenberg-Marquardt optimizer, set the objective function as the weighted mean square error between the theoretical calculated value and the experimental measured value, and obtain the optimized objective function; A regularization term is introduced into the optimization objective function, and a regularization coefficient is determined by an L-curve method to obtain a regularized optimization function; Perform multi-starting point iterative calculation based on the regularized optimization function to obtain optimization parameter data; The optimized parameter data is spatially mapped to generate spatial distribution diagrams of elastic modulus distribution, density distribution and Poisson's ratio distribution, and the material parameter distribution data of the local area of ​​the blade is obtained.

5. The blade structure performance detection method according to claim 1, characterized in that: The method of performing annular scanning analysis on the titanium alloy engine blade based on the material parameter distribution data to obtain local structural feature data includes: Determine scanning parameters based on the material parameter distribution data, and construct annular scanning path data around each scanning point; Performing surface wave measurement on the annular scanning path data, analyzing the attenuation characteristics of the surface wave amplitude in different directions, and obtaining directional attenuation data; Calculating the directional distribution of surface wave energy flow density based on the directional attenuation data, extracting the principal axis direction through principal component analysis, and obtaining material anisotropy data; Performing correlation analysis of adjacent annular regions on the anisotropic data of the material, calculating a correlation coefficient matrix, and obtaining structural uniformity data; Establishing a three-dimensional structural model of a local area according to the structural uniformity data, constructing model parameters including thickness distribution and grain orientation distribution, and obtaining three-dimensional structural model data; The abnormal area detected in the three-dimensional structure model data is scanned in detail to obtain local structure feature data.

6. The blade structure performance detection method according to claim 1, characterized in that: The performing modal decomposition on the local structural characteristic data to obtain fatigue damage accumulation data includes: Performing modal decomposition on the local structural feature data to obtain modal feature data; Performing a generalized Fourier series expansion on the modal characteristic data to obtain a frequency response function; Calculating the kurtosis and skewness of the response signal based on the frequency response function, constructing a nonlinear coefficient by a piecewise linearization method, and obtaining correction coefficient data; Performing a convolution operation on the correction coefficient data and the frequency response function to obtain a stress spectrum density function under a random vibration load; The stress spectrum density function is characterized by eight parameters, and the amplitude distribution of the stress cycle is statistically analyzed by the rain flow counting method to obtain stress cycle data; Based on the stress cycle data, a nonlinear accumulation model is used to calculate fatigue damage values, and the load sequence effect coefficient is set as a correction parameter to obtain fatigue damage accumulation data.

7. The blade structure performance detection method according to claim 1, characterized in that: The fatigue damage accumulation data is input into the dual decoupling diagnosis model for contrast learning and adversarial feature separation to obtain the target diagnosis result, including: Inputting the fatigue damage accumulation data into a comparative decoupling module of a dual decoupling diagnosis model for feature processing to obtain comparative feature data; Performing data enhancement processing on the contrast feature data, constructing enhanced samples by random rotation, translation and scaling operations, and obtaining enhanced feature data; Performing comparative learning based on the comparative feature data and the enhanced feature data, generating a working condition perception mask by minimizing the feature distance, and obtaining working condition feature data; Inputting the operating condition feature data into the generative adversarial network of the adversarial mask decoupling module in the dual decoupling diagnostic model, performing feature extraction through a four-layer convolutional neural network, and obtaining separation feature data; Performing BatchNormalization and Dropout operations on the separated feature data to obtain fault feature data; The fault feature data is input into a support vector machine classifier, and a radial basis kernel function is used to perform fault classification to obtain a target diagnosis result.

8. A blade structure performance testing device, characterized in that: Used to perform the blade structure performance detection method according to any one of claims 1 to 7, the blade structure performance detection equipment comprises: A scanning module is used to perform laser ultrasonic scanning on titanium alloy engine blades to obtain multi-modal spectrum feature data; An inversion module, used to perform parameter inversion and iterative optimization calculation according to the multi-modal spectrum characteristic data to obtain material parameter distribution data of a local area of ​​the blade; An analysis module, used for performing annular scanning analysis on the titanium alloy engine blade based on the material parameter distribution data to obtain local structural feature data; A decomposition module, used for performing modal decomposition on the local structural characteristic data to obtain fatigue damage accumulation data; The separation module is used to input the fatigue damage accumulation data into a dual decoupling diagnosis model for comparative learning and adversarial feature separation to obtain a target diagnosis result.

Images