All-directional operational modal testing method and system based on binocular vision intelligent cooperative tracking for thin-walled workpiece

Abstract

This invention belongs to the field of thin-walled component working modal parameter measurement technology, and discloses a method and system for omnidirectional working modal testing of thin-walled components based on binocular vision intelligent collaborative tracking. The method includes: dimensionality reduction of measurement point data based on principal component analysis; denoising and outlier processing of the data; singular value decomposition based on the power spectral density matrix to generate a singular value spectrum; automatic identification of candidate modal frequencies in the EFDD singular value spectrum using a peak search algorithm, and automatic estimation of their half-power bandwidth to determine the fine analysis frequency band of SSI-DATA; use of an integrated machine learning clustering algorithm for automated interpretation of the stability graph; and addition of a nonlinear diagnosis and identification module for nonlinear detection. This invention constructs a more efficient, intelligent, accurate, and capable modern solution for thin-walled component working modal parameter identification in complex real-world environments, fully leveraging the enormous potential of machine vision non-contact measurement.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of working modal parameter measurement technology for thin-walled components, and particularly relates to a method and system for omnidirectional working modal testing of thin-walled components based on binocular vision intelligent collaborative tracking. Background Technology

[0002] Thin-walled components possess numerous advantages, including light weight, high load-bearing capacity, excellent dynamic performance, and ease of integrated manufacturing. They have been widely applied in high-end, critical fields such as aerospace, automotive, and defense shipbuilding. For instance, thin-walled components have become a primary structural form for aero-engine parts. Efficiently and accurately measuring the operating modal parameters of thin-walled components is the prerequisite and foundation for research on vibration response prediction, vibration characteristic analysis, vibration reduction optimization design, and structural damage identification. Therefore, the need for operating modal testing is extremely urgent.

[0003] Traditional contact modal testing methods can alter the dynamic characteristics of thin-walled components due to the added mass introduced by the attached sensor, especially for lightweight thin-walled structures such as aero-engine components. These methods also suffer from low spatial resolution and difficulties in sensor installation. While some researchers have conducted studies on eliminating the effects of sensor-added mass, the models are complex and lack accuracy. Single-point laser vibrometers, a typical non-contact vibration measurement device, have been widely used in modal testing to overcome the problems of contact modal testing. However, they still suffer from low spatial resolution and time-consuming, labor-intensive point-by-point measurements. Modal testing methods based on continuous single-point laser scanning vibration measurement replace point-by-point single-point laser measurement, improving testing efficiency and spatial resolution. However, these methods require complex optical paths and scanning motion control and execution devices, as well as tedious alignment and debugging in the early stages, and also have high requirements for the surface quality of the object being tested. In addition, there are some full-field optical non-contact vibration measurement methods that can be applied to modal testing of thin-walled parts, such as speckle shear deformation, electronic speckle mode interferometry, and holographic interferometry. Although these methods have good full-field vibration measurement capabilities and high spatial resolution, they generally use lasers as light sources, the optical paths are very complex, and the measurement results are easily affected by external vibrations. They can only be used for scientific research measurements in laboratories equipped with vibration isolation tables, and cannot meet the special needs of working modal testing of thin-walled parts.

[0004] The binocular vision intelligent continuous tracking method for omnidirectional vibration measurement of thin-walled components can meet the requirements of non-contact, omnidirectional, high spatial resolution and high precision measurement of vibration of thin-walled components. In order to obtain the working modal parameters of thin-walled components, it is also necessary to identify the working modal parameters of the omnidirectional vibration measurement data.

[0005] The covariance-driven stochastic subspace method (SSI-COV) and its variants are a classic method for modal parameter identification under environmental excitation. However, when processing vibration measurement data of thin-walled components based on machine vision, its inherent limitations are amplified, mainly in the following four aspects.

[0006] (1) The SSI-COV method is extremely sensitive to noise in vibration measurement data, which is its most fatal weakness when applied to this scenario. The mathematical basis of this method begins with calculating the covariance matrix of the output response data. This calculation process is mathematically equivalent to amplifying the data noise twice. However, the vibration measurement data obtained by machine vision methods inherently has unique noise characteristics. This noise is not traditional Gaussian white noise, but is easily caused by factors such as sudden changes in illumination, missing surface texture, image blurring, or camera shake, resulting in local, sharp data jumps and outliers. When these outliers are fed into the covariance calculation process of SSI-COV, they will seriously pollute the estimation quality of the entire covariance sequence. The direct consequence is that the subsequent system matrix identification based on the polluted covariance matrix will have significant deviations, ultimately leading to serious distortion or even complete errors in the identified modal parameters, especially the damping ratio estimation which is extremely sensitive to noise. At the same time, the purity of the mode shape will also be greatly reduced, and the reliability will be significantly decreased.

[0007] (2) SSI-COV faces a significant computational efficiency bottleneck, making it difficult to handle the massive data challenges brought by machine vision. Full-field vibration measurement of thin-walled components is an advantage of machine vision; to achieve high spatial resolution and high precision, the number of output channels (i.e., the number of measurement points m) can easily reach thousands. The SSI-COV method requires constructing a high-dimensional... (in For user-defined parameters, The Hankel matrix (where the number of time points is 1) is used to calculate a... The covariance matrix of such a huge dimension places almost unbearable demands on computer memory and CPU computing power when storing and performing subsequent singular value decomposition on such a matrix. On ordinary engineering workstations, performing full-channel, full-band SSI-COV analysis often leads to memory exhaustion or unacceptably long computation times, which greatly reduces the feasibility of this method in practical engineering applications.

[0008] (3) The SSI-COV analysis process is relatively rigid and lacks intelligent guidance, belonging to a "blind men and the elephant" style global search. This method requires users to pre-set key parameters and then calculate all data in the same way. It cannot perform targeted analysis based on the characteristics of the data itself. If the researcher is only concerned with a few key modes, or if the signal-to-noise ratio of different frequency bands in the data is significantly different, SSI-COV still needs to consume a lot of resources to process all irrelevant or low-quality data segments. This analysis mode, which lacks focus and applies effort evenly, is undoubtedly a huge waste under the premise of limited computing resources, and it also slows down the entire analysis process.

[0009] (4) The stability graph interpretation process upon which SSI-COV relies becomes more complex and difficult due to its inherent sensitivity to noise. Because data noise is amplified, the stability graph generated by SSI-COV is often filled with numerous spurious stable points caused by noise and outliers. These spurious points are mixed with real physical mode points, requiring analysts to possess extremely rich experience to distinguish valid modes from the chaos. This process is highly subjective, and the interpretation results from different personnel may vary significantly, greatly reducing the objectivity and repeatability of the analysis results.

[0010] Compared to the various shortcomings of SSI-COV, the EFDD+SSI-DATA hybrid strategy, which combines enhanced frequency domain decomposition (EFDD) and data-based random subspace method (SSI-DATA), demonstrates a strong comprehensive advantage. It is not a simple superposition of the two methods, but an organically combined, progressively intelligent analysis process that perfectly matches the characteristics of machine vision vibration measurement data and the needs of engineering applications.

[0011] The core idea of ​​the EFDD+SSI-DATA hybrid strategy is division of labor and cooperation: leveraging the speed of EFDD for global reconnaissance and the precision of SSI-DATA for targeted attacks. This strategy performs exceptionally well in most cases, but its effectiveness is still limited by the inherent characteristics of its components and the way they are combined.

[0012] (1) The computational efficiency bottleneck of this strategy has not been completely overcome, but only shifted and alleviated. As the final accurate identification tool, the computational complexity of the SSI-DATA method is still extremely high in nature. This method requires the construction of a huge data Hankel matrix and a series of QR decomposition and singular value decomposition. When faced with the massive amount of data with tens of thousands of degrees of freedom provided by machine vision, even if only one key frequency band specified by EFDD is analyzed, if there are dense modes in the frequency band or the system itself is complex and requires a high model order for identification, the amount of computation will still increase exponentially, consuming huge amounts of memory and CPU time. This process may take several hours or even days, which is still an unacceptable delay for design verification or online monitoring applications that require rapid iteration. In other words, the hybrid strategy solves the problem of "not wasting computing power on the entire frequency band", but does not solve the fundamental problem of "the high computational cost of SSI-DATA itself".

[0013] (2) The automation and intelligence of this strategy are still insufficient, and it is highly dependent on expert experience. There are several "breakpoints" in the process that require manual intervention. In the EFDD stage, although the singular value spectrum can show the peaks, it is still a challenge for the computer algorithm to automatically and accurately distinguish between the true modal peaks, harmonic peaks, or "humps" and noise peaks caused by nonlinearity. In the SSI-DATA stage, although the stability graph is a powerful tool, its interpretation is highly subjective. Analysts need to set the appropriate number of future / past rows (i value) based on experience and identify the stability bars representing the true modes from the complex "cloud points" in the stability graph. This process is not only time-consuming and laborious, but analysts with different experience may also draw different conclusions, which greatly reduces the repeatability and objectivity of the analysis results. The hybrid strategy optimizes the process, but does not achieve true "one-click" automated output.

[0014] (3) This strategy is based on the assumption of a linear time-invariant system and is helpless against the nonlinear dynamic behavior that thin-walled components may exhibit during operation. Under actual working conditions, thin-walled structures often exhibit strong nonlinearity due to factors such as large deformation, micro-slippage at the connection interface, and the properties of the material itself. For example, the resonant frequency drifts with the excitation amplitude, the mode shape changes with the amplitude, and there are phenomena such as nonlinear damping. EFDD and SSI-DATA are both classic linear system identification methods, which force the nonlinear response to fit to an optimal linear model. As a result, either the identified parameters are some kind of average distortion value that cannot reflect its amplitude dependence characteristics, or a large number of non-convergent and scattered "stable points" are generated in the stability diagram, leading to mode confusion or even failure. The hybrid strategy cannot perceive or quantify these nonlinear phenomena, thus losing important information that reveals the true dynamic behavior of the structure.

[0015] (4) This strategy is highly dependent on the quality of the front-end data and lacks a dedicated processing module for the noise specific to machine vision vibration measurement data. The noise in machine vision vibration measurement data is not traditional Gaussian white noise, but includes spatially correlated noise, local outliers, etc. Although SSI-DATA is slightly more robust to noise than the covariance-based method (SSI-COV), the existing general workflow lacks an intelligent spatial-temporal joint filtering preprocessing step tailored to the characteristics of visual data. These anomalous noise points will directly contaminate the subsequent global correlation function estimation and stability graph generation, leading to a decrease in analysis performance.

[0016] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are:

[0017] Existing methods suffer from problems such as computational efficiency bottlenecks, noise sensitivity, insufficient automation and intelligence, and strong reliance on expert experience. Summary of the Invention

[0018] To address the problems existing in the prior art, this invention provides a method and system for omnidirectional working modal testing of thin-walled components based on binocular vision intelligent collaborative tracking.

[0019] This invention is implemented as follows: a method for omnidirectional working modal testing of thin-walled components based on binocular vision intelligent collaborative tracking, comprising:

[0020] Step 1: Perform principal component analysis on the multi-point displacement data acquired by binocular vision to achieve dimensionality reduction;

[0021] Step 2: Perform temporal and spatial denoising on the dimensionality-reduced data and complete outlier identification.

[0022] Step 3: In the empirical frequency domain decomposition step, singular value decomposition is performed on the power spectral density matrix to generate a singular value spectrum.

[0023] Step 4: Use the peak search algorithm to identify the modal peak frequencies in the singular value spectrum and determine the analysis frequency band range;

[0024] Step 5: Complete the modal identification process by analyzing the obtained candidate frequencies and frequency band input subspace identification data.

[0025] Step 6: Based on the clustering algorithm, automatically distinguish the identified stable points and extract the true modal features;

[0026] Step 7: Use time-frequency analysis to determine the evolution of frequency and damping over time and confirm whether nonlinearity exists;

[0027] Step 8: After confirming the nonlinearity, use a time-varying autoregressive model combined with Hilbert vibration decomposition to extract amplitude-related parameters.

[0028] Furthermore, the data denoising method combines two-dimensional spatial filtering with temporal continuity residual determination, and outlier repair is completed by using moving window kurtosis and skewness statistics.

[0029] Furthermore, the input to the singular value decomposition is the result of the superposition of multiple power spectral density matrices, and a modal candidate set is constructed by eigenvectors with an energy percentage greater than 90%.

[0030] Furthermore, peak search determines the mode boundary by the change in the sign of the first derivative and the half-power energy attenuation ratio, and the resulting frequency band range is used as the subspace identification input parameter.

[0031] Furthermore, the clustering process encodes each stable point into a three-dimensional vector containing frequency, damping, and mode shape similarity, and identifies the clustered regions as effective modal point clusters through distance metrics.

[0032] Furthermore, false modes in the stable graph are eliminated by making the cluster density lower than a threshold, and the true modes are taken as the output mode parameters by taking the center value of each cluster.

[0033] Furthermore, the time-frequency analysis uses continuous wavelet transform or Hilbert time-frequency processing to estimate the instantaneous frequency, and determines whether the system has entered the nonlinear region by the time-segmented damping attenuation slope.

[0034] Furthermore, the parameters of the time-varying autoregressive model in nonlinear identification are obtained by least squares iteration, and the Hilbert envelope is used to obtain the frequency modulation amount caused by amplitude changes.

[0035] Furthermore, the binocular vision data acquisition adopts a synchronous exposure and fixed frame rate method to ensure that the pixel trajectories of the two viewpoints are consistent in the time dimension.

[0036] Another objective of this invention is to provide an omnidirectional modal testing system for thin-walled components, comprising:

[0037] A data processing module used to perform dimensionality reduction in principal component analysis.

[0038] Intelligent filtering module for spatial and temporal noise suppression;

[0039] A linear identification module used to perform singular value spectrum peak frequency extraction and output modal frequency bands;

[0040] An automatic modal parameter extraction module for interpreting stable graph clustering;

[0041] A time-varying model analysis module for identifying nonlinear dynamic characteristics.

[0042] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0043] This invention systematically enhances the EFDD+SSI-DATA hybrid strategy by performing pre-processing intelligent data processing, deep integration of process automation, and expansion of nonlinear capabilities. It constructs a more efficient, intelligent, accurate, and capable modern solution for identifying working modal parameters of thin-walled components in complex real-world environments, thereby fully leveraging the enormous potential of machine vision non-contact measurement.

[0044] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0045] The technical solution of this invention, after transformation, is mainly applied to typical thin-walled components such as aero-engine blades, automotive body-in-white hoods, and satellite battery panels. Efficiently and accurately measuring the working modal parameters of thin-walled components is the prerequisite and foundation for research on vibration response prediction, vibration characteristic analysis, vibration reduction optimization design, and structural damage identification. The demand for working modal testing is extremely urgent. Traditional contact modal testing methods, single-point laser single-point vibration measurement methods, and single-point laser continuous scanning vibration measurement methods are all insufficient to meet the special requirements of working modal testing for thin-walled components. Binocular vision-based omnidirectional vibration measurement of thin-walled components can meet the requirements of non-contact, omnidirectional, high spatial resolution, and high-precision measurement of vibration in thin-walled components. Its key is the identification of working modal parameters based on omnidirectional vibration measurement data. This invention is highly attractive to its target audience, with urgent market demand and a promising future. The specific benefits of this invention rely on product sales, technical support, and subsequent product and technology updates and after-sales service.

[0046] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0047] Covariance-driven stochastic subspace method (SSI-COV) and its variants are a classic method for modal parameter identification under environmental excitation, but their inherent limitations are amplified when processing vibration measurement data of thin-walled components based on machine vision. Compared to the shortcomings of SSI-COV, the EFDD+SSI-DATA hybrid strategy, which combines Enhanced Frequency Domain Decomposition (EFDD) and the data-based random subspace method (SSI-DATA), demonstrates strong comprehensive advantages. It is not a simple superposition of the two methods, but an organically combined, progressively intelligent analysis process that perfectly matches the characteristics of machine vision vibration measurement data and the needs of engineering applications. However, it still has shortcomings such as low computational efficiency, strong dependence on expert experience, insufficient automation and intelligence, inability to handle the nonlinear dynamic behavior of thin-walled parts, and strong dependence on the quality of front-end data. To address these issues, this invention has made systematic enhancements in four dimensions: front-end intelligent data processing, deep integration of process automation, expansion of nonlinear capabilities, and exploration of deep learning paradigms. This has built a more efficient, intelligent, accurate, and modern solution for identifying the working modal parameters of thin-walled parts that can better cope with complex real-world environments, thereby fully leveraging the enormous potential of machine vision non-contact measurement.

[0048] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0049] The demand for modal analysis of typical thin-walled components such as aero-engine blades, automotive body-in-white hoods, and satellite solar panels is urgent. Binocular vision-based omnidirectional vibration measurement of thin-walled components can meet the requirements of non-contact, omnidirectional, high spatial resolution, and high precision measurement. Its key lies in the identification of modal parameters based on omnidirectional vibration measurement data. The EFDD+SSI-DATA hybrid strategy, combining Enhanced Frequency Domain Decomposition (EFDD) and the data-based random subspace method (SSI-DATA), has demonstrated strong comprehensive advantages. However, it still suffers from core shortcomings such as low computational efficiency, strong reliance on expert experience, insufficient automation and intelligence, inability to handle the nonlinear dynamics of thin-walled components, and extreme dependence on the quality of front-end data. These are technical challenges that the industry has long sought to solve but has yet to achieve success in. This invention addresses these core shortcomings by systematically enhancing the system in four dimensions: pre-processing intelligent data, deep integration of automated processes, expansion of nonlinear capabilities, and exploration of deep learning paradigms. This provides a more efficient, intelligent, accurate, and modern solution for identifying modal parameters of thin-walled components in complex real-world environments.

[0050] (4) The technical solution of the present invention overcomes technical bias:

[0051] The demand for modal testing of typical thin-walled components such as aero-engine blades, automotive body-in-white hoods, and satellite solar panels is urgent. Binocular vision-based omnidirectional vibration measurement of thin-walled components can meet the requirements of non-contact, omnidirectional, high spatial resolution, and high-precision measurement of thin-walled component vibration. Its key lies in the identification of modal parameters based on omnidirectional vibration measurement data. However, current modal parameter identification based on omnidirectional vibration measurement data suffers from several core shortcomings: low computational efficiency, strong reliance on expert experience, insufficient automation and intelligence, inability to handle the nonlinear dynamic behavior of thin-walled components, and extreme dependence on the quality of front-end data. Therefore, the industry generally has significant doubts about the application capability and level of binocular vision-based omnidirectional vibration measurement modal testing methods for thin-walled components. This invention addresses these core shortcomings by systematically enhancing the technology in four dimensions. Its application capability and level have been fully verified, completely overcoming industry prejudices against this technology and eliminating doubts about its application. Attached Figure Description

[0052] Figure 1 This is a flowchart of the omnidirectional working modal testing method for thin-walled components based on binocular vision intelligent collaborative tracking provided in an embodiment of the present invention;

[0053] Figure 2 This is a block diagram of a thin-walled component omnidirectional working modal testing system based on binocular vision intelligent collaborative tracking provided in an embodiment of the present invention;

[0054] Figure 3 This is the omnidirectional vibration modal testing platform for thin-walled components provided in the embodiments of the present invention;

[0055] Figure 4 This is a stable graph of EFDD+SSI-DATA provided in the embodiments of the present invention;

[0056] Figure 5 This is a stability graph of the method of the present invention provided in the embodiments of the present invention;

[0057] Figure 6 This is a comparison diagram of vibration modes provided in the embodiments of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0059] The challenges of current thin-walled structures in modal identification stem primarily from the insufficient ability of testing methods to capture the full-field response. Traditional accelerometer placement suffers from mass loading effects, and point measurement methods struggle to fully represent the spatial distribution of higher-order modes. This often leads to issues such as mode omission, parameter drift, and low-frequency distortion in industrial applications. Furthermore, sensor arrays commonly used in engineering projects struggle to maintain sufficient spatial resolution under complex boundary conditions and on highly flexible thin-walled components. This results in modal identification relying on human experience and feature extraction lacking objectivity, further impacting subsequent dynamic parameter inversion and vibration reliability analysis.

[0060] Binocular vision collaborative tracking constructs a non-contact, high-density sampling observation system for the global response of thin-walled components. Its core lies in reconstructing the surface displacement field and synchronously responding to multiple measurement points, enabling modal recognition to transition from low-dimensional discrete signals to high-dimensional continuous dynamic fields. However, such data typically involves millions of pixels, making noise and optical field disturbances the primary sources of interference affecting modal reliability. Therefore, this method uses dimensionality reduction as the front end of the chain, compressing the data dimension through principal component analysis to clearly reveal high-energy response characteristics in the mathematical space, laying a stable input foundation for subsequent frequency and time domain analyses.

[0061] The linear modal identification section employs a hybrid strategy of EFDD and SSI-DATA to construct a two-stage solution path. The power spectral density matrix is ​​decomposed into a singular value spectrum, with energy peaks corresponding to intrinsic modes. A peak search script mines potential frequency bands and automatically feeds them back to SSI-DATA, eliminating the need for manual parameter constraints. This coupling approach improves the sensitivity of frequency extraction while avoiding modal leakage caused by excessively wide bandwidths, demonstrating high adaptability from industrial testing to flexible manufacturing line debugging. The stability graph is then analyzed using a machine learning clustering algorithm. Each estimated point is embedded as an independent vector in a multi-dimensional space, and the cluster centers naturally converge to the true modal parameters, automatically removing stray and spurious modes. This automates the process that has long relied on manual interpretation of stability graphs in engineering.

[0062] When thin-walled components enter nonlinear regions such as large amplitude or boundary loosening, single linear modal frames often fail, with frequencies and damping slowly drifting over time. To achieve real-time insight into this phenomenon, this method introduces time-frequency analysis to track key point responses, directly presenting the frequency evolution trajectory and damping changes, providing observable evidence for determining whether nonlinearity has occurred. Once nonlinearity is triggered, the nonlinearity identification stage begins. Energy transfer and frequency modulation are decomposed using a combination of a time-varying autoregressive model and Hilbert vibration decomposition, extracting amplitude-dependent features and characterizing dynamic nonstationary features that are difficult to obtain in the traditional frequency domain.

[0063] In industrial equipment modal testing platforms, the system manifests as a modular pipeline: a dimensionality reduction module handles data compression, a filtering module removes spatiotemporal disturbances, a hybrid recognition module assigns linear modal solving tasks, and a nonlinear module takes over and outputs evolution curves when needed. The entire process avoids repeated manual parameter iterations, allowing testers to complete only visual calibration and load settings, while the system autonomously completes the entire process from modal extraction to nonlinear recognition. This capability directly supports industrial scenarios such as aerospace thin-walled skins, tool-workpiece elastic coupling, and rail transit aluminum alloy body panels. In the current context of increasingly lightweight structures and shrinking safety margins, it offers significant engineering improvements and enhanced testing reliability.

[0064] like Figure 1 As shown, this embodiment of the invention provides a method for omnidirectional working modal testing of thin-walled components based on binocular vision intelligent collaborative tracking, including:

[0065] S1, Principal component analysis is performed on the multi-point displacement data acquired by binocular vision to achieve dimensionality reduction;

[0066] S2 performs temporal and spatial denoising on the dimensionality-reduced data and completes outlier identification;

[0067] S3, In the empirical frequency domain decomposition step, singular value decomposition is performed on the power spectral density matrix to generate a singular value spectrum.

[0068] S4. A peak search algorithm is used to identify the modal peak frequencies in the singular value spectrum and determine the analysis frequency band range.

[0069] S5, complete the modality recognition process by analyzing the obtained candidate frequencies and frequency band input subspace identification data;

[0070] S6, based on clustering algorithm, automatically distinguishes the identified stable points and extracts the real modal features;

[0071] S7 uses time-frequency analysis to determine the evolution of frequency and damping over time and to confirm the presence of nonlinearity.

[0072] S8, after confirming the nonlinearity, uses a time-varying autoregressive model combined with Hilbert vibration decomposition to extract amplitude-related parameters.

[0073] The omnidirectional working modal testing method for thin-walled components based on binocular vision intelligent collaborative tracking described in this invention takes "non-contact full-field measurement - intelligent dimensionality reduction and noise reduction - frequency domain modal extraction - nonlinear characteristic discrimination" as its core technical route, and realizes high-precision and automated identification of modal parameters of thin-walled components under real working conditions.

[0074] During the testing process, a binocular vision system is first used to synchronously track multiple natural or artificial feature points on the surface of the thin-walled component, acquiring displacement response data of each measurement point in three-dimensional space over time. Since binocular vision acquires high-dimensional data and a large number of measurement points, direct analysis would lead to high computational complexity and easily introduce redundant information. Therefore, in step S1, principal component analysis is introduced to statistically reduce the dimensionality of the multi-measurement point displacement data, effectively compressing the data size while preserving the main structural dynamics information.

[0075] Subsequently, in step S2, in response to the problem that visual measurements are easily affected by changes in illumination, noise interference and local occlusion under actual working conditions, the time domain and spatial data after dimensionality reduction are jointly denoised, and non-structural response or measurement error points are eliminated through an anomaly identification mechanism, thereby ensuring the stability and reliability of subsequent frequency domain analysis data.

[0076] In step S3, the power spectral density matrix of the system is constructed based on the empirical frequency domain decomposition method, and singular value decomposition is performed on it to obtain the singular value spectrum reflecting the structural dynamic characteristics. This singular value spectrum can amplify the true modal energy information in the frequency domain, providing a basis for the automatic identification of modal frequencies.

[0077] Next, in step S4, a peak search algorithm is used to scan the singular value spectrum, automatically identify the peak frequency of each mode, and determine a reasonable analysis frequency band range accordingly, effectively avoiding the uncertainty caused by human experience in selecting the frequency band.

[0078] In step S5, the selected candidate frequencies and corresponding frequency bands are input into the subspace for data analysis. Preliminary identification of modal parameters is achieved through state-space modeling. Subsequently, in step S6, a clustering algorithm is introduced to automatically distinguish the numerous stable points obtained during the identification process, eliminating false modes and extracting the frequency, damping, and mode shape characteristics of the true modes.

[0079] Furthermore, in step S7, the identification results are analyzed using time-frequency analysis to determine whether the modal frequencies and damping parameters change over time, thus confirming whether the structure exhibits nonlinear dynamic behavior. If nonlinear characteristics are confirmed, a time-varying autoregressive model is introduced in step S8, combined with Hilbert vibration decomposition, to refine the vibration signal and extract time-varying parameters related to amplitude, thereby achieving accurate characterization of the nonlinear working modal characteristics of the thin-walled component.

[0080] Through the synergistic effect of the above steps, this invention realizes omnidirectional, non-contact, intelligent modal testing and nonlinear characteristic identification of thin-walled components under complex working conditions.

[0081] The present invention provides a data denoising method that combines two-dimensional spatial filtering with temporal continuity residual determination, and completes outlier repair by using moving window kurtosis and skewness statistics.

[0082] The embodiments of the present invention provide a multi-order superposition result of the power spectral density matrix as input to singular value decomposition, and construct a modal candidate set through eigenvectors with an energy ratio greater than 90%.

[0083] The embodiments of the present invention provide a peak search method that determines the mode boundary by the change of the sign of the first derivative and the half-power energy attenuation ratio, and the resulting frequency band range is used as the subspace identification input parameter.

[0084] The present invention provides a clustering process that encodes each stable point into a three-dimensional vector containing frequency, damping, and mode shape similarity, and identifies clustered regions as effective modal point clusters through distance metrics.

[0085] In this embodiment of the invention, false modes in the stable graph are removed by making the cluster density lower than a threshold, and the true modes are taken as the output mode parameters by taking the center value of each cluster.

[0086] The embodiments of the present invention provide time-frequency analysis that uses continuous wavelet transform or Hilbert time-frequency processing to achieve instantaneous frequency estimation, and determines whether the system has entered the nonlinear region by the time-segmented damping attenuation slope.

[0087] The present invention provides that the parameters of the time-varying autoregressive model in nonlinear identification are obtained by least squares iteration, and the Hilbert envelope is used to obtain the frequency modulation amount caused by the amplitude change.

[0088] The embodiments of the present invention provide binocular vision data acquisition using synchronous exposure and fixed frame rate to ensure that the pixel trajectories of the two viewpoints are consistent in the time dimension.

[0089] The clustering algorithm provided in this embodiment of the invention is as follows: deep embedding clustering learns a low-dimensional representation of the data through an autoencoder, and optimizes the clustering results in the representation space using KL divergence.

[0090] formula

[0091] Soft assignment probability (Student's t-distribution)

[0092]

[0093] Parameter meaning

[0094] : The low-dimensional representation of the i-th data point output by the encoder.

[0095] : The vector representation of the j-th cluster center.

[0096] : The probability that sample i belongs to cluster j.

[0097] Target distribution (auxiliary distribution)

[0098]

[0099] Parameter meaning

[0100] : The probability that sample i belongs to cluster j in the auxiliary target distribution, used to enhance the weight of high-confidence assignments.

[0101] KL divergence loss

[0102]

[0103] Parameter meaning

[0104] $L$: KL divergence loss, used to measure the difference between the target distribution P and the soft-assigned distribution Q.

[0105] The core advantage of the EFDD+SSI-DATA hybrid strategy used in this invention lies in achieving a perfect balance and unity between efficiency and accuracy. The hybrid strategy cleverly divides tasks, with the EFDD method acting as the "scout." EFDD, based on the singular value decomposition of the power spectral density matrix, boasts extremely high computational efficiency, enabling rapid scanning and analysis of full-field visual data from tens of thousands of measurement points, generating a clear singular value spectrum in a very short time. This spectrum acts like a "modal map," intuitively displaying the frequency positions and energy levels of all possible modes, answering the global question of "where are these modes?" The SSI-DATA method then functions as a "precision-guided weapon." It no longer needs to process all the massive amounts of data, but instead, based on the intelligence provided by EFDD, it delves deeply into the narrow frequency bands containing the key modes of interest. SSI-DATA operates directly on the original data matrix, avoiding the noise amplification effect of SSI-COV covariance calculation, resulting in stronger noise resistance and widely recognized as having one of the highest damping ratio identification accuracy among all time-domain methods. This collaborative model ensures that valuable computational resources are precisely allocated to the most critical areas, achieving both rapid and accurate analysis. This strategy significantly enhances the reliability and confidence of the analysis results. EFDD (frequency domain method) and SSI-DATA (time domain method) analyze the same set of data based on fundamentally different mathematical principles. Therefore, the modal frequencies initially identified by EFDD and the final results precisely identified by SSI-DATA can constitute a powerful cross-validation. If the main modal parameters obtained by the two methods are highly consistent, analysts can have great confidence in the final results. This mutual verification based on different principles is something that no single method can provide. Furthermore, the prior knowledge provided by EFDD offers crucial guidance for interpreting SSI-DATA stability plots. Analysts no longer need to interpret the entire stability plot aimlessly; instead, they can focus directly on the vicinity of the frequencies indicated by EFDD, looking for clear stability bars. This greatly reduces over-reliance on human experience, minimizes the possibility of misjudgments and omissions, and makes the results more objective and reliable.

[0106] The hybrid strategy demonstrates high flexibility and strong engineering applicability. It breaks down a complex modality recognition problem into two logically clear stages: "macroscopic understanding" and "microscopic focus," providing engineers with a clear and actionable workflow. Researchers can flexibly adjust the strategy according to the actual needs of the project: if only a rapid baseline assessment is required, the EFDD stage can be completed; if precise quantification of specific modalities is needed, SSI-DATA can be initiated for detailed analysis. This flexibility makes resource allocation more rational than ever before. More importantly, the SSI-DATA stage can directly analyze the principal components after PCA dimensionality reduction, thereby compressing tens of thousands of channels of data into dozens of dimensions, achieving an order-of-magnitude improvement in computational efficiency. This makes processing ultra-large visual datasets possible, transforming the "impossible" into the "possible."

[0107] This hybrid strategy lays a solid foundation for the future introduction of more advanced data processing and automation algorithms. The peak frequency and bandwidth automatically identified in the EFDD stage are automatically passed to SSI-DATA as input parameters via scripts, achieving automated process integration. Intelligent tools based on clustering algorithms are developed to automatically interpret the stability graph generated by SSI-DATA, automatically filtering out the true physical modes. Ultimately, this achieves intelligent and automated operation of the entire process from data input to parameter output, minimizing human intervention—something that is difficult to achieve with a single SSI-COV process.

[0108] Instead of directly inputting thousands of measurement point data into SSI-DATA, it is better to first extract features from the spatial domain in S1. Using Principal Component Analysis (PCA), an optimal set of orthogonal bases can be found, projecting the original massive displacement field data onto the coordinates of fewer than 100 principal components. Subsequent SSI-DATA analysis is only performed on the time series of these principal components, which can reduce the computational load by several orders of magnitude while preserving the dominant modal information to the maximum extent. This is the preferred "slimming" solution for processing full-field visual data.

[0109] S2, spatially, utilizes the spatial continuity of the thin-walled component's mode shapes, employing a two-dimensional adaptive filter or wavelet threshold denoising technique to smooth out abrupt changes that violate physical laws. Temporally, it employs a statistical outlier detection algorithm to automatically identify and repair instantaneous jumps caused by optical flow tracking failures. This will provide a "clean," high-quality dataset for subsequent modal analysis, fundamentally improving input quality.

[0110] S6 integrates a machine learning clustering algorithm for automated interpretation of the stability graph. Each point in the stability graph is treated as a multi-dimensional data point containing frequency, damping, and mode shape correlation coefficients. Unsupervised clustering automatically separates clustered, true stable points from scattered spurious points and automatically outputs the modal parameters of the cluster centers. This significantly reduces reliance on expert experience, improving analytical efficiency and the objectivity of the results.

[0111] like Figure 2 As shown in the figure, an embodiment of the present invention provides a thin-walled component omnidirectional working modal testing system based on binocular vision intelligent collaborative tracking, specifically including:

[0112] A data processing module used to perform dimensionality reduction in principal component analysis.

[0113] Intelligent filtering module for spatial and temporal noise suppression;

[0114] A linear identification module used to perform singular value spectrum peak frequency extraction and output modal frequency bands;

[0115] An automatic modal parameter extraction module for interpreting stable graph clustering;

[0116] A time-varying model analysis module for identifying nonlinear dynamic characteristics.

[0117] The thin-walled component omnidirectional working modal testing system based on binocular vision intelligent collaborative tracking provided by this invention is not a simple parallel combination of existing single modal recognition technologies or signal processing methods. Instead, it is an overall technical solution that is multi-mechanism collaborative, progressive, and mutually feedback-based, based on the weak stiffness, high coupling, and strong nonlinear vibration characteristics of thin-walled components under actual working conditions. Its functional modules form an organic coupling relationship at the data flow, parameter flow, and criterion levels to jointly complete the stable identification of working modes.

[0118] During system operation, high-dimensional spatiotemporal response data of the thin-walled component in omnidirectional working modes is first acquired through binocular vision intelligent collaborative tracking. This data simultaneously contains the true dynamic information of the structure and redundant components introduced by illumination changes, parallax errors, and background disturbances. To avoid the high-dimensional redundant data from compromising the stability of subsequent modal recognition, the system does not directly enter frequency domain analysis. Instead, it first performs principal component analysis for dimensionality reduction through the data processing module. While maintaining the main energy characteristics and modal correlations of the structure, the original observation space is orthogonally reconstructed, establishing a low-redundancy, highly consistent feature subspace foundation for subsequent filtering and recognition.

[0119] Based on this, the intelligent filtering module does not suppress noise in isolation, but rather works in concert to suppress spatial and temporal noise in combination, taking into account the statistical characteristics and temporal correlation of the dimensionality-reduced data. This allows the micro-amplitude modal response of the thin-walled component under weak excitation conditions to be highlighted, thus avoiding the excessive smoothing or weakening of the real modal information by traditional filtering.

[0120] Subsequently, the linear identification module performs singular value spectrum peak frequency extraction in the noise-controlled data subspace. This process is not an independent frequency search, but rather forms a consistent constraint with the results of the preceding dimensionality reduction and filtering. It outputs candidate mode frequency bands only under the conditions of satisfying structural energy accumulation and temporal stability, thereby effectively suppressing the generation of false spectral peaks.

[0121] Furthermore, the automatic modal parameter extraction module, based on the stability graph clustering mechanism, interprets the identification results of multiple working conditions and multiple time periods in a consistent manner, and automatically completes the extraction of modal frequencies, damping ratios and mode shape parameters. This process relies on the stability support of the aforementioned frequency band identification results, avoiding the subjectivity and uncertainty brought about by manual interpretation.

[0122] Finally, based on the acquired reliable modal parameters, the time-varying model analysis module performs time-varying identification of the nonlinear dynamic characteristics of thin-walled components, thereby characterizing the evolution of working conditions, structural coupling effects, and nonlinear response, and forming a complete closed loop description of the working modes.

[0123] In summary, this invention forms an inseparable overall working mechanism through the synergistic cooperation of the above modules in terms of data structure, processing order, and criteria. Its technical effect does not come from any single module, but from the synergistic effect of multiple technical features under a unified goal. It significantly improves the stability, accuracy, and engineering applicability of omnidirectional working mode testing of thin-walled parts, and has made significant progress that cannot be expected by existing technology combinations.

[0124] Evidence related to the technical effects obtained by the embodiments of the present invention.

[0125] Using a self-made T-shaped thin-walled component as the experimental object, an experimental platform for omnidirectional vibration modal testing of the thin-walled component was built, such as... Figure 3 As shown, an omnidirectional vibration measurement method for thin-walled components based on binocular vision intelligent continuous tracking is used to obtain omnidirectional vibration data of the thin-walled components. Operating modal parameters are identified using a hybrid strategy of SSI-COV, EFDD+SSI-DATA, and the improved EFDD+SSI-DATA hybrid strategy of this invention, respectively. Figure 4 , Figure 5 , Figure 6As shown. Experimental results show that the modal parameter recognition accuracy rates of the SSI-COV, EFDD+SSI-DATA hybrid strategy, and the improved EFDD+SSI-DATA hybrid strategy of this invention are 68.6%, 82.5%, and 99.2%, respectively.

[0126]

[0127] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0128] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for omnidirectional working modal testing of thin-walled components based on binocular vision intelligent collaborative tracking, characterized in that, include: Step 1: Perform principal component analysis on the multi-point displacement data acquired by binocular vision to achieve dimensionality reduction; Step 2: Perform temporal and spatial denoising on the dimensionality-reduced data and complete outlier identification. Step 3: In the empirical frequency domain decomposition step, singular value decomposition is performed on the power spectral density matrix to generate a singular value spectrum. Step 4: Use the peak search algorithm to identify the modal peak frequencies in the singular value spectrum and determine the analysis frequency band range; Step 5: Complete the modal identification process by analyzing the obtained candidate frequencies and frequency band input subspace identification data. Step 6: Based on the clustering algorithm, automatically distinguish the identified stable points and extract the true modal features; Step 7: Use time-frequency analysis to determine the evolution of frequency and damping over time and confirm whether nonlinearity exists; Step 8: After confirming the nonlinearity, use a time-varying autoregressive model combined with Hilbert vibration decomposition to extract amplitude-related parameters.

2. The method according to claim 1, characterized in that, The data denoising method combines two-dimensional spatial filtering with temporal continuity residual determination, and outlier repair is completed by using moving window kurtosis and skewness statistics.

3. The method according to claim 1, characterized in that, The input to the singular value decomposition is the result of the superposition of multiple power spectral density matrices, and a modal candidate set is constructed by eigenvectors with an energy percentage greater than 90%.

4. The method according to claim 1, characterized in that, Peak search determines the mode boundary by the change in the sign of the first derivative and the half-power energy attenuation ratio, and the resulting frequency band range is used as the subspace identification input parameter.

5. The method according to claim 1, characterized in that, Clustering encodes each stable point into a three-dimensional vector containing frequency, damping, and mode shape similarity, and identifies clustered regions as effective modal point clusters through distance metrics.

6. The method according to claim 1, characterized in that, In the stable graph, false modes are removed by making the cluster density lower than a threshold, and the true modes are taken as the output mode parameters by taking the center value of each cluster.

7. The method according to claim 1, characterized in that, Time-frequency analysis uses continuous wavelet transform or Hilbert time-frequency processing to estimate instantaneous frequency, and determines whether the system has entered the nonlinear region by the time-segmented damping attenuation slope.

8. The method according to claim 1, characterized in that, In nonlinear identification, the parameters of the time-varying autoregressive model are obtained by least squares iteration, and the Hilbert envelope is used to obtain the frequency modulation amount caused by amplitude changes.

9. The method according to claim 1, characterized in that, The binocular vision data acquisition uses synchronous exposure and a fixed frame rate to ensure that the pixel trajectories of the two viewpoints are consistent in the time dimension.

10. A binocular vision intelligent collaborative tracking-based omnidirectional working modal testing system for thin-walled components, based on the method of any one of claims 1 to 9, characterized in that, include: A data processing module used to perform dimensionality reduction in principal component analysis. Intelligent filtering module for spatial and temporal noise suppression; A linear identification module used to perform singular value spectrum peak frequency extraction and output modal frequency bands; An automatic modal parameter extraction module for interpreting stable graph clustering; A time-varying model analysis module for identifying nonlinear dynamic characteristics.

Images