A method and measurement system for reconstructing blade tip clearance measurement data
By combining mean singular value decomposition dictionary learning and minimum angle regression algorithm, an adaptive dictionary is trained and sparse representation is performed, which solves the data reconstruction problem under undersampling conditions in the measurement of blade tip clearance of aero-engines and achieves high-precision and robust signal recovery.
Patent Information
- Application Number
- CN202611229057.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-13
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for measuring blade tip clearance in aero-engines suffer from incomplete and distorted measurement data under under-sampling conditions, as well as insufficient anti-interference capabilities and adaptability to complex operating conditions, making it difficult to achieve high-precision blade tip clearance data reconstruction.
By employing mean singular value decomposition dictionary learning and minimum angle regression algorithm, an adaptive dictionary is trained by alternately executing sparse coding and dictionary updates. The minimum angle regression algorithm is then used for sparse representation, and the dictionary atoms most relevant to the signal are selected step by step and the sparse coefficients are updated to achieve high-precision reconstruction of the leaf tip gap signal.
Under under-sampling conditions, it can accurately recover the blade tip gap data, improve anti-interference ability and adaptability to complex working conditions, reduce the computing resource requirements, and achieve high-precision signal reconstruction.
Smart Images

Figure CN123116206A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aero-engine blade tip clearance measurement technology, and in particular to a method and system for reconstructing blade tip clearance measurement data. Background Technology
[0002] Aero-engine tip clearance refers to the radial distance between the tip of the engine rotor blade and the inner wall of the casing, and is one of the key parameters for monitoring engine operating status. The size of the tip clearance directly affects the engine's efficiency, safety, and service life—excessive clearance leads to increased airflow leakage and decreased efficiency, while insufficient clearance may cause blade-casing contact failures, and in severe cases, even catastrophic accidents. Therefore, achieving accurate online measurement of tip clearance is of great significance for aero-engine performance evaluation, fault diagnosis, and safe operation.
[0003] In recent years, capacitive sensors, as a non-contact measurement method, have demonstrated significant advantages in real-time online monitoring of blade tip clearance. Capacitive sensors accurately reflect dynamic changes in clearance by capturing minute capacitance variations between the blade tip and the casing. Current measurement methods based on the root mean square (RMS) value of the capacitance signal, by constructing a signal model and deriving the relationship between the RMS value and the blade tip clearance, have achieved clearance measurement under certain conditions. However, under the high-speed rotation conditions of aero-engines, the blade tip sweeps past the sensor at extremely high linear velocities, resulting in extremely rapid changes in the clearance signal. According to the Nyquist sampling theorem, to accurately capture complete signal information, the sampling frequency must be at least twice the highest frequency of the signal. For high-speed rotating bladed disks, meeting this requirement necessitates a very high sensor sampling rate, which is often difficult to achieve in practical engineering.
[0004] When the sampling frequency is insufficient to meet the requirements of the Nyquist sampling theorem, undersampling occurs, leading to incomplete measurement data and signal distortion, thus making it impossible to effectively extract accurate tip clearance values. Some capacitive sensor solutions attempt to ensure data accuracy by increasing the sampling frequency, but this not only results in enormous computational load and severe data redundancy but also significantly increases the hardware burden and data processing pressure on the system. For equipment like aero engines, which have strict requirements on weight, power consumption, and reliability, simply relying on increasing the sampling frequency is not practical.
[0005] Existing technologies attempt to reduce the sampling rate through compressed sensing (CS) and recover the signal using sparse reconstruction. However, existing methods have significant shortcomings: (1) It is sensitive to noise and has poor reconstruction accuracy under strong engine interference. (2) Large reconstruction error in high dynamic range measurements; (3) Sparse representation is inefficient, relies on a fixed sparse base or a simple dictionary, and has poor adaptability; (4) Lacks adaptive capability when signal characteristics change under complex operating conditions.
[0006] Therefore, there is an urgent need for a method that can reconstruct high-precision blade tip clearance data under conditions of undersampling, strong noise, and variable operating conditions. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for reconstructing blade tip clearance measurement data. This system aims to accurately reconstruct blade tip clearance data even when the aero-engine bladed disk is running at high speed and the sensor sampling frequency is insufficient, resulting in undersampling. It also improves anti-interference capability and adaptability to complex operating conditions.
[0008] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for reconstructing blade tip clearance measurement data, comprising the following steps: Step S1: Perform mean singular value decomposition dictionary learning on the collected leaf tip gap signal; Step S2: Use the least angle regression algorithm to perform sparse representation on the processed signal; Step S3: Determine whether the relative error meets the termination condition. If not, continue iteratively updating the dictionary and sparse coefficients until the reconstruction error is minimized.
[0009] Specifically, in step S1, an adaptive dictionary is trained by alternately executing sparse coding and dictionary update steps. Each column of the adaptive dictionary constitutes a dictionary atom, and the dictionary atom is adapted to the structural features of the leaf tip gap signal. In step S2, based on the adaptive dictionary, the least angle regression algorithm is used to sparsely represent the leaf tip gap signal under undersampling conditions. The dictionary atom most related to the residual of the signal is selected step by step, and the sparse coefficients are recursively updated in an equiangular direction to realize the sparse representation of the signal. In step S3, the relative error of the reconstruction result is evaluated. When the square of the L2 norm of the residual is less than a predetermined threshold or the number of non-zero terms of the sparse coefficient reaches the maximum limit, the iteration is terminated and the reconstructed leaf tip gap data is output.
[0010] Further, step S1 includes: S101. Initialize the dictionary: Randomly select from the collected leaf tip gap dataset. Using 10 samples as initial dictionary atoms, we obtain the initial dictionary matrix. Set the dictionary size to That is, the dictionary contains atoms; the dataset is represented as ,in For the dimensions of the data, The number of data points; the dictionary matrix is represented as Each column represents an atom of the dictionary, and the initial dictionary is denoted as . .
[0011] S102, Sparse Coding: Fixing the dictionary in each iteration The least angle regression algorithm is used to calculate the sparse coefficient matrix for each data point. This makes each data point Sparse representation using dictionary atoms: ,in It is the first The sparse coefficient vector of n data points is obtained by minimizing the loss function: The constraints are ,in The number of non-zero elements in the vector. The sparsity is set. This constraint guarantees the sparsity of the sparse coefficients, meaning that each signal is represented by only a linear combination of a few dictionary atoms.
[0012] S103. Update the dictionary: using the previously calculated sparse coefficient matrix. The dictionary is updated by performing singular value decomposition on each column of the dictionary to obtain new dictionary atoms. The dictionary update process is optimized as follows: ,in Let be the Frobenius norm, which is the square root of the sum of the squares of all elements in the matrix. The dictionary update process utilizes the mean singular value decomposition algorithm: singular value decomposition is performed on the residual matrix corresponding to the current dictionary atom; the left and right singular vectors corresponding to the largest singular value are used to update the dictionary atom, updating each atom sequentially to minimize the overall reconstruction error. Specifically, for the ... With the dictionary atoms fixed and all other atoms unchanged, only the first one is optimized. The row vectors of each atom and its corresponding sparse coefficients are used to find the update direction that minimizes the reconstruction error through singular value decomposition.
[0013] The dictionary learning and updating process described above is executed iteratively and alternately: in each iteration, sparse encoding is performed with the dictionary fixed first, and then dictionary updates are performed with the sparse coefficients fixed, and so on until the dictionary converges or the preset maximum number of iterations is reached. This alternating optimization strategy can gradually improve the matching degree between the dictionary and the training signal.
[0014] Furthermore, the dictionary learning stage in step S1 and the sparse representation stage in step S2 constitute a two-stage processing flow: the first stage trains an adaptive sparse basis that matches the characteristics of the leaf tip gap signal through mean singular value decomposition dictionary learning; the second stage uses the adaptive sparse basis and the minimum angle regression algorithm to solve the sparse coefficients to reconstruct the undersampled data.
[0015] Further, step S2 includes: S201. Initialization: Set the initial residual. ,in For the current data point to be reconstructed, This is the residual vector for the current data point. Initially, all elements of the sparse coefficient vector are set to zero, and the residual is orthogonal to all dictionary atomic vectors.
[0016] S202. Select the most relevant dictionary atom: In each iteration, calculate the current residual. The atom with the largest absolute value of the inner product with each atom in the dictionary is selected and added to the active set. ; in dictionary matrix The first in One atom, This is for inner product operations.
[0017] This step ensures that each selected atom is most correlated with the current residual, thus achieving the most efficient signal representation with the fewest atoms.
[0018] S203, Update Coefficients: The sparse coefficients of all atoms in the active set are updated jointly and recursively, so that the coefficients gradually increase along the equiangular direction while minimizing the reconstruction error; assuming the selected atom set is... Then the formula for updating the coefficients is: ; in For the selected subset of dictionary atoms, This represents the corresponding sparse coefficient vector. The characteristic of the least angle regression algorithm is that in each iteration, the coefficients are gradually increased along an equal angle direction, rapidly reducing the residual while maintaining sparsity.
[0019] S204, Update Residual: Subtract the contribution of the selected dictionary atoms to the data from the current residual. This yields the updated residual vector.
[0020] Steps S202 to S204 are executed cyclically. In each cycle, a new dictionary atom is selected to be added to the active set and the coefficients are updated until the termination condition in step S3 is met.
[0021] Furthermore, in step S3, the termination condition is one of the following two: the squared L2 norm of the residual. Less than the predetermined threshold The reconstruction accuracy has met the requirements; or the number of non-zero terms in the sparse coefficients has reached the preset maximum limit. This indicates that the sparsity has reached its upper limit. When any condition is met, the iteration stops, and the current sparsity coefficients and the reconstructed signal are output.
[0022] The above method is applied to the measurement of blade tip clearance in aero-engines. The blade tip clearance signal is acquired by a capacitive sensor. The capacitive sensor reflects the dynamic characteristics of the clearance by capturing the capacitance change between the blade tip and the casing. The method is performed under undersampling conditions, with a sampling frequency lower than the Nyquist sampling frequency.
[0023] Secondly, the present invention also provides a blade tip clearance measurement system, comprising: The sensor module is used to acquire blade tip clearance signals during the operation of the aero-engine via a capacitive sensor. The dictionary learning module is used to learn the dictionary by performing mean singular value decomposition on the collected leaf tip gap signal. An adaptive dictionary is trained by alternately performing sparse coding and dictionary updates. The sparse representation module is used to perform sparse representation of the signal to be reconstructed based on the adaptive dictionary and the least angle regression algorithm, and to progressively select the most relevant dictionary atoms and recursively update the sparse coefficients. The evaluation module is used to evaluate the relative error of the reconstruction results. When the square of the L2 norm of the residual is less than a predetermined threshold or the number of non-zero terms of the sparse coefficients reaches the maximum limit, the iteration is terminated and the reconstructed tip gap data is output.
[0024] Due to the adoption of the above technical solution, the beneficial effects of the present invention are as follows: (1) It can perform high-precision signal reconstruction using limited observation data under various environments and complex conditions, solving the problems of high-frequency sampling requirements and computational resource consumption in traditional methods. It can accurately recover tip gap data even when the sampling frequency is insufficient.
[0025] (2) This invention utilizes the least angle regression algorithm to efficiently solve for sparse coefficients, achieving high-precision data reconstruction even under undersampling conditions, enabling compressed sensing technology to recover accurate leaf tip gap data at lower sampling frequencies. The least angle regression algorithm has the advantages of low computational complexity and good numerical stability, making it particularly suitable for handling high-dimensional sparse optimization problems.
[0026] (3) The method provided by the present invention trains an adaptive dictionary and uses the minimum angle regression algorithm to calculate the sparsity coefficients, which solves the problem that traditional methods cannot effectively handle the blade tip gap measurement under low sampling frequency or high rotation speed conditions. Attached Figure Description
[0027] Figure 1 This is a flowchart illustrating a method for reconstructing blade tip clearance measurement data according to the present invention.
[0028] Figure 2 This is the raw tip gap data obtained at a high sampling rate.
[0029] Figure 3 This is a magnified view of a portion of the original data at a high sampling rate.
[0030] Figure 4 This is a comparison chart of undersampled data (red scatter dots) at a sampling frequency of 1 / 100 and the original data (blue curve).
[0031] Figure 5 This is a comparison chart of the data reconstructed using undersampled data and the original data.
[0032] Figure 6 This is a schematic diagram of undersampled data with phase shift.
[0033] Figure 7 A comparison diagram of the reconstructed signal using undersampled data with phase shift and the original signal. Detailed Implementation
[0034] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0035] This invention provides a method for reconstructing blade tip clearance measurement data, applicable to online monitoring of blade tip clearance during aero-engine operation. In this scenario, a capacitive sensor is mounted on the engine casing, capturing minute capacitance changes between the rotating blade tip and the casing to reflect the dynamic characteristics of the clearance. Because the engine operates at high speeds (blade speeds can reach tens of thousands of revolutions per minute), the signals acquired by the sensor have a high dynamic range and complex time-varying characteristics.
[0036] The method specifically includes the following steps: S1. Perform K-mean Singular Value Decomposition (K-SVD) dictionary learning on the collected leaf tip gap signal to train an adaptive dictionary. Specifically, this includes: S101. Initialize the dictionary. Randomly select from the collected leaf tip gap dataset. Using 10 samples as initial dictionary atoms, we obtain the initial dictionary matrix. The initial dictionary is denoted as Dictionary size The dataset is determined based on the complexity of the signal and the required reconstruction accuracy. ,in For the dimensions of the data, The number of data points; the dictionary matrix is represented as Each column represents an atom of the dictionary. The quality of the initial dictionary has some impact on the final learning outcome, but it will be gradually optimized during the iteration process.
[0037] S102, Sparse Coding: In each iteration, the current dictionary is fixed. The sparse coefficient matrix for each data point is calculated using the Least Angle Regression (LARS) algorithm. This makes each data point Sparse representation using dictionary atoms: ,in It is the first The sparse coefficient vector of n data points is obtained by minimizing the loss function: The constraints are ,in The number of non-zero elements in the vector; To determine the desired sparsity, the number of non-zero elements in the sparse coefficients is controlled. The minimum angle regression algorithm efficiently finds the optimal sparse coefficients that meet the given conditions. The minimum angle regression algorithm iteratively selects the dictionary atoms most relevant to the target residual and updates the coefficients in an equiangular direction, thus efficiently approximating the optimal solution while maintaining sparsity.
[0038] S103. Update the dictionary: using the previously calculated sparse coefficient matrix. The dictionary is updated by performing singular value decomposition on each column of the dictionary. ,in The Frobenius norm is used. The update process employs the mean singular value decomposition algorithm: the singular value decomposition is performed on the residual matrix corresponding to the current atom, and the left and right singular vectors corresponding to the largest singular value are used to update the dictionary atom. This process is performed column by column, updating only one atom at a time while keeping the other atoms fixed, until all dictionary atoms have been updated.
[0039] The above steps S102 and S103 are executed alternately and iteratively until the dictionary converges (i.e., the change in reconstruction error between two adjacent iterations is less than a preset threshold) or the preset maximum number of iterations is reached.
[0040] S2. Sparse representation of the processed signal is performed using the least angle regression algorithm. Specifically, based on the adaptive dictionary, the least angle regression algorithm is used to sparsely represent the leaf tip gap signal under undersampling conditions. The dictionary atoms most correlated with the residuals of the signal are selected progressively, and the sparse coefficients are recursively updated in an equal-angle direction to achieve a sparse representation of the signal. This includes: S201, Initialization. For the undersampled data points to be reconstructed. Set initial residual At this point, the sparse coefficient vector All elements are initialized to zero, the active set is empty, and the residual is orthogonal to all dictionary atoms.
[0041] S202. Select the most relevant dictionary atom: In each iteration, calculate the current residual. The atom with the largest absolute value of the inner product with each atom in the dictionary is selected and added to the active set. ; in dictionary matrix The first in One atom, This is for inner product operations.
[0042] This step ensures that each selected atom has the highest correlation with the current residual, thus achieving the most efficient signal representation with the fewest atoms. The core advantage of the least angle regression algorithm lies in its ability to efficiently identify the atom that contributes the most to the current residual.
[0043] S203, Update Coefficients: The sparse coefficients of all atoms in the active set are updated jointly and recursively, so that the coefficients gradually increase along the equiangular direction while minimizing the reconstruction error; assuming the selected atom set is... Then the formula for updating the coefficients is: ; in For the selected subset of dictionary atoms, This is the corresponding sparse coefficient vector.
[0044] The characteristic of the least-angle regression algorithm is that it gradually increases the coefficients along an equal-angle direction in each iteration, rapidly reducing the residuals while maintaining sparsity. Specifically, the algorithm calculates update paths for all coefficients in the current active set with equal-angle step sizes, and then updates these coefficients simultaneously until the correlation between the next dictionary atom and the residuals reaches the same level.
[0045] S204, Update Residual: Subtract the contribution of the selected dictionary atoms to the data from the current residual. This yields the updated residual vector.
[0046] Repeat steps S202 to S204, selecting a new dictionary atom in each loop to add to the active set and updating the coefficients, until the termination condition is met.
[0047] Step S3: Determine if the relative error meets the termination condition. If not, continue iterating and updating the dictionary and sparse coefficients until the reconstruction error is minimized. Specifically, evaluate the relative error of the reconstruction result. When the squared L2 of the residual is less than a predetermined threshold or the number of non-zero terms in the sparse coefficients reaches the maximum limit, terminate the iteration and output the reconstructed leaf tip gap data.
[0048] The termination condition is one of the following two: the square of the L2 norm of the residuals. Less than the predetermined threshold The reconstruction accuracy has met the requirements; or the number of non-zero terms in the sparse coefficients has reached the preset maximum limit. This indicates that the sparsity has reached its upper limit. When any condition is met, the iteration stops, and the current sparsity coefficients and the reconstructed signal are output.
[0049] This embodiment also provides a blade tip clearance measurement system, including: The sensor module is used to acquire blade tip clearance signals during the operation of the aero-engine via a capacitive sensor. The dictionary learning module is used to learn the dictionary by performing mean singular value decomposition on the collected leaf tip gap signal. An adaptive dictionary is trained by alternately performing sparse coding and dictionary updates. The sparse representation module is used to perform sparse representation of the signal to be reconstructed based on the adaptive dictionary and the least angle regression algorithm, and to progressively select the most relevant dictionary atoms and recursively update the sparse coefficients. The evaluation module assesses the relative error of the reconstruction results. It terminates the iteration and outputs the reconstructed tip clearance data when the squared L2 of the residual is less than a predetermined threshold or the number of non-zero terms in the sparse coefficients reaches a maximum limit. To verify the effectiveness of the present invention, the following experiment was conducted.
[0050] Experimental data: Signals of blade tip clearance from an aero-engine were collected, with a data length N = 3200. An adaptive dictionary was trained using the mean-singular-value decomposition dictionary learning method, and the dictionary size K was set according to the signal characteristics. The sampling matrix was constructed at 1 / 100 of the normal sampling frequency, i.e., acquiring undersampled data at an extremely low sampling rate. This means that only about 32 sampling points were selected from the 3200 data points for reconstruction, with a sampling rate only one percent of the normal sampling frequency. Experimental results are as follows: Figures 2 to 7 As shown.
[0051] Figure 2The original tip gap data obtained at a high sampling rate are shown. The signal exhibits clear periodicity, reflecting the gap variation during blade rotation. Figure 3 () is a magnified view of a portion of the original data, allowing for a more detailed observation of the signal's waveform characteristics and structural details. Figure 4 The comparison between undersampled data (red scatter dots) and original data (blue curve) at a sampling frequency of 1 / 100 shows that the undersampled data retains only a very small number of original signal sample points, and traditional interpolation methods cannot recover meaningful signal waveforms from it. Figure 5 The comparison between the reconstructed signal from undersampled data using the method of the present invention and the original data is shown. It can be seen that the reconstructed signal is basically the same as the original signal with a very small error, which verifies the excellent reconstruction capability of the method of the present invention at extremely low sampling rates. Figure 6 This demonstrates the case of undersampled data with phase shift, where the sampling time has a phase deviation relative to the original signal. Figure 7 The paper presents a comparison between the reconstructed signal using undersampled data with phase shift and the original signal. The results show that even with phase shift, the method of the present invention can still achieve high-precision signal reconstruction, demonstrating the robustness of the method.
[0052] The experimental results above show that, under the learning of mean singular value decomposition dictionary, combined with the compressed sensing framework and the minimum angle regression algorithm, the leaf tip gap data can be successfully reconstructed at an extremely low sampling rate (1 / 100 of the normal sampling frequency) with a small reconstruction error, which verifies the effectiveness and superiority of the method of this invention.
[0053] To further illustrate the technical advantages of this invention, the following analysis will focus on the algorithm principles.
[0054] The core idea of K-SVD dictionary learning is to progressively optimize dictionary atoms by alternately performing sparse coding and dictionary updates, enabling them to better match the local structural features of the training signal. Compared to traditional fixed basis methods (such as DCT and wavelet basis methods), the dictionary learned by K-SVD has stronger adaptability, achieving more accurate sparse representation with fewer atoms. This is particularly important for leaf tip gap signals, as these signals typically exhibit complex non-stationary characteristics and multi-scale structures, which fixed basis functions struggle to effectively capture. The key innovation of the K-SVD algorithm lies in the dictionary update step: instead of updating the entire dictionary simultaneously as in traditional methods, it updates atom by atom, optimizing only one atom and its corresponding sparse coefficients at a time, finding the optimal update direction through singular value decomposition. This column-by-column update strategy ensures that each update effectively reduces reconstruction error.
[0055] The core idea of the Least Angle Regression (LARS) algorithm is to efficiently approximate the optimal solution while maintaining sparsity by progressively selecting the features (dictionary atoms) most relevant to the target residual and updating the coefficients in an equiangular direction. The LARS algorithm has the advantages of low computational complexity and good numerical stability, making it particularly suitable for handling high-dimensional sparse optimization problems. Within the framework of this invention, the LARS algorithm is used in conjunction with an adaptive dictionary learned from K-SVD, fully leveraging the advantages of both: the adaptive dictionary provides a sparse representation basis that highly matches the signal, while the LARS algorithm efficiently finds the sparsest representation of the signal under this basis. The combination of the two achieves high-precision reconstruction of the leaf tip gap signal under undersampling conditions.
[0056] The overall technical process of this invention is as follows: First, an adaptive dictionary is learned from a large number of training signals using the K-SVD algorithm; then, the dictionary and the LARS algorithm are used to sparsely reconstruct the undersampled observation data; finally, the reconstruction quality is judged through an error evaluation mechanism, and a result that meets the accuracy requirements is output. This scheme fundamentally solves the problem of insufficient measurement accuracy of traditional methods under undersampled conditions, and provides an efficient and reliable technical solution for real-time monitoring of aero-engine blade tip clearance.
[0057] The method of this invention can also be applied to clearance measurement scenarios in other types of rotating machinery, such as gas turbines and turbochargers. These devices also face the challenge of clearance monitoring under high-speed rotation conditions.
[0058] Furthermore, the parameters involved in the method of this invention (such as dictionary size) sparsity Number of iterations, termination threshold Number of maximum sparsity coefficients The dictionary size and sparsity can be flexibly adjusted according to the actual application scenario and requirements. For example, in scenarios with high real-time requirements, the dictionary size and sparsity can be appropriately reduced to speed up the calculation; in scenarios with high accuracy requirements, these parameters can be increased to obtain better reconstruction quality. This parameter adjustability further enhances the practicality and adaptability of the method of the present invention.
[0059] Those skilled in the art will understand that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications and variations can be made to the present invention without departing from its core principles. For example, the specific parameters of dictionary learning can be adjusted according to the actual application scenario; the threshold of the termination condition can also be flexibly set according to accuracy requirements; the minimum angle regression algorithm can also be replaced by other sparse optimization algorithms (such as orthogonal matching pursuit, basis pursuit, etc.), and these modifications and variations all fall within the scope of protection of the present invention.
[0060] Contents not described in detail in this specification are existing technologies known to those skilled in the art. The specific embodiments described above further illustrate the purpose, technical solutions, and beneficial effects of this invention. It should be understood that the above descriptions are merely specific embodiments of this invention and are not intended to limit the scope of protection of this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A method of tip clearance measurement data reconstruction, characterized by the steps of include: Step S1: Perform mean singular value decomposition dictionary learning on the collected leaf tip gap signal; Step S2: Use the least angle regression algorithm to perform sparse representation on the processed signal; Step S3: Determine whether the relative error meets the termination condition. If not, continue iteratively updating the dictionary and sparse coefficients until the reconstruction error is minimized.
2. The method for reconstructing tip clearance measurement data according to claim 1, characterized in that: In step S1, an adaptive dictionary is trained by alternately executing sparse coding and dictionary update steps. Each column of the adaptive dictionary constitutes a dictionary atom, and the dictionary atom is adapted to the structural features of the leaf tip gap signal. In step S2, based on the adaptive dictionary, the least angle regression algorithm is used to sparsely represent the leaf tip gap signal under undersampling conditions. The dictionary atom most related to the residual of the signal is selected step by step, and the sparse coefficients are recursively updated in an equiangular direction to realize the sparse representation of the signal. In step S3, the relative error of the reconstruction result is evaluated. When the square of the L2 norm of the residual is less than a predetermined threshold or the number of non-zero terms of the sparse coefficient reaches the maximum limit, the iteration is terminated and the reconstructed leaf tip gap data is output.
3. The tip clearance measurement data reconstruction method of claim 2, wherein, Step S1 includes: S101, initializing a dictionary: randomly selecting one sample from the collected tip clearance data set as an initial dictionary atom, obtaining an initialized dictionary matrix , setting the dictionary size to , that is, the dictionary contains atoms; the data set is represented as , wherein is the dimension of the data, and is the number of data points; the dictionary matrix is represented as , each column being an atom of the dictionary, and the initial dictionary being denoted as ; S102, sparse coding: fixing the dictionary in each iteration , compute the sparse coefficient matrix of each data point using the least angle regression algorithm , so that each data point is sparsely represented by the dictionary atoms: , where is the sparse coefficient vector of the th data point, is solved by minimizing the loss function: , with the constraint , where is the number of non-zero elements of the vector, is the set sparse degree; S103, updating the dictionary: using the computed sparse coefficient matrix updating the dictionary by performing singular value decomposition on each column atom of the dictionary to update the dictionary atoms: where is the Frobenius norm.
4. The tip clearance measurement data reconstruction method according to claim 3, characterized in that, In step S103, the dictionary update process utilizes the mean singular value decomposition algorithm: singular value decomposition is performed on the residual matrix corresponding to the current dictionary atom, and the left singular vector and right singular vector corresponding to the maximum singular value are taken to update the dictionary atom. Each atom is updated sequentially to minimize the overall reconstruction error.
5. The tip clearance measurement data reconstruction method of claim 2, wherein, The dictionary learning stage in step S1 and the sparse representation stage in step S2 constitute a two-stage processing flow: the first stage trains an adaptive sparse basis that matches the characteristics of the leaf tip gap signal through mean singular value decomposition dictionary learning; the second stage uses the adaptive sparse basis and the minimum angle regression algorithm to solve the sparse coefficients to reconstruct the undersampled data.
6. The tip clearance measurement data reconstruction method of claim 3, wherein, Step S2 includes: S201, initialization: set initial residual wherein is the current data point to be reconstructed, is the residual vector of the current data point, all elements of the initial sparse coefficient vector are set to zero, and the residual is orthogonal to all dictionary atom vectors; S202, select the most relevant dictionary atom: in each iteration, compute the inner product of the current residual with each atom in the dictionary, select the atom with the largest absolute value of the inner product to add to the active set: ; wherein is a dictionary matrix of the th atom, is an inner product operation; S203, updating the coefficients: jointly updating the sparse coefficients of all atoms in the active set in a recursive manner, so that the coefficients gradually increase along the isometric direction while minimizing the reconstruction error; assuming that the atom set has been selected as The formula for updating the coefficients is: ; wherein is the selected subset of dictionary atoms, is the corresponding sparse coefficient vector; S204, update the residual: subtract the contribution of the selected dictionary atom pair data from the current residual: to obtain the updated residual vector.
7. The tip clearance measurement data reconstruction method according to claim 6, characterized in that, Steps S202 to S204 are executed cyclically. In each cycle, a new dictionary atom is selected to be added to the active set and the coefficients are updated until the termination condition in step S3 is met.
8. The tip clearance measurement data reconstruction method of claim 6, wherein, In step S3, the termination condition is one of the following two: The square of the two-norm of the residual Less than a predetermined threshold , the reconstruction accuracy is satisfied; or The number of non-zero terms of the sparse coefficient reaches a preset maximum limit value , indicating that the sparsity has reached an upper limit.
9. The tip-tolerance measurement data reconstruction method of claim 1, wherein, The method is applied to the measurement of blade tip clearance in aero-engines. The blade tip clearance signal is acquired by a capacitive sensor. The capacitive sensor reflects the dynamic characteristics of the clearance by capturing the capacitance change between the blade tip and the casing. The method is performed under undersampling conditions, with a sampling frequency lower than the Nyquist sampling frequency.
10. A tip clearance measurement system, characterized by include: The sensor module is used to acquire blade tip clearance signals during the operation of the aero-engine via a capacitive sensor. The dictionary learning module is used to learn the dictionary by performing mean singular value decomposition on the collected leaf tip gap signal. An adaptive dictionary is trained by alternately performing sparse coding and dictionary updates. The sparse representation module is used to perform sparse representation of the signal to be reconstructed based on the adaptive dictionary and the least angle regression algorithm, and to progressively select the most relevant dictionary atoms and recursively update the sparse coefficients. The evaluation module is used to evaluate the relative error of the reconstruction results. When the square of the L2 norm of the residual is less than a predetermined threshold or the number of non-zero terms of the sparse coefficients reaches the maximum limit, the iteration is terminated and the reconstructed tip gap data is output.
