High-precision particle image velocity measurement method of self-adaptive regulation and control mechanism

Abstract

The invention discloses a high-precision particle image velocity measurement method based on a self-adaptive regulation and control mechanism, and relates to the field of particle image velocity measurement, and the method specifically comprises the steps: 1, carrying out the preprocessing of an original image sequence through collecting a particle image, taking a key index as a feedback quantity, dynamically adjusting a collection parameter, and screening an effective frame; step 2, delay embedding is carried out on the pixel gray scale time sequence of the standardized effective frame to construct a matrix, a low-rank subspace is determined, coarse displacement is projected to the subspace to obtain an initial displacement field, a joint cost function is constructed to optimize the displacement field and window parameters, and the optimized displacement field is determined; and step 3, performing phase-space reconstruction on the optimized displacement time sequence to obtain a matrix, building VAE training, optimizing by taking reconstruction loss, KL divergence and total correlation as loss functions, screening a main mode, and outputting a final result. Through phase-space reconstruction and a variational auto-encoder mode unmixing method, main vibration and background disturbance as well as artifacts and random noise are distinguished; and the related peak quality is substantially improved.

Description

Technical Field

[0001] This invention belongs to the field of particle image velocimetry technology, specifically relating to a high-precision particle image velocimetry method with an adaptive control mechanism. Background Technology

[0002] Particle image velocimetry (PIV), a major fluid velocity field measurement technique, relies on the collaborative work of a light source, camera, and image processing software to reconstruct the velocity field by tracking the inter-frame displacement of tracer particles. Although the PIV technology is relatively mature, several sources of error can still be introduced in practical testing: first, complex scattering and local occlusion caused by differences in flow facilities and boundary conditions; second, drift and mismatch in the experimental system's optical path, camera settings, and synchronization triggering; third, uneven particle density, depth-of-field variations, image blurring, and underexposure during image recording; and fourth, limitations in data evaluation methods regarding cross-correlation window selection, parameter fixation, sub-pixel interpolation, and robust regularization. Furthermore, digital camera background noise, photon statistical noise, and equipment electronic noise all reduce the signal-to-noise ratio of the correlation peak and distort its shape, directly affecting the accuracy of displacement estimation. The aforementioned problems are particularly prominent in high-speed and micro-scale flow, low-light and sub-pixel displacement scenarios. The traditional process of "cross-correlation + optical flow + fixed parameters + linear noise reduction" often suffers from matching drift, window mismatch, noise dominance and unstable performance across scenarios, making it difficult to consistently obtain high-precision and verifiable velocity field results.

[0003] Therefore, there is an urgent need for a high-precision particle image velocimetry method with an adaptive control mechanism to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to provide a high-precision particle image velocimetry method with an adaptive control mechanism to solve the technical problem of difficulty in continuously obtaining high-precision and verifiable velocity field results in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: High-precision particle image velocimetry methods with adaptive control mechanisms include: Step 1: By acquiring particle images and dynamically adjusting the acquisition parameters with key indicators as feedback, the original image sequence is preprocessed and valid frames are selected. Step 2: Perform delayed embedding matrix construction on the pixel grayscale temporal sequence of the standardized effective frame, determine the low-rank subspace, project the coarse displacement onto the subspace to obtain the initial displacement field, construct the joint cost function to optimize the displacement field and window parameters, and determine the optimized displacement field. Step 3: Reconstruct the phase space matrix of the optimized displacement time series, build a VAE training system, optimize the loss function using reconstruction loss, KL divergence, and total correlation, select the main modes, and output the final results.

[0006] Furthermore, collect particle images and dynamically adjust the acquisition parameters with the key indicators as the feedback quantities. The specific method is as follows: Select a high-speed camera that meets the basic requirements of PIV technology; for fluids, use tracer particles that match the physical properties; if the natural texture imaging on the solid surface meets the standard, directly use it, otherwise, form a trackable feature by spraying random textures or attaching particles. Set the initial value of the exposure time as T1 and set the sampling frequency according to the flow field type; Calculate the cross-correlation peak shape index and the coarse displacement △u between adjacent frames in real time. The cross-correlation peak shape index includes the peak height and the full width at half maximum, and calculate the cross-correlation peak signal-to-noise ratio Sc; When △u > k×w, increase the sampling frequency or shorten the exposure time until △u ≤ k×w, where w is the side length of the cross-correlation window and k is a preset proportional threshold; When Sc < s1, increase the illumination intensity, increase the camera gain or extend the exposure time until Sc ≥ s1. s1 is the preset system minimum signal-to-noise ratio threshold. Record the parameter adjustment information to form an acquisition log, and use the particle image after parameter adjustment as the original image sequence.

[0007] Furthermore, preprocess the original image sequence and screen out the valid frames. The specific method is as follows: Through flat-field correction, background subtraction, and distortion correction, obtain the image sequence after geometric and gray-scale correction; calculate the sharpness of the corrected image, that is, MTF(0.5) at the Nyquist frequency, calculate the background gray-scale variance ab of the corrected image, and set the sharpness threshold M1 and the background noise threshold b1 based on the conventional PIV historical data and the actual situation; Only retain the valid frames with sharpness ≥ M1 and ab ≤ b1, de-trend and normalize the valid frames, record the frame deletion rate, and output the standardized valid frames.

[0008] Furthermore, perform delay embedding on the pixel gray-scale time series of the standardized valid frames to form a matrix and determine the low-rank subspace. The specific method is as follows: Record the standardized valid frames as , where t represents different time points and is used as the index of the valid frames, T is the total number of frames. Determine the pixel coordinates (x, y) of each valid image frame, and for the gray-scale time series of each pixel Perform delay embedding and splice the embedding vectors to form a delay embedding matrix , where d represents the embedding dimension, which is determined by the false nearest neighbor ratio, and R represents the set of real numbers; Perform SVD decomposition on , that is = UΣV^T, where U is the left singular matrix, V^T is the right singular matrix, Σ is the diagonal matrix, and its diagonal elements are The singular values, where i = 1, 2, ..., min(T-d+1, d), are set with a threshold. Where ω(β) is a threshold coefficient based on random matrix theory, β=K / N, and K is... The number of columns; N is the number of rows, The median and rank of the singular values ​​are represented by the median and rank of the singular values. To meet The number of singular values, and the corresponding low-rank subspace. For the front of U A matrix composed of columns; the subspace is updated in real time using incremental SVD, as shown in the formula. ,in, Let be the subspace at time t. For the subspace of the previous moment, The first step obtained by SVD decomposition at time t Left singular matrix, The forgetting factor is set based on historical data and actual needs.

[0009] Furthermore, the coarse displacement is projected onto the subspace to obtain the initial displacement field, specifically as follows: Coarse displacement estimation is obtained using the basic cross-correlation method. Then, the coarse displacement estimate is projected onto the low-rank subspace, according to the formula. The initial low-rank displacement field u(0) is obtained, where Let represent the orthogonal basis matrix of the low-rank subspace. This is the result of coarse displacement estimation.

[0010] Furthermore, a joint cost function is constructed to optimize the displacement field and window parameters, and the optimized displacement field is determined. The specific method is as follows: Using formula Construct a joint cost function, where u is the displacement field, θ is the affine parameter of the deformable window, including scaling, shearing, rotation, and translation parameters; Jc(u, θ) is the normalized negative peak value of cross-correlation, Jo(u) and Js(u) are the mean values ​​of optical flow reprojection error, Ji(u) is the mean value of the squared gradient of the displacement field, and Jm(u) is the consistency error between the displacement field and the low-rank subspace. , , , , These represent the weighting coefficients of each cost item, set according to historical data and actual conditions; Gradient descent is used to jointly optimize θ and u. During the iteration process, the step size and number of iterations are adjusted to optimize the cost function. It gradually converges to the minimum value; Combining a multi-scale pyramid strategy, the top layer calculates the coarse displacement field based on u(0), which is then upsampled and used as the initial solution for the next layer, up to the original resolution layer. The original resolution layer performs sub-pixel correction through cubic spline interpolation, and the fusion result yields the optimized displacement field. .

[0011] Furthermore, the phase space is reconstructed from the optimized displacement time series to obtain the matrix, and a VAE is built for training. The specific method is as follows: Let the optimized displacement field time series be denoted as... The sequence length is N. The optimal delay J is determined by the minimum of the average mutual information, and the optimal dimension D is determined by the false nearest neighbor criterion. Embedding vectors are constructed based on J and D, and arranged to form a delay embedding matrix X with a dimension of (T-(D-1)J)×D. A VAE is built, and the encoder is a two-layer fully connected network that outputs the mean of the latent variables. Given the log-variance logσ², the decoder is a symmetric fully connected network that outputs a reconstruction matrix Xh.

[0012] Furthermore, the reconstruction loss, KL divergence, and total relevance are used as loss functions for optimization. The specific method is as follows: Let the total loss function be Lt = Lr + Lk + f × Lc (where f is the total relevance weight); Where Lr represents the reconstruction loss, Lk represents the KL divergence, and Lc represents the total correlation. Mean squared error is used to measure the difference between the original matrix X and the reconstructed matrix Xh, denoted as reconstruction loss, and the specific formula is denoted as follows: , where Nx is the total number of elements in X.

[0013] By extracting the mean of the latent variables from the encoder output With log-variance logσ², for each latent variable dimension r, ; A three-layer fully connected discriminant is constructed and trained to distinguish between the true latent variable distribution and the independent latent variable distributions in each dimension. Using the trained discriminant, the dependence of the true latent variable distribution on the independent distribution is calculated; this dependence is the total correlation coefficient, where f is the weighting coefficient, set based on historical data and the actual situation. In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention robustly improves the accuracy of displacement and velocity field reconstruction by introducing a collaborative optimization framework of modal guidance and residual feedback in sub-pixel level displacement estimation. Through phase space reconstruction and modal unmixing methods, it distinguishes between the main vibration / main flow components and background disturbances, artifacts and random noise, significantly improving the quality of correlation peaks and estimation reliability. It achieves parameter adaptation and process closed-loop control under different velocity scales, particle densities and imaging conditions, making the method applicable to high-precision velocimetry of conventional flows, wide velocity ranges and micro flows. 2. This invention achieves efficient dynamic displacement field estimation of standardized effective frame pixel grayscale temporal sequence through delayed embedding matrix construction, low-rank subspace determination, joint cost function optimization, and multi-scale pyramid strategy. The output results have physical rationality (low-rank subspace constraint), temporal consistency (optical flow reprojection constraint), and spatial smoothness (gradient constraint). It has significant application value in fields such as unsteady flow field analysis, large displacement scene matching, and real-time monitoring. At the same time, the resource consumption is reduced by optimizing computational efficiency (incremental SVD, multi-scale pyramid), which improves engineering practicality. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 The diagram illustrates the steps of the high-precision particle image velocimetry method with adaptive control mechanism of the present invention. Figure 2 The diagram illustrates the steps of the method for determining the optimized displacement field according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1 , Figure 2 The high-precision particle image velocimetry method with adaptive control mechanism shown includes the following steps: Step 1: By acquiring particle images and dynamically adjusting the acquisition parameters with key indicators as feedback, the original image sequence is preprocessed and valid frames are selected. Select a high-speed camera with a full resolution of at least A×B pixels and a grayscale depth of at least H1 (refer to the minimum requirements of PIV for grayscale resolution). The focal length and working distance of the lens need to be adjusted so that the equivalent pixel diameter of the particles is H2, and the particle areal density is H3 (determined by statistically analyzing the cross-correlation signal-to-noise ratio at different densities). For fluids, use tracer particles that match the physical properties; for solids, if the natural texture on the surface meets the imaging standard, directly utilize it, otherwise, form traceable features by spraying random textures or attaching particles. The lighting system uses pulsed lasers or high-power LEDs, combined with flat-field illumination and dark-field masking to reduce background non-uniformity. Synchronize the camera and lighting through TTL / FPGA triggering. Set the initial value of the exposure time to T1, and set the sampling frequency according to the flow field type (for example, for conventional flows: 100 - 500 fps; for high-speed flows: 1000 - 5000 fps). Calculate the cross-correlation peak shape indexes of two adjacent frames in real time, including peak height and full width at half maximum. By accumulating the grayscale projections of the image horizontally and vertically, statistically analyze the offset of the projection curve as the coarse displacement △u between frames, determine the standard deviation of the cross-correlation background noise, and calculate the cross-correlation peak signal-to-noise ratio Sc through the peak height and the standard deviation of the cross-correlation background noise. When △u > k×w, increase the sampling frequency or shorten the exposure time until △u ≤ k×w, where w is the side length of the current cross-correlation window, and the k value is set as the proportional threshold for judging whether the inter-frame displacement exceeds the window side length, preset according to the actual application scenario. When Sc < s1 (in this embodiment, s1 is set as an empirical threshold of 6), increase the lighting intensity, camera gain (increase by 1 - 2 dB each time to avoid introducing electronic noise due to excessive gain), or extend the exposure time until Sc ≥ s1, where s1 is the lowest signal-to-noise ratio threshold acceptable for the system. Below this value, it will lead to matching failures or a significant increase in the false detection rate. Record the parameter adjustment time, the values before and after adjustment, and the triggering conditions in real time to form an acquisition log for subsequent result auditing. Use the particle images collected after completing the parameter adjustment as the original image sequence.

[0018] Perform preprocessing and quality screening on the original image sequence. First, use a standard flat-field plate to take a background image. Subsequently, divide each frame of the original image by this background image to achieve flat-field correction. Use the method of Gaussian blur to determine the background of the image, and control the smoothness of the extracted background by adjusting the intensity of Gaussian blur. Then subtract this estimated background from the flat-field corrected image to highlight the foreground particles. Calibrate the camera in advance using a two-dimensional calibration board (such as a checkerboard) to obtain a set of accurate lens distortion parameters. Then, use the distortion correction function provided in the open-source computer vision library or mathematical software to process each frame of the image according to this set of parameters, and finally obtain a corrected image sequence with accurate geometric shapes.

[0019] The modulation transfer function (MTF(0.5)) at the Nyquist frequency is calculated for the corrected image sequence. The calculated MTF(0.5) is recorded as the sharpness. The background gray variance ab is calculated. Based on the historical data of the conventional PIV scene, the sharpness threshold M1 and the background noise threshold b1 are set. Only when the sharpness of a frame is not lower than the set threshold and its background noise level is not higher than the set threshold (this threshold is related to the inherent background noise of the camera) is it determined to be a valid frame. Abnormal frames that do not meet the conditions will be automatically removed. The retained valid frames are detrended (to eliminate slow grayscale drift) and normalized, the frame deletion rate (number of frames removed / total number of frames) is recorded, and the normalized valid frames are output.

[0020] Step 2: Perform delayed embedding matrix construction on the pixel grayscale temporal sequence of the standardized effective frame, determine the low-rank subspace, project the coarse displacement onto the subspace to obtain the initial displacement field, construct the joint cost function to optimize the displacement field and window parameters, and determine the optimized displacement field. The standardized valid frame is denoted as Where t represents different time points, serving as the index of valid frames (e.g., t=1, 2, ..., T, where T is the total number of valid frames), used to distinguish valid image frames acquired at different time points, determine the pixel coordinates (x, y) of each valid image frame, and generate a grayscale time series for each pixel. Delayed embedding is performed, with embedding parameters including delay τ and embedding dimension d. τ is determined by minimizing the first order of average mutual information (AMI), and d is determined by reducing the false nearest neighbor ratio (FNN) to below 5%. The delayed embedding vectors of all pixels are concatenated into a delayed embedding matrix. , ∈R^((T-d+1)×d), where R represents the set of real numbers and T-d+1 represents the number of vectors generated by delayed embedding; right Perform SVD decomposition, i.e. =UΣV^T, where U is a left singular matrix, V^T is a right singular matrix, and Σ is a diagonal matrix with diagonal elements for The singular values, where i = 1, 2, ..., min(T-d+1, d), are set with a threshold. Where ω(β) is the threshold coefficient based on random matrix theory (determined by looking up β in a table; for a square matrix, ω(β) = 2.858), β = K / N, and K is... The number of columns; N is the number of rows, The median and rank of the singular values ​​are represented by the median and rank of the singular values. To meet The number of singular values, and the corresponding low-rank subspace. For the front of U A matrix composed of columns; To adapt to the dynamic changes in the flow field, the incremental SVD method is used to update the low-rank subspace in real time, as shown in the following formula: ,in, Let be the subspace at time t. For the subspace of the previous moment, Let be the left singular matrix of the first γ columns obtained by SVD decomposition at time t, and ρ be the forgetting factor (set according to historical data and actual needs, and is 0.985 in this embodiment).

[0021] A coarse displacement estimate is obtained through the basic cross-correlation method. This coarse displacement estimate is then projected onto a low-rank subspace to obtain the initial low-rank displacement field u(0), as shown in the specific formula. ,in Let represent the orthogonal basis matrix of the low-rank subspace. This is the result of coarse displacement estimation.

[0022] By constructing a joint cost function that integrates multiple physical constraints and combining it with a deformable window model, we can achieve coordinated optimization of the displacement field and window shape, balancing matching accuracy and physical rationality. The specific formula for the joint cost function is as follows: ; in This represents the overall joint cost function, which comprehensively reflects the optimization degree of the displacement field and window parameters. This represents the displacement field to be solved. The affine parameter set for a deformable window specifically includes the scaling factor in the x-direction, the scaling factor in the y-direction, the shearing factor, the rotation angle, the translation amount in the x-direction, and the translation amount in the y-direction. This represents the normalized negative cross-correlation peak value of the deformable window. This represents the mean optical flow reprojection error, which is obtained by calculating the average difference between the grayscale value of a pixel in the current frame and the grayscale value of the corresponding pixel at the same displacement position in the previous frame. This represents the mean optical flow reprojection error, which is obtained by calculating the average difference between the grayscale value of a pixel in the current frame and the grayscale value of the corresponding pixel at the same displacement position in the previous frame. This represents the mean of the sum of squares of the displacement field gradient, which is obtained by calculating the average of the sum of squares of the gradients of the displacement components of adjacent pixels. This represents the consistency error between the displacement field and the low-rank subspace, which is obtained by calculating the degree of deviation between the displacement field and the low-rank subspace. , , , , These represent the weighting coefficients of each cost item, set based on historical data and actual conditions.

[0023] Gradient descent is used to jointly optimize θ and u. During the iteration process, the step size and number of iterations are adjusted to optimize the cost function. It gradually converges to the minimum value.

[0024] By employing a multi-scale pyramid strategy, the displacement field is progressively optimized from low to high resolution. Fine-tuning is achieved through sub-pixel interpolation, balancing computational efficiency and estimation accuracy. This generates an S-layer Gaussian pyramid, with each layer obtained from the upper layer image through Gaussian blurring and downsampling. This captures flow field features at different scales. At the top of the pyramid, the cost function is solved based on the initial displacement field u(0). The coarse displacement field is obtained and upsampled as the initial solution for the next layer. This optimization process is repeated until the original resolution layer is reached, ensuring matching accuracy in large displacement scenarios. At the original resolution layer, cubic spline interpolation is used to perform sub-pixel-level correction on the displacement field. The optimization results from each layer are then fused to obtain the final optimized displacement field. .

[0025] Step 3: Reconstruct the phase space matrix of the optimized displacement time series, build a VAE training system, optimize the loss function using reconstruction loss, KL divergence, and total correlation, select the main modes, and output the final results.

[0026] Optimize displacement field The time series is denoted as , representing the displacement value at time t (such as the velocity component in optical flow estimation), with a sequence length of N, calculated based on average mutual information. Its delayed sequence Mutual information, delayed sequence By sequence The optimal delay is obtained by shifting backward by j time steps, and the j at which the mutual information first reaches its minimum value is selected as the optimal delay. Using the pseudo-nearest neighbor (FNN) criterion, initialize d=1, and calculate the phase space vector F(t) = For each vector F(t), find its nearest neighbor F(a), increase the dimension to d+1, obtain a new vector and find its nearest neighbor, calculate the distance ratio Rd, if Rd>Rt, then F(a) is considered a false nearest neighbor in the low-dimensional space, Rt is the distance ratio threshold, set according to historical data and actual situation, repeat the above steps, calculate the FNN rate (proportion of false nearest neighbors) under different dimensions d, select the smallest d that drops below 5% and tends to be stable (difference between adjacent dimensions of FNN ≤1%) as the optimal dimension, and at the same time satisfy d≤T / 10 (to avoid overfitting). Based on the determined optimal delay J and optimal dimension D, the sequence is... Construct embedding vectors, each containing the displacement values ​​of the current time and the previous (D-1) delayed time. Arrange all embedding vectors in rows to form a delayed embedding matrix X with dimensions (T-(D-1)J)×D, where rows represent the number of samples and columns represent the embedding dimensions.

[0027] Using the delayed embedding matrix X as input, feature extraction is performed through a two-layer fully connected network. The first layer contains 256 neurons and employs the ReLU activation function to enhance nonlinear fitting capabilities; the second layer contains 128 neurons and employs the LeakyReLU activation function to alleviate the gradient vanishing problem. The final output consists of two vectors: the mean μ and the log-variance logσ² of the latent variables (with the same dimension as μ).

[0028] The latent variable z (sampled from μ and logσ² using a reparameterization technique) is used as input and reconstructed through a two-layer fully connected network symmetric to the encoder. The first layer contains 128 neurons and uses the LeakyReLU activation function; the second layer contains 256 neurons and uses the ReLU activation function. The final output is a reconstructed matrix Xh (with the same dimensions as the input X), which is the reconstruction result of the original delayed embedding matrix.

[0029] The total loss function is Lt = Lr + Lk + f × Lc; Where Lr represents the reconstruction loss, Lk represents the KL divergence, and Lc represents the total correlation.

[0030] Mean squared error is used to measure the difference between the original matrix X and the reconstructed matrix Xh, denoted as reconstruction loss, and the specific formula is denoted as follows: , where Nx is the total number of elements in X.

[0031] By extracting the mean μ and log-variance logσ² of the latent variables from the encoder output, for each latent variable dimension r, ; Construct a 3-layer fully connected discriminator (input is latent variable samples, output is the probability that the sample comes from the "true latent variable distribution"), train the discriminator to distinguish between the true latent variable distribution and the independent latent variable distributions of each dimension; use the trained discriminator to calculate the degree of dependence of the true latent variable distribution on the independent distribution, which is the total correlation, and f is the weight coefficient, which is set according to historical data and actual situation. An ADM optimizer was used for parameter updates with a learning rate of 0.001 to balance convergence speed and stability; the batch size was set to 128 to reduce training fluctuations; the number of training epochs was set to 100-300 epochs, and a KL divergence linear warm-up strategy was adopted (the KL loss weights were gradually increased from 0 to 1 in the first 20 epochs) to avoid excessive suppression of reconstruction accuracy in the early stages. An early stopping mechanism was introduced, which stopped training when the reconstruction loss on the validation set did not decrease for 20 consecutive epochs to prevent overfitting. After training, the dimensions of other latent variables in the decoder are fixed to have a mean of 0. Only a single latent variable is scanned within its distribution range (e.g., from -3σ to 3σ), and the corresponding reconstruction matrix Xhat(z) is generated by the decoder. The reduction ratio of reconstruction error for each z is calculated (i.e., leave-one-out reconstruction evaluation). The latent variable with the highest reduction ratio and whose corresponding modal spectral features (e.g., main peak frequency) fall within the physical prior bandwidth is retained. The reconstruction result corresponding to this latent variable is the main flow / main vibration mode that needs to be retained. The modes corresponding to the remaining latent variables are determined as background perturbations or image artifacts.

[0032] The extracted candidate modes are quantitatively screened, and the main modes with significant physical significance are retained. The variance of each mode is calculated as a proportion of the total variance of all modes (i.e., energy proportion E). Only modes with E ≥ 5% are retained to ensure that their contribution to the overall flow characteristics is significant. The power spectrum is obtained by performing a Fourier transform on the modes to identify the main peak frequency; the sliding variance of the main peak frequency is calculated by using a sliding window (the window size is 5%-10% of the total number of frames), and only modes with a sliding variance ≤10%×the main peak frequency are retained to ensure stable frequency characteristics. Calculate the difference between the signal-to-noise ratio (SNR) of the mode and the SNR of the original displacement sequence, and retain only the modes with an SNR difference ≥ 3dB to ensure that the noise component in the mode is effectively suppressed; Based on prior knowledge of the flow field (such as the expected velocity range and vibration frequency bandwidth), only modes whose main peak frequency fi falls within the physically reasonable range [fmin, fmax] are retained, while pseudo-modes with no physical meaning are eliminated.

[0033] Through the above filtering, a high-precision velocity field (v=ü / Δt, where Δt is the inter-frame time interval, including x / y components, velocity magnitude, and vector diagram), dynamic modal sequence (primary modal time series curves and spectrograms), and full-process parameter log (collection / filtering / estimation / demixing parameters) are output.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0035] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A high-precision particle image velocimetry method with adaptive control mechanism, characterized in that, Comprise: Step one, by collecting particle image, and with key indicators as feedback dynamic adjustment of acquisition parameters, pre-processing and screening effective frame of original image sequence; Step two, the pixel gray time series of the normalized effective frame is delayed embedded to construct a matrix, the low rank subspace is determined, the initial displacement field is obtained by projecting the coarse displacement into the subspace, the joint cost function is constructed to optimize the displacement field and window parameters, and the optimized displacement field is determined; Step three, the phase space reconstruction is done to the optimized displacement time series to obtain a matrix, the VAE training is built, the loss function optimization is done with reconstruction loss, KL divergence and total correlation, the main mode is screened, and the final result is output.

2. The adaptive regulation mechanism high-precision particle image velocimetry method according to claim 1, characterized in that, Collect particle image, and with key indicators as feedback dynamic adjustment of acquisition parameters, the specific method is: Select a high-speed camera that meets the basic requirements of PIV technology; for fluid, use tracer particles that match the physical properties; if the natural texture imaging of the solid surface meets the standard, it can be directly used, otherwise, random texture can be sprayed or particles can be attached to form a traceable feature, the initial value of exposure time is set as T1, the sampling frequency is set according to the type of flow field; Real-time calculation of adjacent frame cross-correlation peak shape index and inter-frame coarse displacement △u, the cross-correlation peak shape index includes peak height and half-width, and the cross-correlation peak signal-to-noise ratio Sc is calculated; When △u>k×w, increase the sampling frequency or shorten the exposure time to △u≤k×w, wherein w is the length of the cross-correlation window, and k is a preset proportion threshold; When Sc<s1, increase the illumination intensity, improve the camera gain or prolong the exposure time until Sc≥s1, s1 is a preset minimum system signal-to-noise ratio threshold, record the parameter adjustment information to form the acquisition log, and the particle image after parameter adjustment is used as the original image sequence.

3. The adaptive regulation mechanism's high-precision particle image velocimetry method according to claim 1, characterized in that, Pre-processing and screening effective frame of original image sequence, the specific method is: Through flat field correction, background subtraction and distortion correction, the geometric and gray corrected image sequence is obtained; the definition of the corrected image, i.e. MTF(0.5) at Nyquist frequency, is calculated, the background gray variance ab of the corrected image is calculated, and the definition threshold M1 and the background noise threshold b1 are set based on the historical data of conventional PIV and the actual situation; Only the effective frames with definition ≥M1 and ab≤b1 are retained, the effective frames are detrended and normalized, the frame deletion rate is recorded, and the normalized effective frames are output.

4. The adaptive control mechanism's high-precision particle image velocimetry method according to claim 1, wherein, Delay embedding is done to the pixel gray time series of the normalized effective frame to construct a matrix, and the low rank subspace is determined, the specific method is: Let the normalized valid frames be denoted as , t represents different time points, as the index of valid frames, T is the total number of frames, define the pixel coordinates (x, y) of each valid image frame, and the gray time series of each pixel is delayed embedded, and the embedded vectors are spliced to form a delay embedding matrix , where d represents the embedding dimension, determined by the false neighbor rate, and R represents the real set; right Perform SVD decomposition, i.e. =UΣV^T, where U is a left singular matrix, V^T is a right singular matrix, and Σ is a diagonal matrix with diagonal elements for The singular values, where i = 1, 2, ..., min(T-d+1, d), are set with a threshold. Where ω(β) is a threshold coefficient based on random matrix theory, β=K / N, and K is... The number of columns; N is the number of rows, The median and rank of the singular values ​​are represented by the median and rank of the singular values. To meet The number of singular values, and the corresponding low-rank subspace. For the front of U A matrix composed of columns; the subspace is updated in real time using incremental SVD, as shown in the formula. ,in, Let be the subspace at time t. For the subspace of the previous moment, The first step obtained by SVD decomposition at time t The left column is a singular matrix, where ρ is the forgetting factor, set according to historical data and actual needs.

5. The adaptive regulation mechanism's high-precision particle image velocimetry method according to claim 1, characterized in that, The initial displacement field is obtained by projecting the coarse displacement into the subspace, the specific method is: Coarse displacement estimation by basic cross-correlation method The coarse displacement estimation is projected to the low-rank subspace to get the initial low-rank displacement field u(0) by formula , where represents the orthogonal basis matrix of the low-rank subspace, is the coarse displacement estimation result.

6. The adaptive regulation mechanism's high-precision particle image velocimetry method according to claim 1, characterized in that, The joint cost function is constructed to optimize the displacement field and window parameters, and the optimized displacement field is determined, the specific method is: Using the formula A joint cost function is constructed, where u is the displacement field, θ is the affine parameters of the deformable window, including scaling, shearing, rotation, and translation parameters, Jc(u, θ) is the normalized negative peak of cross-correlation, Jo(u), Js(u) are the mean of optical flow re-projection errors, Ji(u) is the mean of the square sum of displacement field gradients, and Jm(u) is the consistency error of the displacement field with the low-rank subspace. , , , , respectively represent the weight coefficients of each cost term, which are set according to historical data and actual situation; The gradient descent method is used to combine and optimize θ and u. In the iteration process, the step size and the number of iterations are adjusted, so that the cost function gradually converges to the minimum value; Combining the multi-scale pyramid strategy, the top layer is based on The coarse displacement field is calculated, and after upsampling, it is used as the initial solution of the lower layer until the original resolution layer. The original resolution layer is corrected to the sub-pixel through cubic spline interpolation, and the fusion result is optimized displacement field .

7. The adaptive regulation mechanism's high-precision particle image velocimetry method according to claim 1, characterized in that, The phase space reconstruction is done to the optimized displacement time series to obtain a matrix, the VAE training is built, the specific method is: Let the optimized displacement field time series be denoted as , with sequence length N, the optimal delay J is determined by the average mutual information minimum value, the optimal dimension D is determined by the false nearest neighbor rate criterion, the embedding vector is constructed based on J and D, and the delay embedding matrix X with dimension (T- (D-1) J) × D is formed by arrangement, the VAE is built, the encoder is a two-layer fully connected network, and the output latent variable mean and the logarithmic variance logσ², the decoder is a symmetric fully connected network, and the reconstructed matrix Xh is output.

8. The adaptive regulation mechanism's high-precision particle image velocimetry method according to claim 1, characterized in that, The loss function optimization is done with reconstruction loss, KL divergence and total correlation, the specific method is: Let the total loss function Lt=Lr+Lk+f×Lc, wherein Lr represents the reconstruction loss, Lk represents the KL divergence, Lc represents the total correlation, f is the total correlation weight, and the setting is based on historical data and actual situation; The difference between the original matrix X and the reconstructed matrix Xh is measured by the mean square error, denoted as reconstruction loss, and the specific formula is where Nx is the total number of elements of X. by extracting the latent variable mean of the encoder output with the log variance logσ2, for each latent variable dimension r, ; A three-layer fully connected discriminator is constructed to distinguish the real latent variable distribution from the independent latent variable distribution in each dimension. The dependence of the real latent variable distribution on the independent distribution is calculated using the trained discriminator, and the degree is the total correlation degree. f is a weight coefficient, which is set according to historical data and actual situation.

Images