A method for denoising and reconstructing space-time graph signals

Through the combination of deep algorithm expansion technology and conjugate gradient method, the denoising and reconstruction objective function of the space-time graph signal is constructed, which solves the problems of high computational complexity and low hyperparameter optimization efficiency in the existing technology, and achieves efficient and accurate space-time Graph signal denoising and reconstruction.

CN117874444BActive Publication Date: 2024-11-22CHINA UNIV OF MINING & TECH
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202410074355.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-18
Publication Date
2024-11-22
Estimated Expiration
2044-01-18

AI Technical Summary

Technical Problem

The existing spatiotemporal and spatial graph signal denoising reconstruction methods have high computational complexity, and optimization of hyperparameters depends on the calculation-intensive grid search method, which may lead to the problem of suboptimal selection.

Method used

Using the depth algorithm expansion technology, the denoising reconstruction objective function is constructed through the spatial domain graph Laplace matrix and the time domain differential matrix, and iteratively solves it using the conjugate gradient method, combining diagonal assisted weighting matrix and hyperparameter optimization to achieve Noise reduction reconstruction of spatiotemporal graph signals.

Benefits of technology

It significantly reduces the calculation overhead, improves the convergence speed and accuracy of the algorithm, effectively recognizes and removes noise, improves the quality and reliability of the signal, and is suitable for large-scale data processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117874444B_ABST
    Figure CN117874444B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of signal denoising and reconstruction, and in particular to a method for denoising and reconstructing a space-time graph signal; collecting and inputting a complete space-time graph signal without noise, and dividing the complete space-time graph signal without noise into a training set and a verification set; respectively performing denoising and reconstruction on the complete space-time graph signal without noise in the training set and the verification set by using the same method, obtaining the optimal model parameters after the training, and verifying the algorithm effect on the verification set data; the method for denoising and reconstructing a space-time graph signal in the present invention performs well in processing noisy and incomplete space-time graph signals, and has significant advantages; the method for denoising and reconstructing a space-time graph signal in the present invention can effectively identify and remove noise, especially in space-time sensor data, and by effectively processing noise, the method helps to improve the quality, accuracy and reliability of the signal.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of seawater temperature signal acquisition, and in particular to a method for denoising and reconstructing space-time graph signals. Background Art

[0002] The spatiotemporal graph structure is a dynamic representation of a network, in which the connections between nodes change over time. Existing research on denoising and reconstruction of spatiotemporal graph signals involves multiple fields, including graph signal processing, spatiotemporal signal processing, and machine learning. In existing research, some spatiotemporal graph signal processing algorithms have high computational complexity, which may limit their use in practical applications, especially in real-time or resource-constrained environments; in some spatiotemporal graph signal processing algorithms, the performance of spatiotemporal graph signal processing is limited by the quality of the input data. If the input spatiotemporal graph signal is affected by noise, distortion, or incompleteness, the accuracy of denoising and reconstruction will be affected; for machine learning methods, the generalization ability of the model on unseen data is a key issue. If the training data is insufficient or not representative enough, the model may perform poorly in practical applications; choosing an appropriate model structure is crucial for spatiotemporal graph signal processing tasks, and the choice of model may be affected by factors such as the particularity of the problem and the limitations of available data; spatiotemporal graph signals usually contain temporal and spatial correlations, and accurately modeling these correlations is a challenge. Existing methods simplify temporal or spatial correlations, resulting in insufficient modeling of real scenes.

[0003] In addition, the sampling and reconstruction problems of graph signals have been widely studied in the field of graph signal processing for many years. So far, the denoising and reconstruction methods of spatiotemporal graph signals can be summarized into two main branches: 1) static methods, which are based on the application of time-vertex Fourier transform on each measurement time dimension; 2) time-varying methods that simultaneously reconstruct the entire time-varying data, and apply convex optimization theory to solve the denoising and reconstruction objective function composed of signal smoothness penalty terms. These methods usually formulate optimization problems that jointly solve spatial and temporal smoothing, usually by iterative solution, using computationally expensive conjugate gradient methods, involving eigenvalue decomposition or matrix inversion of the graph Laplacian matrix of the calculated spatiotemporal graph signal. This computational burden is particularly challenging for graphs with a large number of nodes. In addition, optimizing hyperparameters usually relies on computationally intensive grid search methods, which may face the problem of suboptimal selection when finding the best hyperparameter combination. This is mainly because grid search uniformly samples the hyperparameter space, while the actual hyperparameter performance may have large nonlinearity or irregularity. Therefore, grid search may miss some potential excellent hyperparameter combinations, resulting in the final selected hyperparameter configuration not being globally optimal. A method for denoising and reconstructing spatiotemporal graph signals is proposed to solve the problems existing in the prior art. Summary of the invention

[0004] The purpose of the present invention is to provide a method for denoising and reconstructing a spatiotemporal graph signal, so as to solve the problems in the prior art that the existing spatiotemporal graph signal denoising and reconstruction methods have high computational complexity, and the optimization of hyperparameters usually relies on computationally intensive grid search methods, which may lead to suboptimal selection.

[0005] The technical solution of the present invention is: a method for denoising and reconstructing a space-time graph signal, collecting and inputting a noise-free and complete space-time graph signal, dividing the noise-free and complete space-time graph signal into a training set and a verification set; respectively performing denoising and reconstruction on the noise-free and complete space-time graph signal in the training set and the verification set by using the same method, and the denoising and reconstruction method is:

[0006] S1. Add noise and binary sampling to the noise-free complete space-time graph signal to synthesize the noisy incomplete space-time graph signal;

[0007] S2. Based on the noisy and incomplete spatiotemporal graph signal, the spatial domain adjacency matrix and the spatial domain graph Laplacian matrix are used to characterize the graph structure in the spatial domain, and the denoising and reconstruction objective function is designed;

[0008] S3. The optimization variables of the denoising and reconstruction objective function are iterated and gradients are obtained by using a deep algorithm, and the diagonalizable auxiliary weighted matrix, the time domain difference matrix and the denoising and reconstruction objective function hyperparameters are solved. The diagonalizable auxiliary weighted matrix, the time domain difference matrix and the denoising and reconstruction objective function hyperparameters obtained by the above solutions are substituted into the denoising and reconstruction objective function, and the conjugate gradient method is used until the function value of the denoising and reconstruction objective function converges to obtain the denoised and reconstructed space-time graph signal;

[0009] After training, the optimal model parameters are obtained, and the algorithm effect is verified on the validation set data.

[0010] Preferably, in step S2, based on the noisy and incomplete spatiotemporal graph signal, the graph structure in the spatial domain is characterized by the spatial domain adjacency matrix and the spatial domain graph Laplacian matrix, and the method for designing the denoising and reconstruction objective function is:

[0011] S21. Constructing a spatial domain adjacency matrix by using a threshold Gaussian kernel weighting function;

[0012] S22. Based on the spatial domain adjacency matrix, construct a spatial domain graph Laplacian matrix and a time domain difference matrix, and a diagonalizable auxiliary weighting matrix;

[0013] S23. Based on the noisy and incomplete microseismic signals obtained by modeling, the spatial domain graph Laplace matrix, the time domain difference matrix, and the diagonalizable auxiliary weighted matrix, the space-time graph denoising and reconstruction objective function is constructed.

[0014] Preferably, in step S1, the method of adding noise and binary sampling to the noise-free complete space-time graph signal to synthesize the noisy incomplete space-time graph signal is:

[0015]

[0016] in, represents the Hadamard product, Y is the observed space-time graph signal, X is the complete space-time graph signal without noise, V is the space-time noise, and J∈{0,1} is the sampling matrix.

[0017] Preferably, the spatial domain adjacency matrix is ​​denoted as W, and the expression of the elements in the spatial domain adjacency matrix W is:

[0018]

[0019] where θ represents the scaling factor or kernel bandwidth and dist(i,j) represents the Euclidean distance between sample i and sample j.

[0020] Preferably, the spatial domain graph Laplacian matrix is ​​denoted as L, and the expression of L is:

[0021] L=DW,

[0022] Where D is the degree matrix, and the element expression of D is:

[0023] D ij =∑ i W ij ,

[0024] The time domain difference matrix D L Defined as:

[0025]

[0026] Among them, d1,…,d L ∈R ≥0 Indicates that the filter coefficients satisfy

[0027] Preferably, the expression of the denoising and reconstruction objective function is:

[0028]

[0029] Among them, R is the diagonalizable auxiliary weighting matrix introduced N is the number of nodes in the space domain of the spatiotemporal graph signal; α is a hyperparameter; the quadratic penalty term of the graph Laplacian matrix

[0030] Preferably, the optimization variables in the noise reduction and reconstruction objective function are split into parameters that can be learned through deep learning algorithms, that is, k is the iteration number of the denoising and reconstruction objective function; for each obtained deep algorithm, the optimization parameter Θ(k) is expanded (k=1,...,K), and the denoising and reconstruction objective function is solved by the preset number of iterations of the denoising and reconstruction objective function to obtain the denoising and reconstruction space-time graph signal

[0031] Preferably, a parameterized iterative solution is performed based on a supervised learning deep algorithm to optimize the variable hyperparameter α, the diagonalizable auxiliary weighting matrix R, and the time domain difference matrix D. L ; Set the number of learning rounds E for the deep algorithm to expand, and for each round of update of the deep algorithm, a set of parameterized optimization variables is obtained:

[0032]

[0033] Use the loss function to update the optimization variable. The loss function is defined as:

[0034]

[0035] in, is a complete and noise-free space-time graph signal, N and M are the spatial and temporal dimensions of the signal respectively;

[0036] Using a set of parameters Θ e (k) Solve the above denoising and reconstruction objective function.

[0037] Preferably, the utilization of a set of parameters θ obtained e (k) Solving the above denoising and reconstruction objective function specifically includes the following methods:

[0038] Initial value setting:

[0039] Optimize step size setting:

[0040] Optimize direction update: X k+1 =X k +τΔX k ,

[0041]

[0042]

[0043] Among them, the gradient of the denoising and reconstruction objective function is defined as:

[0044] Through the loss function Further update Θ e (k), until the set number of deep algorithm learning rounds E is reached, and a set of parameters Θ that makes the denoising and reconstruction objective function optimal is obtained E (k) k=1,…,K,and bring it into the denoising and reconstruction objective function to obtain the denoising and reconstruction space-time graph signal

[0045] Preferably, after the training, the optimal model parameters are obtained: the diagonalizable auxiliary weighting matrix R, the time domain difference matrix D L and the denoising and reconstruction objective function hyperparameter α; using the validation set to construct a graph structure including: adjacency matrix W, spatial domain graph Laplacian matrix L, through the denoising and reconstruction objective function, solve the denoising and reconstructed validation set data.

[0046] Compared with the prior art, the advantages of the present invention are:

[0047] (1) A spatiotemporal graph signal denoising and reconstruction method in the present invention uses a graph signal model to explore the covariance relationship between the time and space domains, and uses the spatial domain graph Laplacian matrix L and the time domain difference matrix D L A denoising and reconstruction objective function is constructed, and the deep algorithm expansion technology is used to iteratively solve the optimization variables; it performs well in processing noisy and incomplete spatiotemporal graph signals, and its advantages are significant; a spatiotemporal graph signal denoising and reconstruction method in the present invention can effectively identify and remove noise, especially in spatiotemporal sensor data, and by effectively processing noise, the method helps to improve the quality, accuracy and reliability of the signal.

[0048] (2) The present invention introduces a quadratic penalty term from the diagonalizable auxiliary weight matrix R to the graph Laplacian matrix L in the reconstruction task. The performance is significantly improved; we use deep algorithm unfolding to optimize the proposed denoising reconstruction objective function, providing an efficient and accurate solution to the problem. Through comparative evaluation with state-of-the-art methods on synthetic and real datasets, we demonstrate the excellence of deep algorithm unfolding in signal reconstruction, highlighting its potential for practical applications in various fields. Deep algorithm unfolding is particularly good at handling the denoising reconstruction problem of spatiotemporal graph signals, providing an advanced and effective solution to this problem.

[0049] (3) The present invention constructs a spatiotemporal graph structure of the signal and utilizes deep algorithm expansion technology to simultaneously achieve the purpose of denoising and reconstruction of spatiotemporal distributed signals, fully integrate the spatiotemporal characteristics of the signal, and provide more comprehensive information for the reconstruction process, which helps to more accurately restore the spatiotemporal distribution and pattern of the signal.

[0050] (4) The present invention adopts deep algorithm expansion technology, which has significant advantages in solving alternative optimization problems; by reducing the number of loop iterations, deep algorithm expansion can effectively reduce computational overhead and improve the convergence speed of the algorithm. Its feature of improving instruction-level parallelism helps to accelerate the processor to execute multiple instructions, especially in optimization problems of large-scale data processing. In addition, algorithm expansion optimizes loop structure, improves cache utilization, and reduces data access latency, which is particularly critical for processing optimization problems of large data sets. The effect of reducing branch costs makes the execution efficiency of complex optimization algorithms higher. Overall, deep algorithm expansion technology provides a powerful performance optimization method for solving alternative optimization problems through its many advantages, and is particularly suitable for large-scale and complex problem solving scenarios.

[0051] (5) The present invention is based on the deep algorithm expansion, and the parameterized diagonalizable auxiliary weighting matrix R and the time domain difference matrix D are updated in an alternating iterative manner. L and hyperparameter α, and use a closed-form solution to achieve denoising and reconstruction of the spatiotemporal graph signal; the core of this process is to expand through a deep algorithm and continuously iterate and optimize the diagonalizable auxiliary weighting matrix R and the time domain difference matrix D L and hyperparameter α, and uses the conjugate gradient method to solve the problem, ensuring the accuracy and efficiency of spatiotemporal graph signal processing. The closed-form solution technology ensures the accuracy of noise reduction and reconstruction on a global scale; this alternating iterative update method ensures higher accuracy and efficiency when processing spatiotemporal graph signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0053] Figure 1 The root mean square error (the smaller the error, the better) of the reconstruction of the spatiotemporal graph signal by the present invention under different sampling rates (the proportion of the number 0 in the sampling matrix J) on the synthetic data set in comparison with other methods;

[0054] Figure 2 The root mean square error of the denoising of the spatiotemporal graph signal by the present invention under different V spatiotemporal noise matrix variances on the synthetic data set compared with other methods (the smaller the error, the better);

[0055] Figure 3 The root mean square error (the smaller the error, the better) of the reconstruction of the spatiotemporal graph signal by the present invention at different sampling rates (the proportion of the number 0 in the sampling matrix J) on the seawater temperature data set in comparison with other methods;

[0056] Figure 4For the seawater temperature data set, compared with other methods, the root mean square error of the present invention for denoising the space-time graph signal under different V space-time noise matrix variances (the smaller the error, the better). DETAILED DESCRIPTION

[0057] The present invention is further described in detail below in conjunction with specific embodiments:

[0058] like Figure 1 As shown, a method for denoising and reconstructing a space-time graph signal is provided. A noise-free and complete space-time graph signal is collected and input, and the noise-free and complete space-time graph signal is divided into a training set and a validation set. In this embodiment, the training set is 70% and the validation set is 30%. In the data collection process, multiple sensors distributed in space independently record time series to form a graph structure, and then form a space-time graph. The noise-free and complete space-time graph signals in the training set and the validation set are respectively subjected to denoising and reconstruction by the same method. The denoising and reconstruction method is:

[0059] 1. De-noise and reconstruct the noise-free and complete spatiotemporal graph signals in 70% of the training set:

[0060] S1. Add noise and binary sampling to the noise-free and complete space-time graph signal in the training set to synthesize the noisy and incomplete space-time graph signal;

[0061] In step S1, the method of adding noise and binary sampling to the noise-free complete space-time graph signal to synthesize the noisy incomplete space-time graph signal is as follows:

[0062]

[0063] in, represents the Hadamard product, Y is the observed space-time graph signal, X is the complete space-time graph signal without noise, V is the space-time noise, and J∈{0,1} is the sampling matrix.

[0064] S2. Based on the noisy and incomplete spatiotemporal graph signal, the graph structure in the spatial domain is constructed and the denoising and reconstruction objective function is designed;

[0065] In step S2, based on the noisy and incomplete spatiotemporal graph signal, the graph structure of the spatial domain is characterized by the spatial domain adjacency matrix and the spatial domain graph Laplacian matrix, and the method of designing the denoising and reconstruction objective function is as follows: S21. Construct a spatial domain adjacency matrix by a threshold Gaussian kernel weighted function; the spatial domain adjacency matrix is ​​denoted as W, and the expression of the elements in the spatial domain adjacency matrix W is:

[0066]

[0067] Where θ represents the scale factor or kernel bandwidth, and dist(i,j) represents the Euclidean distance between sample i and sample j. The spatial domain adjacency matrix W realizes the spatial correlation analysis between the data signals collected by the sensor, and then describes the spatial correlation between the data signals collected by the sensor through the spatial domain adjacency matrix W, which improves the accuracy of the mathematical description of the spatial dimension correlation of the incomplete signal obtained under noise conditions, and ensures the accuracy of the subsequent denoising and reconstruction of the spatiotemporal graph signal. Based on the amplitude of the time series data collected by the sensors with arbitrary spatial distribution, the spatial domain adjacency matrix W of the spatial dimension is constructed to measure the spatial distance of the sensor sampling signal. The spatial dimension implicit graph structure is used to fuse the information of adjacent or non-adjacent sensor sampling points in the spatial dimension to obtain highly robust features, improve the accuracy of the mathematical description of the correlation between the spatiotemporal signal sampling points obtained under noise conditions, and ensure the accuracy of the subsequent denoising and reconstruction of the spatiotemporal graph signal.

[0068] S22. Based on the spatial domain adjacency matrix, construct a spatial domain graph Laplacian matrix and a time domain difference matrix, and a diagonalizable auxiliary weighting matrix;

[0069] The spatial domain graph Laplacian matrix is ​​denoted as L, and the expression of L is:

[0070] L=DW,

[0071] Where D is the degree matrix, and the element expression of D is:

[0072] D ij =∑ i W ij ,

[0073] The time domain difference matrix D L Defined as:

[0074]

[0075] Among them, d1,…,d L ∈R ≥0 Indicates that the filter coefficients satisfy

[0076] The present invention uses the time domain difference matrix D L To explore the relationship between data in the time dimension, it can accurately capture the evolution and association of data in the time series; this method helps to analyze the dynamic changes of data at different time points, detect trends and seasonality; by analyzing the data in the time dimension, it can better understand the time law and dynamic association of data. Time domain difference matrix D LBy filtering the time series vibration waveforms collected by spatially distributed sensors, the original noise-free time series waveform can be retained as much as possible while reducing noise. By utilizing the topological structure and connection relationship of the graph, the correlation between nodes is considered during the filtering process, making the filtering more in line with the intrinsic structural characteristics of the time series data, and being able to adjust the weight of the filter according to the local signal characteristics to adapt it to different signal changes, achieving more accurate filtering of noise, expanding the search range of similar points in the time dimension, and avoiding the traditional filtering method's excessive reliance on the strength of adjacent signals in the time dimension, so that the time series signal waveform itself can be retained to the greatest extent, greatly improving the accuracy of the time series waveform denoising reconstruction. The technical effect of this implementation will combine graph structure information and adaptive filtering to comprehensively process the time series waveform, and have the ability to intelligently select the best method, making the processing more efficient and accurate.

[0077] S23. Based on the noisy and incomplete microseismic signal obtained by modeling, the spatial domain graph Laplace matrix, the time domain difference matrix, and the diagonalizable auxiliary weighting matrix, a space-time graph denoising and reconstruction objective function is constructed; the expression of the denoising and reconstruction objective function is:

[0078]

[0079] Among them, R is the diagonalizable auxiliary weighting matrix introduced N is the number of nodes in the space domain of the spatiotemporal graph signal; α is a hyperparameter; the quadratic penalty term of the graph Laplacian matrix

[0080] The present invention uses a diagonalizable auxiliary weighting matrix R to further strengthen the correlation between data points and re-weight the data in the graph structure, which helps to more accurately capture the association and mutual influence between data points, and more easily identify and highlight the importance of data points, so as to better understand the structure and characteristics of the data set, improve the efficiency of data analysis and mining algorithms, and provide more powerful support for data processing. The purpose of introducing the diagonalizable auxiliary weighting matrix R when establishing the denoising reconstruction objective function of the spatiotemporal graph signal denoising reconstruction is to realize the data time dimension sample point weighting technology; in the process of spatiotemporal graph signal denoising reconstruction, the denoising reconstruction objective function hopes to accurately reconstruct the original signal and retain as much useful information as possible. In order to achieve this goal, introducing the diagonalizable auxiliary weighting matrix R is an effective method. The diagonalizable auxiliary weighting matrix R is a diagonal matrix, and its diagonal elements represent the weight of each sample point in the time dimension. By adjusting these weights, the purpose of controlling the contribution of different sample points in the reconstruction process can be achieved; specifically, a larger weight means that the sample point has a higher importance in the reconstruction, while a smaller weight means that the sample point has a lower importance in the reconstruction. By introducing the diagonalized auxiliary weighting matrix R, we can multiply each sample point in the denoising and reconstruction objective function by its corresponding weight, thereby realizing the data time dimension sample point weighting technology. The advantage of this is that the importance of each sample point can be adjusted according to the actual situation, thereby better adapting to different denoising and reconstruction tasks.

[0081] The quadratic penalty term based on the graph Laplacian matrix is ​​defined without additional constraints. Without introducing additional parameters, the optimization variables are solved through a combination of neural networks to expand the hyperparameter search range, thereby greatly improving the solution accuracy.

[0082] S3. The optimization variables of the denoising and reconstruction objective function are iteratively expanded using a deep algorithm to obtain gradients and solve the diagonalizable auxiliary weighting matrix, the time domain difference matrix and the denoising and reconstruction objective function hyperparameters. The diagonalizable auxiliary weighting matrix, the time domain difference matrix and the denoising and reconstruction objective function hyperparameters obtained based on the above solutions are substituted into the denoising and reconstruction objective function, and the denoised and reconstructed spatiotemporal image signal is obtained by the conjugate gradient method until the function value of the denoising and reconstruction objective function converges; the deep algorithm expansion represents a deep learning technology that aims to learn the parameterized diagonalizable auxiliary weighting matrix R and the time domain difference matrix D L , and adjust the hyperparameter α to improve the effect of denoising and reconstruction of the spatiotemporal graph signal. The key features of the present invention include constructing a minimization optimization problem, and the deep algorithm expansion method learns the parameters of the model by expressing the optimization problem in a differentiable form and solving it by gradient descent. The advantages of the deep algorithm expansion in the present invention are as follows: 1) parameterizing the diagonalizable auxiliary weighting matrix R and the time domain difference matrix D L:The deep algorithm expansion method introduces a parameterized diagonal auxiliary weighting matrix R and a time domain difference matrix D L , allowing the model to learn more adaptive signal processing capabilities; 2) Self-iteration and optimization process: The deep algorithm expansion method expands the optimization problem into a step-by-step optimization process, allowing the model to automatically adjust parameters in each iteration and generate new inputs using the output of the current model. This self-iteration and optimization process can help the model gradually optimize and learn better representation and reconstruction capabilities. Specifically, the deep algorithm expansion method first constructs a loss function, which includes the difference between the original spatiotemporal graph signal and the denoised reconstructed spatiotemporal graph signal, the regularization term, and the penalty term of the hyperparameter, and then iteratively updates the parameters of the model to minimize the loss function and continuously improve the quality of the denoised reconstructed spatiotemporal graph signal. The conjugate gradient method processes the denoised reconstructed spatiotemporal graph signal through convex optimization theory. It uses the convex optimization principle to quickly and accurately find the global optimal solution, avoiding the problem of falling into the local optimal solution, thereby improving the processing efficiency and the reliability of the results. The conjugate gradient method combines convex optimization theory and signal processing technology, greatly improving the processing efficiency and ensuring the accuracy and reliability of the denoised reconstructed spatiotemporal graph signal.

[0083] The optimization variables in the denoising and reconstruction objective function are split into parameters that can be learned through deep learning algorithms, namely k is the iteration round of the denoising and reconstruction objective function; for each obtained depth algorithm, the optimization parameter Θ is expanded. (k) , (k=1,...,K), solve the denoising and reconstruction objective function through the preset number of iterations of the denoising and reconstruction objective function to obtain the denoising and reconstruction space-time graph signal

[0084] Furthermore, a parametric iteration is performed based on a supervised learning deep algorithm to solve the optimization variable hyperparameter α, the diagonalizable auxiliary weight matrix R, and the time domain difference matrix D L ; Set the number of learning rounds E for the deep algorithm to expand, and for each round of update of the deep algorithm, a set of parameterized optimization variables is obtained:

[0085]

[0086] Use the loss function to update the optimization variable. The loss function is defined as:

[0087]

[0088] in, is a complete and noise-free space-time graph signal, N and M are the spatial and temporal dimensions of the signal respectively;

[0089] The use of a set of parameters θ obtained e(k) Solving the above denoising and reconstruction objective function specifically includes the following methods:

[0090] Initial value setting:

[0091] Optimize step size setting:

[0092] Optimize direction update: X k+1 =X k +τΔX k ,

[0093]

[0094]

[0095] Among them, the gradient of the denoising and reconstruction objective function is defined as:

[0096] Through the loss function Further update Θ e (k) , until the set number of deep algorithm learning rounds E is reached, and a set of parameters Θ that makes the denoising and reconstruction objective function optimal is obtained E (k) k=1,…,K,and bring it into the denoising and reconstruction objective function to obtain the denoising and reconstruction space-time graph signal

[0097] In short, the optimization parameter Θ is obtained by expanding the deep algorithm, and then Θ is obtained by each update based on the deep algorithm. The denoising and reconstruction objective function is solved by the preset number of iterations of the denoising and reconstruction objective function to obtain the denoising and reconstruction space-time graph signal Finally, through K-times iterative deep algorithm expansion, the optimal denoising and reconstruction of the spatiotemporal graph signal is learned.

[0098] 2. Perform denoising and reconstruction on the noise-free and complete space-time graph signals in 30% of the validation set. The method for denoising and reconstruction on the noise-free and complete space-time graph signals in the validation set is the same as the method for denoising and reconstruction on the noise-free and complete space-time graph signals in the training set, and will not be repeated here.

[0099] After the training, the optimal model parameters are obtained: the diagonalizable auxiliary weighting matrix R, the time domain difference matrix D L and the denoising and reconstruction objective function hyperparameter α; using the validation set to construct a graph structure including: adjacency matrix W, spatial domain graph Laplacian matrix L, through the denoising and reconstruction objective function, solve the denoising and reconstructed validation set data.

[0100] Down Figure 1In the figure, the horizontal axis represents the sampling rate (the proportion of the number 0 in the sampling matrix J), and the vertical axis represents the root mean square error; Figure 2 In the figure, the horizontal axis represents the variance of the V space-time noise matrix, and the vertical axis represents the root mean square error; Figure 3 In the figure, the horizontal axis represents the sampling rate (the proportion of the number 0 in the sampling matrix J), and the vertical axis represents the root mean square error; Figure 4 In the figure, the horizontal axis represents the variance of the spatiotemporal noise matrix V, and the vertical axis represents the root mean square error. By comparing with the existing research methods, it can be seen that the method proposed in the present invention has a higher accuracy in denoising and reconstructing spatiotemporal image signals.

[0101] In the present invention Figure 1 and Figure 2 The synthetic dataset generation process in is as follows: First, 100 points are randomly selected in a square area to create a nearest neighbor spatial domain graph structure, where each node is connected to its five nearest neighbor nodes by edges in the graph structure, and the weight of the edge is inversely proportional to the square of the Euclidean distance. Specific synthetic data time domain generation method: The initial signal, denoted as x1, is a low-frequency graph structure with an energy level of 104; then, at each time step t≥2, an independent and identically distributed (iid) white Gaussian signal ft is generated and rescaled to achieve a 2-norm of ∈ (in this study, ∈ is set to 1); then the signal xt is updated using the equation xt=xt-1+L-1 / 2ft, where L is the spatial domain graph Laplacian matrix of the spatial domain graph structure.

[0102] In addition, we use a monthly sea temperature dataset obtained from the NOAA Physical Sciences Laboratory website. Specifically, we focus on a subset of 100 selected points located in the Pacific Ocean over a period of 600 months; the synthetic dataset and the sea temperature dataset are divided into 70% of the data as a training set and 30% as a validation set, respectively, and the research results are obtained by averaging 20 random assignments.

[0103] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with this technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the scope of protection of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is limited by the attached claims rather than the above description, and it is intended to include all changes within the meaning and scope of the equivalent elements of the claims in the present invention.

Claims

1. A method for denoising and reconstructing a spatiotemporal image signal, characterized in that: Collect and input a noise-free and complete space-time graph signal, and divide the noise-free and complete space-time graph signal into a training set and a validation set; respectively, perform noise reduction and reconstruction on the noise-free and complete space-time graph signal in the training set and the validation set by the same method, and the noise reduction and reconstruction method is: S1. Add noise and binary sampling to the noise-free complete space-time graph signal to synthesize the noisy incomplete space-time graph signal; the method for synthesizing the noisy incomplete space-time graph signal is: in, represents the Hadamard product, Y is the observed space-time graph signal, X is the complete space-time graph signal without noise, V is the space-time noise, and J∈{0,1} is the sampling matrix; S2. Based on the noisy and incomplete spatiotemporal graph signal, the spatial domain graph structure is characterized by the spatial domain adjacency matrix and the spatial domain graph Laplacian matrix, and a denoising and reconstruction objective function is designed; the expression of the denoising and reconstruction objective function is: Among them, R is the diagonalizable auxiliary weighting matrix introduced N is the number of nodes in the space domain of the spatiotemporal graph signal; α is a hyperparameter; the quadratic penalty term of the graph Laplacian matrix is the space-time image signal after noise reduction and reconstruction; D L is the time domain difference matrix; The optimization variables in the denoising and reconstruction objective function are split into parameters that can be learned through deep learning algorithms, namely k is the iteration round of the denoising and reconstruction objective function; for each obtained depth algorithm, the optimization parameter Θ is expanded. (k) ,k=1,...,K, solve the denoising and reconstruction objective function through the preset number of iterations of the denoising and reconstruction objective function to obtain the denoising and reconstruction space-time graph signal S3. Use the deep algorithm to iterate and find the gradient and solve the diagonalizable auxiliary weight matrix R and the time domain difference matrix D L And the denoising reconstruction objective function hyperparameter α, based on the above solution, the diagonalizable auxiliary weighting matrix R and the time domain difference matrix D L Substitute the denoising reconstruction objective function hyperparameter α into the denoising reconstruction objective function, and use the conjugate gradient method until the function value of the denoising reconstruction objective function converges to obtain the denoised and reconstructed space-time graph signal and the optimized variables after denoising and reconstruction; After training, the optimal model parameters are obtained, and the algorithm effect is verified on the validation set data.

2. The method for denoising and reconstructing a spatiotemporal image signal according to claim 1, characterized in that: In step S2, based on the noisy and incomplete spatiotemporal graph signal, the spatial domain graph structure is characterized by the spatial domain adjacency matrix and the spatial domain graph Laplacian matrix, and the method of designing the denoising and reconstruction objective function is as follows: S21. Constructing a spatial domain adjacency matrix by using a threshold Gaussian kernel weighting function; S22. Based on the spatial domain adjacency matrix, construct a spatial domain graph Laplacian matrix and a time domain difference matrix, and a diagonalizable auxiliary weighting matrix; S23. Based on the noisy and incomplete microseismic signals obtained by modeling, the spatial domain graph Laplace matrix, the time domain difference matrix, and the diagonalizable auxiliary weighted matrix, the space-time graph denoising and reconstruction objective function is constructed.

3. The method for denoising and reconstructing a spatiotemporal image signal according to claim 2, characterized in that: The spatial domain adjacency matrix is ​​recorded as W, and the expression of the elements in the spatial domain adjacency matrix W is: where θ represents the scaling factor or kernel bandwidth and dist(i,j) represents the Euclidean distance between sample i and sample j.

4. A method for denoising and reconstructing a spatiotemporal image signal according to claim 3, characterized in that: The spatial domain graph Laplacian matrix is ​​denoted as L, and the expression of L is: L=DW, Where D is the degree matrix, and the element expression of D is: D ij =∑ i W ij , The time domain difference matrix D L Defined as: Among them, d1,…,d L ∈R ≥0 Indicates that the filter coefficients satisfy 5. The method for denoising and reconstructing a spatiotemporal image signal according to claim 4, characterized in that: Through the parametric iteration based on supervised learning deep algorithm, the optimization variable hyperparameter α, the diagonalizable auxiliary weight matrix R, and the time domain difference matrix D are solved. L ; Set the number of learning rounds E for the deep algorithm to expand, and for each round of update of the deep algorithm, a set of parameterized optimization variables is obtained: Use the loss function to update the optimization variable. The loss function is defined as: in, is a complete and noise-free space-time graph signal, N and M are the spatial and temporal dimensions of the signal respectively; X (k) Refers to the denoised reconstructed space-time graph signal obtained at the Kth iteration; Using a set of parameters Θ e (k) Solve the above denoising and reconstruction objective function.

6. The method for denoising and reconstructing a spatiotemporal image signal according to claim 5, characterized in that: The use of a set of parameters θ obtained e (k) Solving the above denoising and reconstruction objective function specifically includes the following methods: Initial value setting: Optimize step size setting: Optimize direction update: X k+1 =X k +τΔX k , Among them, the gradient of the denoising and reconstruction objective function is defined as: Through the loss function Further update Θ e (k) , until the set number of deep algorithm learning rounds E is reached, and a set of parameters Θ that makes the denoising and reconstruction objective function optimal is obtained E (k) k=1,…,K,and bring it into the denoising and reconstruction objective function to obtain the denoising and reconstruction space-time graph signal 7. The method for denoising and reconstructing a spatiotemporal image signal according to claim 6, characterized in that: After the training, the optimal model parameters are obtained: the diagonalizable auxiliary weighting matrix R, the time domain difference matrix D L and the denoising reconstruction objective function hyperparameter α; The graph structure constructed using the validation set includes: adjacency matrix W, spatial domain graph Laplacian matrix L, and the objective function is reconstructed through denoising to solve the validation set data after denoising and reconstruction.

Citation Information

Patent Citations

  • Time sequence signal efficient denoising and high-precision reconstruction modeling method and system

    CN111046824A

  • Graph signal distributed online reconstruction method based on truncated Taylor series approximation

    CN111242867A

  • Image clustering method of multi-view intrinsic low-rank structure

    CN112148911A

  • Nuclear power circulating water pump planetary gearbox detection method based on space-time diagram

    CN115541232A

  • Graph convolutional neural network training method and device, electronic equipment and storage medium

    CN117035025A