A sensor network graph structure learning method
Through the stationary graph signal model and ADMM optimization algorithm, the problem of insufficient graph signals in the sensor network is solved, and the graph structure is accurately learned in the observation data, which improves the representation ability and accuracy of the graph structure.
Patent Information
- Application Number
- CN202210191628.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-28
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2042-02-28
AI Technical Summary
In sensor networks, when the number of graph signals is insufficient, it is difficult to accurately infer the graph structure, resulting in poor graph learning effect, especially when the number of nodes is less than ten times.
The stationary graph signal model is adopted, and the observed data is converted into a stationary signal through factor analysis and graph filters. The ADMM method and alternating optimization algorithm are used to express the problem of signal denoising as smooth regularization, and the parameters are adjusted to control the sparseness of the graph structure. Ensure the accuracy of the learned graph structure.
By introducing global smoothness and sparseness constraints, the graph structure information of the sensor network can be accurately learned and inferred from the observation data, improving the representation ability and accuracy of the graph structure.
Smart Images

Figure CN114564994B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of graph signal processing, and in particular to a sensor network graph structure learning method. Background Art
[0002] A sensor network is a computer network consisting of many spatially distributed automatic devices that use sensors to collaboratively monitor physical or environmental conditions at different locations. Graphs, also known as networks, can be extracted from a large number of entities in real relationships. Some common graphs are used to describe different relationships. In addition to sensor networks, it can also be applied to biological networks, transportation networks, social networks, citation networks, etc.
[0003] The task of modern data processing is usually to process structured data, and graphs are used as mathematical tools to describe such data structures. They provide a flexible way to represent the relationship between data entities. Such data is called graph signal, and the signal value of the graph signal is defined on the vertex set of the weighted undirected graph. Graph learning, in general, refers to machine learning on graphs. The method of graph learning is to map the attributes of the graph to the space of feature vectors with the same dimension. The goal of graph learning is to extract the desired features of the graph. Therefore, the representation of the graph can be easily used for downstream tasks such as node classification and link prediction without the need for an explicit embedding process. Therefore, graph learning is a more powerful and meaningful graph analysis technique.
[0004] In graph learning, graphs can capture the actual geometry of structured data, which is essential for efficient processing, analysis, and visualization; learning the relationships between data entities is beneficial to many application areas, such as understanding the functional connections between brain regions or the behavioral influence between a group of people; the inferred graph helps predict future data evolution. The construction of meaningful graph topology plays a vital role in the effective representation, processing, analysis, and visualization of structured data. When a natural choice of graph cannot be obtained from the dataset, the graph topology needs to be inferred or learned from the data. Summary of the invention
[0005] The present invention proposes a sensor network graph structure learning method, which fits the observed sensor network data through a stable and smooth graph signal, so as to accurately infer the sensor graph structure.
[0006] A sensor network graph structure learning method, the method steps are as follows:
[0007] Step 1: Convert the observation data of M groups of N nodes in the sensor network into a data matrix Y∈R N×M ;
[0008] Step 2: Based on the factor analysis model, a stationary graph signal representation model is established;
[0009] Step 3: Use the established stationary signal model as input and pass it through a graph filter;
[0010] Step 4: According to the data matrix Y and the stationary graph signal representation model, the signal denoising problem is expressed as a smooth signal regularization problem according to the graph signal learning framework, so that the signal has global smoothness;
[0011] Step 5: Use the ADMM method to solve the constrained optimization problem, and use the idea of alternating optimization to solve the filter coefficients, graph structure and input stable signal;
[0012] Step 6: Adjust the parameters to control the sparsity of the learned graph structure Laplacian matrix and delete the edges with small weight values to learn the accurate graph structure.
[0013] Furthermore, the expression of the input stationary graph signal in step 2 is:
[0014] X=HC+u x
[0015] Where C is a stationary signal, defined as the covariance matrix ∑ c =E[cc T ]=I zero mean image signal; H is FIR image filter, u x is the mean of the signal X.
[0016] Furthermore, in step 3, the expression of the FIR graph filter is:
[0017]
[0018] where h k is the coefficient of the graph filter, the matrix S is the graph shift operator, α k is the filter coefficient, and k is the filter order.
[0019] Furthermore, in step 4, the stationary graph signal learning problem is expressed as a stationary and smooth signal regularization sparsity problem, and the input signal is operated to learn the graph structure, and the steps are:
[0020] Step 4-1: In the context of graph learning, the graph shift operator S is used as the Laplacian matrix L to achieve the diffusion of stationary signals;
[0021]
[0022] Step 4-2, perform least square estimation fitting on the input signal X passing through the FIR filter and the observation data matrix Y to remove noise;
[0023]
[0024] Step 4-3: Apply smoothness and sparsity constraints to the graph learning problem to control the learned graph Laplacian matrix;
[0025] Among them, α and β are the parameters of adjusting the learned Laplace matrix, which control the scale of smoothness and sparsity; tr is the abbreviation of trace, which means finding the trace of the matrix.
[0026] Step 4-4: Considering the non-convex nature of the solution process, the idea of alternating optimization is adopted to respectively k ,C solve;
[0027]
[0028] This is a convex problem, and the closed-form solution is calculated:
[0029] C * =(H T H+αH T LH) -1 H T Y
[0030] To α k Find the gradient and use the gradient descent algorithm. After multiple iterations or when the error between two iterations is within 10 -4 The following stops;
[0031]
[0032] To α k The obtained gradient is as follows:
[0033] Use the gradient to get the convergence result to update the filter coefficient α k .
[0034] Furthermore, in step 5, the ADMM algorithm is used to solve the optimization problem, and the steps are as follows:
[0035] Step 5-1, define the constraint set L about the Laplace matrix L:
[0036] trace(L)=N
[0037] L*1=0
[0038] L i,j =L j,i ≤0
[0039] Introducing auxiliary variable Z∈L , then the constraint condition satisfies LZ=0;
[0040] Step 5-2: Write the original problem as an augmented Lagrangian function to reduce the number of iterations:
[0041]
[0042] Where P is the Lagrange multiplier, ρ is a parameter that controls the scale of regularization, and Z is an auxiliary variable.
[0043] Step 5-3, calculate the gradient of the Laplacian matrix L:
[0044]
[0045] Update L using the gradient:
[0046]
[0047] Where γ is the step size, k' is the introduced variable which is equal to the filter order, and r is the parameter introduced for partial derivative;
[0048] Step 5-4: Update the auxiliary variable Z according to the converged Laplace matrix L:
[0049]
[0050] Where Π represents the Euclidean projection, which projects L onto the set L.
[0051] Step 5-5, combine Z and the learned L to get the new Lagrange multiplier P:
[0052] P i+1 =P i +ρ(Z i+1 -L i+1 )
[0053] Furthermore, in step 6, by adjusting the coefficients, a graph structure similar in sparsity to the actual graph structure is obtained.
[0054] The beneficial effects achieved by the present invention are: solving the problem of sensor network graph learning when the number of graph signals is small (when the number of signals is less than ten times the number of nodes), adopting a method of introducing global smoothness to the input stationary signal, thereby learning and inferring more accurate graph structure information in the observed data. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 The present invention is a flowchart of a sensor network graph structure learning method according to an embodiment of the present invention. DETAILED DESCRIPTION
[0056] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.
[0057] The present invention proposes an analysis method for fitting observed sensor network data with a stable and smooth graph signal to infer the Laplace matrix of the sensor graph structure. The steps of the method are as follows:
[0058] Step 1: Convert the observation data of M groups of N nodes in the sensor network into a data matrix Y∈R N×M .
[0059] Step 2: Based on the traditional factor analysis model, a new stationary graph signal representation model is proposed.
[0060] The expression of the input stationary graph signal is:
[0061] X=HC+u x
[0062] where C is a stationary signal. A random process on a graph is called (generally) stationary if its first moment is constant over the vertex set and its covariance matrix is invariant with respect to the localization operator. Usually, C is assumed to have a covariance matrix of ∑ c =E[cc T ]=I zero mean image signal. Where H is the FIR image filter, u x is the mean of signal X. Without loss of generality, we can assume that u x is 0.
[0063] Step 3: Use the established stationary signal as input and pass it through a graph filter to fit our observed data.
[0064] The expression of general FIR graph filter is:
[0065]
[0066] where h k are the coefficients of the graph filter, and the matrix S is the graph shift operator. S can be replaced by the Laplacian matrix L or the graph adjacency matrix W. k is the filter order.
[0067] Step 4: According to the graph signal learning framework, the signal denoising problem is formulated as a smooth signal regularization problem, so that the signal has global smoothness.
[0068] The problem of learning stationary graph signals is expressed as a regularized sparse problem of stationary and smooth signals. The input signal is operated and the graph structure is learned. The steps are as follows:
[0069] Step 4-1: In the context of graph learning, the graph shift operator S is used as the Laplacian matrix L to achieve the diffusion of stationary signals;
[0070]
[0071] Step 4-2, perform least square estimation fitting on the input signal X passing through the FIR filter and the observation data matrix Y to remove noise;
[0072]
[0073] Step 4-3: Apply smoothness and sparsity constraints to the graph learning problem to control the learned graph Laplacian matrix;
[0074]
[0075] Step 4-4: Considering the non-convex nature of the solution process, the idea of alternating optimization is adopted to respectively k , C solves:
[0076]
[0077] This is a convex problem, and the closed-form solution is calculated:
[0078] C * =(H T H+αH T LH) -1 H T Y
[0079] To α k Find the gradient and use the gradient descent algorithm. After multiple iterations or when the error between two iterations is within 10 -4 The following stops;
[0080]
[0081] To α k The obtained gradient is as follows:
[0082]
[0083] The gradient can be used to obtain the converged result in order to update the filter coefficient α k .
[0084] Step 5: Use the ADMM method to solve the constrained optimization problem, and use the idea of alternating optimization to solve the filter coefficients, graph structure and input stable signal.
[0085] The ADMM algorithm is used to solve the optimization problem. The steps are as follows:
[0086] Step 5-1: Set the constraint condition set L about the Laplace matrix L:
[0087] trace(L)=N
[0088] L*1=0
[0089] L i,j =L j,i ≤0
[0090] Introducing auxiliary variable Z ∈L , then the constraint condition satisfies LZ=0.
[0091] Step 5-2: Write the original problem as an augmented Lagrangian function to reduce the number of iterations:
[0092]
[0093] Step 5-3, calculate the gradient of the Laplacian matrix L:
[0094]
[0095] Update L using the gradient:
[0096]
[0097] Where γ is the step size.
[0098] Step 5-4: Update the auxiliary variable Z according to the converged Laplace matrix L:
[0099]
[0100] Where Π represents the Euclidean projection, which projects L onto the set L.
[0101] Step 5-5, combine Z and the learned L to get the new Lagrange multiplier P:
[0102] P i+1 =P i +ρ(Z i+1 -L i+1 )
[0103] Step 6: Adjust the parameters to control the sparsity of the learned graph structure Laplacian matrix, delete the edges with small weight values, and learn a more accurate graph structure. By adjusting the coefficients, a graph structure with similar sparsity to the real graph structure is obtained, thus learning a graph structure with better performance.
[0104] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.
Claims
1. A sensor network graph structure learning method, characterized by: The steps of the method are as follows: Step 1: Convert the observation data of M groups of N nodes in the sensor network into a data matrix Y∈R N×M ; Step 2: Based on the factor analysis model, a stationary graph signal representation model is established; Step 3: Use the established stationary signal model as input and pass it through a graph filter; Step 4: According to the data matrix Y and the stationary graph signal representation model, the signal denoising problem is expressed as a smooth signal regularization problem according to the graph signal learning framework, so that the signal has global smoothness; In step 4, the stationary graph signal learning problem is expressed as a stationary and smooth signal regularization sparsity problem. The input signal is operated to learn the graph structure. The steps are: Step 4-1: In the context of graph learning, the graph shift operator S is used as the Laplacian matrix L to achieve the diffusion of stationary signals; Step 4-2, perform least square estimation fitting on the input signal X passing through the FIR filter and the observation data matrix Y to remove noise; Step 4-3: Apply smoothness and sparsity constraints to the graph learning problem to control the learned graph Laplacian matrix; Where α and β are the parameters that adjust the learned Laplacian matrix, controlling the scale of smoothness and sparsity; tr is the abbreviation of trace, which means finding the trace of the matrix; Step 4-4: Considering the non-convex nature of the solution process, the idea of alternating optimization is adopted to respectively k , C is solved; This is a convex problem, and the closed-form solution is calculated: C * =(H T H+αH T LH) -1 H T THE To α k Find the gradient and use the gradient descent algorithm. After multiple iterations or when the error between two iterations is within 10 -4 The following stops; To α k The obtained gradient is as follows: Use the gradient to get the convergence result to update the filter coefficient α k ; Step 5: Use the ADMM method to solve the constrained optimization problem, and use the idea of alternating optimization to solve the filter coefficients, graph structure and input stable signal; Step 6: Adjust the parameters to control the sparsity of the learned graph structure Laplacian matrix and delete the edges with small weight values to learn the accurate graph structure.
2. A sensor network graph structure learning method according to claim 1, characterized in that: The expression of the input stationary graph signal is: X=HC+u x Where C is a stationary signal, defined as the covariance matrix ∑ c =E[cc T ]=I zero mean image signal; H is FIR image filter, u x is the mean of the signal X.
3. A sensor network graph structure learning method according to claim 1, characterized in that: In step 3, the expression of the FIR graph filter is: where h k is the coefficient of the graph filter, the matrix S is the graph shift operator, α k is the filter coefficient, and k is the filter order.
4. A sensor network graph structure learning method according to claim 1, characterized in that: In step 5, the ADMM algorithm is used to solve the optimization problem, and the steps are as follows: Step 5-1, define the constraint set L about the Laplace matrix L: trace(L)=N L*1=0 L i,j =L j,i ≤0 Introducing auxiliary variables Z∈L, the constraint condition satisfies LZ=0; Step 5-2: Write the original problem as an augmented Lagrangian function to reduce the number of iterations: Where P is the Lagrange multiplier, ρ is a parameter that controls the scale of regularization, and Z is an auxiliary variable; Step 5-3, calculate the gradient of the Laplacian matrix L: Update L using the gradient: Where γ is the step size, k' is the introduced variable which is equal to the filter order, and r is the parameter introduced for partial derivative; Step 5-4: Update the auxiliary variable Z according to the converged Laplace matrix L: Where Π represents the Euclidean projection, which projects L onto the constraint set L; Step 5-5, combine Z and the learned L to get the new Lagrange multiplier P: P i+1 =P i +ρ(Z i+1 -L i+1 )。 5. A sensor network graph structure learning method according to claim 1, characterized in that: In step 6, by adjusting the coefficients, a graph structure similar to the sparsity of the actual graph structure is obtained.
