Distributed coupling tensor flow decomposition method based on sensor network
By adopting the distributed coupled tensor flow decomposition method in the sensor network, using the alternating direction multiplier method and the average consensus algorithm, the problem of coupled tensor flow data processing in the sensor network is solved, and efficient decomposition and low communication complexity are achieved. .
Patent Information
- Application Number
- CN202410963056.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2024-11-08
Smart Images

Figure CN118917075A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of signal processing, and in particular relates to a distributed coupled tensor flow decomposition method based on a sensor network. Background Art
[0002] Tensors generalize matrices to more than two dimensions. High-order tensors and their decompositions are widely used in signal processing, data analysis, and machine learning. In sensor networks (SNs), the features observed by different sensors may overlap. The multidimensional data measured can be represented by a coupled tensor model to capture multidimensional features. Since most tensor decomposition algorithms are proposed for batch processing, and there are relatively few studies on data streams, the study of coupled tensor flow decomposition is of great significance. Summary of the invention
[0003] In view of the above problems, the present invention proposes a distributed coupled tensor flow decomposition method based on sensor networks.
[0004] The technical solution adopted by the present invention is:
[0005] A distributed coupled tensor flow decomposition method based on a sensor network comprises the following steps:
[0006] S1. For Figure 1 The distributed sensor network shown in the figure sets Represents a directed dynamic graph composed of sensor nodes, and the node set is Edge sets that change over time represents a bidirectional data link, and the neighborhood of sensor node k at time t is denoted as and represents the degree of node k at time t;
[0007] S2. Model the tensor slice received at sensor node k at time t as:
[0008]
[0009] in, is an N-order tensor flow at time t and node k∈V, with rank R; is a unit diagonal tensor of order N, satisfying r1=…=r if and only if r1=…=r N hour otherwise represents the p-th shared factor matrix of node k, p=1,…,P; represents the private factor matrix of node k, I q It is tensor flow Middle I q -The number of 1-dimensional tensors, q=P+1,…,N; is the nth factor matrix of node k at time t; is the tensor measurement noise;
[0010] S3. Consider only the random walk of the factor matrix, fix the modulus n and assume that when n′≠n and n′=1,2,...,N It is known that the convex optimization problem for establishing coupled tensor flow decomposition is:
[0011]
[0012] in, is based on is the time-varying factor matrix; F(·) is the inherent dynamic model of the factor matrix; Each sensor node Process noise at is a general convex loss function, ω k is the weight;
[0013] S4, using the alternating direction multiplier method to solve the convex optimization problem established in S3, under the condition of the Frobenius norm, the augmented Lagrangian function Initialize the three parameters in in is the dual variable, n=1,...,N, then the factor matrix is shared in the jth iteration The updates are as follows:
[0014]
[0015] Among them, Y (p),k (t) is the expansion of the tensor received by any node k along the modulo-p and z⊙ denotes the Khatri-Rao product, represents transpose, is the p-th shared factor matrix of node k obtained in the j-1-th iteration, is the dual variable obtained in the j-1th iteration, X is the optimization variable, j = 1, 2, …, ρ is the regularization factor;
[0016] The pth shared factor matrix is updated as follows:
[0017]
[0018] The dual variable obtained in the jth iteration K is the number of nodes in the sensor network, p = 1,…,P;
[0019] If the shared factor matrix estimated by the current iteration is similar to the result estimated by the previous iteration, and the difference between the two meets the set convergence condition, the iterative optimization is terminated;
[0020] S5. Under the relaxed condition, the tensor decomposition is unique, and the alternating least squares method can be used to estimate the factor matrix of each sensor node. It should be noted that if the tensor decomposition is not unique, the obtained factor matrix does not have reliable research significance. The factor matrix estimated by the alternating least squares method for the jth iteration is:
[0021]
[0022] Where: is the pth shared factor matrix estimated by node k after the jth iteration, is the qPth private factor matrix estimated by node k after the jth iteration, I R is the identity matrix of rank R,
[0023] p=1,…,P and q=P+1,…,N;
[0024] S6. Use the average consensus algorithm to improve the accuracy of the shared factor matrix between nodes. First, set And p = 1, ..., P, then perform CP decomposition according to the factor matrix of S5, and obtain the estimated factor matrix of node k only by relying on the tensor slices received by each node Then the shared factor matrix is averaged and consensus is performed on each node, where w k,k (t) is the current self-weight of node k, w k,k′ (t) is the consensus weight that node k gives to neighbor node k′, specifically:
[0025] In the initial stage of the average consensus iteration algorithm, the pth factor matrix of node k is set and its dual variable for have:
[0026]
[0027] in, and are the pth factor matrix and dual variable obtained by node k after the lth consensus iteration, and They are the pth factor matrix and dual variable obtained by the neighboring point k′ after the l-1th consensus iteration;
[0028] After completing L iterations, let p=1,…,P have:
[0029]
[0030] in, and are the pth factor matrix and dual variable obtained by node k after the Lth consensus iteration, is the factor matrix of the p-th estimate of node k;
[0031] Use the consensus result as the initial value of the alternating least squares algorithm at time t+1, and repeat the process for P iterations;
[0032] S7. After the iterative optimization is completed, the estimated factor matrix of node k is obtained as follows:
[0033] Beneficial effects of the present invention: The present invention is a method for fully distributed coupled tensor flow decomposition. First, it is assumed that the measurement data of all sensors are available and cannot be processed using a distributed method, thereby deriving a coupled tensor flow decomposition algorithm; then the average consensus (AC) algorithm is used to fuse the same features, thereby realizing tensor flow decomposition in a fully distributed manner. This method does not require tensor measurement data from previous time points, and the computational complexity will not increase over time, because the parameters estimated for each dimension at the current moment will be used for initialization at the next moment. Computer simulation results show that the present invention uses a fully distributed approach to maintain the private dimensions of each node; only the shared factor matrix is exchanged through the sensor network to ensure low communication complexity. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is a schematic diagram of a distributed sensor network constructed by the present invention.
[0035] Figure 2 It is a curve showing the NMSE of load factor estimation for different network types changing over time.
[0036] Figure 3 It is the curve of NMSE estimated by load factor of different number of nodes changing with time. DETAILED DESCRIPTION
[0037] The technical solution of the present invention is described in detail below.
[0038] The process of the method of the present invention is specifically as follows:
[0039] In order to maintain generality, we assume that there are P shared factor matrices, the first P factor matrices of each node are shared, and there is a hard coupling relationship between the shared factor matrices, that is, in p = 1, ..., P, then the tensor slice received at the sensor node k at each time point can be expressed as:
[0040]
[0041] Where: is the private factor matrix of node k, where q = P + 1, ..., N. For any The expansion of the received tensor along modulo-n can be expressed as in is the nth factor matrix of the tensor received by node k at time t,
[0042] Assuming that the measurement data of all sensors can be used, the decomposition problem of coupled tensor flow can be transformed into the optimization problem shown in formula (2). Since the dynamic model F(·) of the factor matrix is actually complex and unknown, for simplicity, only the random walk of the factor matrix is considered, and the modulus n is fixed and it is assumed that when n′≠n It is known that, thus the transformed simplified convex optimization problem is obtained, where: is a general convex loss function, is the nth factor matrix of the tensor received by node k at time t-1, and ω k ,λ k is the appropriate weight.
[0043]
[0044] The dynamic model F(·) of the factor matrix is actually complex and unknown. For simplicity, we only consider the random walk of the factor matrix, that is:
[0045]
[0046] in, Each node The process noise at the
[0047] Therefore, formula (2) can be simplified as:
[0048]
[0049] In this case, the problem described by equation (4) is non-convex for all parameters, but if we fix the modulus n and assume that when n′≠n It is known that equation (4) can be transformed into a convex optimization problem as shown in equation (5), where n = 1,…,N.
[0050]
[0051] The alternating direction method of multipliers (ADMM) can be used to solve this convex optimization problem, where the augmented Lagrangian function is:
[0052]
[0053] where ρ is the regularization factor, and are the nth and pth dual variables of node k at time t, n=1,...,N.
[0054] Under the condition of Frobenius norm, after initializing the three parameters in the augmented Lagrangian function, we get The update of the shared factor matrix in the jth iteration is shown in equation (7), where j = 1, 2, ...,
[0055]
[0056] Then the update method of the pth shared factor matrix is shown in formula (8), where p=1,…,P. If the shared factor matrix estimated by the current iteration is similar to the result estimated by the previous iteration, and the difference between the two meets the set convergence condition, the iterative optimization can be terminated.
[0057]
[0058] The average processing of node k according to equation (8) shows that AC can be used without any fusion center to achieve fully distributed estimation. First, the factor matrix of each node is estimated using alternating least squares (ALS), and then AC is used to improve the accuracy of the shared factor matrix between nodes. Once the tensor slice of each node is obtained, according to equation (7), the factor matrix estimated at the jth iteration is:
[0059]
[0060] Where: p = 1, ..., P and q = P + 1, ..., N. Equations (9) and (10) are calculated independently at each node. In essence, the low-rank tensor decomposition is unique under relaxed conditions and has ordering and scale ambiguity as shown in Equation (11).
[0061]
[0062] Where: k (t) is a permutation matrix used to ensure that the actual tensor is consistent with the estimated tensor in each dimension; is such that Λ (1) (t)…Λ (N) (t) = I full rank diagonal scaling matrix.
[0063] In order to aggregate the tensor slice information provided by all nodes, AC is used to calculate equation (8).
[0064] First set And p = 1, ..., P, then perform CP decomposition according to equations (9) and (10), and obtain only the tensor slices received by each node Then AC is performed on the shared factor matrix at each node by the following steps, where w k,k′ (t) is the consensus weight of node k.
[0065] set up have:
[0066]
[0067] After completing L iterations, we have:
[0068]
[0069] Then the consensus result is used as the initial value of the alternating least squares algorithm at time t+1, and the above iterative process is repeated P times. After the iterative optimization is completed, the estimated factor matrix of node k is
[0070] Simulation Example
[0071] This example uses MATLAB to simulate and verify the above-mentioned distributed coupled tensor flow decomposition method based on average consensus. The specific execution process is:
[0072]
[0073] Simulation conditions and parameters: The elements in are independent of each other and follow a Gaussian distribution with a mean of 0 and a variance of 1. The weighting parameter, regularization parameter and tolerance parameter of are set to ω k =1, ρ =1 and η k =10 -6 ; At t = 0, randomly initialize and And the elements in the two are independent of each other and obey a Gaussian distribution with a mean of 0 and a variance of 1. In order to simplify the model, the limited resolution of the sensor is not considered and the detection probability P is assumed to be d =1; In order to evaluate the accuracy of factor matrix estimation, the normalized mean square error (NMSE) of the factor matrix is used. When the network types are non-connected network, fully connected network, adjacent network and dynamic network, and when the number of nodes is 5, 15, 25 and 35 respectively, N=4 is set. And assuming that the factor matrix is reinitialized at t = 100, all simulation results are the average of 1000 Monte Carlo tests. Figure 2 and Figure 3 The simulation results show that the algorithm has good convergence and the parameter changes have little effect on the load factor estimation performance, which verifies the effectiveness of this method.
Claims
1. A distributed coupled tensor flow decomposition method based on sensor networks, It is characterized in that The following steps are involved: S1. For distributed sensor networks, set Represents a directed dynamic graph composed of sensor nodes, and the node set is Edge sets that change over time represents a bidirectional data link, and the neighborhood of sensor node k at time t is denoted as and represents the degree of node k at time t; S2. Model the tensor slice received at sensor node k at time t as: in, is an N-order tensor flow at time t and node k∈V, with rank R; is a unit diagonal tensor of order N, satisfying if and only if r 1 =…=r N hour otherwise represents the p-th shared factor matrix of node k, p=1,…,P; represents the qPth private factor matrix of node k, I q It is tensor flow Middle I q -The number of 1-dimensional tensors, q=P+1,…,N; is the nth factor matrix of node k at time t; is the tensor measurement noise; S3. Consider only the random walk of the factor matrix, fix the modulus n and assume that when n′≠n and n′=1,2,...,N It is known that the convex optimization problem for establishing coupled tensor flow decomposition is: in, is based on is the time-varying factor matrix; F(·) is the inherent dynamic model of the factor matrix; Each sensor node Process noise at is a general convex loss function, ω k is the weight; S4, using the alternating direction multiplier method to solve the convex optimization problem established in S3, under the condition of the Frobenius norm, the augmented Lagrangian function Initialize the three parameters in in is the dual variable, n=1,...,N, then the factor matrix is shared in the jth iteration The updates are as follows: Among them, Y (p),k (t) is the expansion of the tensor received by any node k along the modulo-p and ⊙ denotes the Khatri-Rao product, represents transpose, is the p-th shared factor matrix of node k obtained in the j-1-th iteration, is the dual variable obtained in the j-1th iteration, X is the optimization variable, j = 1, 2, …, ρ is the regularization factor; The pth shared factor matrix is updated as follows: The dual variable obtained in the jth iteration K is the number of nodes in the sensor network, p = 1,…,P; If the shared factor matrix estimated by the current iteration is similar to the result estimated by the previous iteration, and the difference between the two meets the set convergence condition, the iterative optimization is terminated; S5. Under the more relaxed condition, the tensor decomposition is unique. The factor matrix of each sensor node is estimated using the alternating least squares method. The factor matrix estimated by the jth iteration is obtained using the alternating least squares method: In the formula, is the pth shared factor matrix estimated by node k after the jth iteration, It is the qPth private factor matrix estimated by node k after the jth iteration, I R is the identity matrix of rank R, with p = 1,…,P and q = P+1,…,N; S6. Use the average consensus algorithm to improve the accuracy of the shared factor matrix between nodes. First, set And p = 1, ..., P, then perform CP decomposition according to the factor matrix of S5, and obtain the estimated factor matrix of node k only by relying on the tensor slices received by each node Then the shared factor matrix is averaged and consensus is performed on each node, where w k,k (t) is the current self-weight of node k, w k,k′ (t) is the consensus weight that node k gives to neighbor node k′, specifically: In the initial stage of the average consensus iteration algorithm, the pth factor matrix of node k is set and its dual variable have: in, and are the pth factor matrix and dual variable obtained by node k after the lth consensus iteration, and They are the pth factor matrix and dual variable obtained by the neighboring point k′ after the l-1th consensus iteration; After completing L iterations, let p=1,…,P have: in, and They are the pth factor matrix and dual variable obtained by node k after the Lth consensus iteration, is the factor matrix of the p-th estimate of node k; Use the consensus result as the initial value of the alternating least squares algorithm at time t+1, and repeat the process for P iterations; S7. After the iterative optimization is completed, the estimated factor matrix of node k is obtained as follows:
