Distributed multi-sensor tracking method based on privacy protection and anomaly suppression

Abstract

The invention provides a distributed multi-sensor tracking method based on privacy protection and anomaly suppression. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression comprises the following steps: acquiring original measurement information of local sensors; according to the original measurement information, adopting a pre-established robust Kalman filter to suppress random abnormal interference to obtain a local state estimation value; and encrypting the local state estimation value by adopting a dynamic mask mechanism, and fusing the encrypted estimation values transmitted between the adjacent sensor nodes by adopting an average consensus filter according to the topological structure of the network system, thereby obtaining global consistency state estimation of the multi-sensor system. According to the method, the local state estimation value is encrypted, privacy information is protected, and the method is applied to a multi-sensor scene through global consistency state estimation of a multi-sensor system, so that the technical problem that a distributed sensor tracking method is lacked in the prior art to avoid privacy leakage is solved.

Description

Technical Field:

[0001] The present invention relates to the field of artificial intelligence, and particularly to a distributed multi-sensor tracking method based on privacy protection and anomaly suppression. Background Art:

[0002] Currently, multi-sensor network systems have been widely applied in fields such as target tracking and environmental monitoring. Among them, one of the key issues is how to design a suitable state estimator that can accurately estimate the state by using the information of sensor nodes.

[0003] Existing state estimation schemes can be roughly divided into two categories: centralized state estimation and distributed state estimation algorithms. The centralized scheme uses a central estimator to batch process the measurement values of multiple sensors to generate a globally optimal state estimation vector, thus greatly increasing the communication cost; in addition, the centralized scheme also requires complex matrix solution operations. Different from the centralized scheme, there is no fusion center in the distributed estimation, and the filtering algorithm is deployed on each sensor node, using the information propagated by neighbor nodes and local measurement values to provide state estimation. Compared with the centralized scheme, the distributed scheme has lower communication overhead and more flexible application scenarios.

[0004] It can be seen from the execution process of the distributed scheme that obtaining local observations and exchanging data between nodes are crucial for the effective operation of the distributed Kalman filter. However, in engineering practice, due to accidental environmental disturbances, intermittent sensor failures, and potential network attacks, sensor nodes are usually interfered by outliers. Ignoring these interference factors will lead to an increase in estimation errors and even divergence. In addition, distributed estimation requires the exchange of sensitive information between neighbor nodes, which raises concerns about privacy issues. Since directly exchanging information easily causes information leakage, it thus limits the application scenarios of distributed estimation.

[0005] Therefore, there is an urgent need for a distributed multi-sensor tracking method based on privacy protection and anomaly suppression, which helps to solve the technical problem in the prior art that there is a lack of a distributed sensor tracking method to avoid privacy leakage. Summary of the Invention:

[0006] The present invention provides a distributed multi-sensor tracking method based on privacy protection and anomaly suppression, which encrypts the local state estimation values to protect privacy information, and through the global consistency state estimation of the multi-sensor system, it is applied to multi-sensor scenarios, which helps to solve the technical problem in the prior art that there is a lack of a distributed sensor tracking method to avoid privacy leakage.

[0007] The distributed multi-sensor tracking method based on privacy protection and anomaly suppression includes:

[0008] Obtain the original measurement information of local sensors;

[0009] A pre-established robust Kalman filter is used to suppress random abnormal interference according to the original measurement information to obtain a local state estimation value;

[0010] A dynamic mask mechanism is used to encrypt the local state estimation value, and an average consensus filter is used to fuse the encrypted estimation values transmitted between adjacent sensor nodes according to the topological structure of the network system, thereby obtaining a global consistent state estimation of the multi-sensor system.

[0011] In one embodiment, the step of suppressing random abnormal interference using a pre-established robust Kalman filter according to the original measurement information to obtain a local state estimate includes:

[0012] Determine the topology of the sensor network system and the system state space model;

[0013] Initialize the initial state value of the sensor node and the initial covariance matrix Μ0;

[0014] Based on the pre-established robust Kalman filter, combined with the system dynamics model, the state prediction is performed to obtain the prior state vector and the prior covariance matrix M i,k / k-1 ;

[0015] Get the measurement information y of each local sensor node i,k , integrating the information shared by neighboring nodes to obtain the local intermediate state vector and the local posterior covariance matrix M i,k .

[0016] In one embodiment, a connected multi-sensor network is modeled as an undirected graph G = (N, ε) with a node set N, where |N| represents the number of nodes in the multi-sensor network and the neighborhood of a node, i.e., the set of nodes connected to node i, is defined as N i .

[0017] In one embodiment, the system state model and the measurement model are established as follows:

[0018] x k =Ax k-1 +w k (17)

[0019] y i,k =H i,k x i,k +v i,k (18)

[0020] Among them, x kis the state vector at time k; y i,k is the measurement vector of the i-th sensor at time k; w k ∈(0, Q) represents system noise with a mean of 0 and a variance of Q; v i,k ∈(0, R) is measurement Gaussian noise with a mean of 0 and a variance of R; A is the system matrix; H is the measurement matrix.

[0021] In one embodiment, for sensor node i:

[0022] Perform state prediction:

[0023]

[0024] where is the global posterior state estimate value of sensor node i at time k - 1;

[0025] Perform error covariance matrix prediction:

[0026] M i,k / k-1 = AM i,k-1 A T + Q (20)

[0027] where M i,k-1 is the global posterior state estimate value at time k - 1, and A T represents the transposed form of the system matrix A.

[0028] In one embodiment, node i shares the information matrix Γ with adjacent nodes i,k , and fuses the information matrices of neighbor nodes to obtain the posterior covariance matrix M i,k , and the formula is as follows:

[0029]

[0030] Calculate the Kalman filter gain K i,k :

[0031]

[0032] Perform measurement update to obtain the local intermediate state vector and the local posterior covariance matrix M i,k , as follows:

[0033]

[0034] where sat(·) is the outlier suppression function.

[0035] In one embodiment, the measurement information y of each local sensor node is obtained i,k , and the information shared by neighbor nodes is fused to obtain the local intermediate state vector and the local posterior covariance matrix M i,k During the process, the average consensus filter The specific design process is as follows:

[0036]

[0037] where Γ i,k (m) is the information matrix of the i-th sensor after the m-th consensus iteration at time k, and w i,l is the weight between the i-th sensor node and the l-th sensor node, is the inverse matrix of the local posterior covariance, and Γ i,k (n) is the information matrix after n iterations.

[0038] In one embodiment, the specific design method of the outlier suppression mechanism is:

[0039]

[0040] where represents the innovation at time k; is the upper bound of the saturated innovation, which is usually preset by historical experience.

[0041] In one embodiment, a dynamic masking mechanism is used to encrypt the local estimated value, and the average consensus filter is used to fuse the encrypted estimated values transmitted between adjacent sensor nodes according to the topology of the network system to obtain the global consistent state estimation of the multi-sensor system. The specific process is as follows:

[0042] Use the dynamic masking function h(·) to encrypt the local intermediate state estimated value, so that:

[0043]

[0044] where r i,k (m - 1) is the encrypted intermediate state estimated value of the i-th sensor node after m - 1 iterations at time k; h i (·) is the masking function of sensor node i; m - 1 is the number of consensus iterations; π i is the input parameter of the masking function, and then we get:

[0045]

[0046] where is the intermediate state estimation vector of the i-th sensor after the m-th consensus iteration at time k, and w i,l is the weight between the i-th sensor node and the j-th sensor node;

[0047] Obtain the global consensus state estimation of the multi-sensor network system after n iterations

[0048] In one embodiment, the specific method for designing the dynamic mask is as follows:

[0049] (1) Define the dynamic mask function h(·):

[0050]

[0051] (2) Design the dynamic mask function as:

[0052]

[0053] where φ, σ, δ are positive scalars; γ is a vector, and the vector dimension is consistent with the dimension of the state of the state. Description of the drawings:

[0054] Figure 1 is a schematic flow chart of a distributed multi-sensor tracking method based on privacy protection and anomaly suppression in an embodiment of the present invention;

[0055] Figure 2 is a schematic diagram of the multi-sensor information fusion process in another embodiment of the present invention;

[0056] Figure 3 is a schematic diagram of the privacy protection mechanism in another embodiment of the present invention. Detailed implementation manners:

[0057] Figure 1 is a schematic flow chart of a distributed multi-sensor tracking method based on privacy protection and anomaly suppression in an embodiment of the present invention; Figure 2 is a schematic diagram of the multi-sensor information fusion process in another embodiment of the present invention; Figure 3 is a schematic diagram of the privacy protection mechanism in another embodiment of the present invention.

[0058] The object of the present invention is to overcome the deficiencies in the prior art, and provide a distributed multi-sensor tracking method and system, which adopts an anomaly suppression measurement outlier and uses an output mask mechanism to achieve privacy protection of node sensitive information. Specifically, according to the original measurement information of local sensors, a pre-established robust Kalman filter is used to suppress random anomaly interference to obtain a local state estimation; a dynamic mask mechanism is used to encrypt the local estimation value, and according to the topological structure of the network system, an average consensus filter is used to fuse the encrypted estimation values transmitted between adjacent sensor nodes to obtain the global consensus state estimation of the multi-sensor system. The suppression of anomaly interference and the privacy protection of sensitive data are achieved while ensuring the estimation accuracy.

[0059] As Figures 1 to 3 shown, in one embodiment, the present invention provides a distributed multi-sensor tracking method based on privacy protection and anomaly suppression. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression includes:

[0060] S101. Obtain the original measurement information of local sensors.

[0061] In this step, a specific step for obtaining the original measurement information of local sensors is provided.

[0062] S102. According to the original measurement information, use a pre-established robust Kalman filter to suppress random anomaly interference to obtain a local state estimate.

[0063] In this step, a specific step for using a pre-established robust Kalman filter to suppress random anomaly interference according to the original measurement information to obtain a local state estimate is provided.

[0064] S103. Use a dynamic masking mechanism to encrypt the local state estimate, and use an average consensus filter to fuse the encrypted estimates transmitted between adjacent sensor nodes according to the topological structure of the network system, thereby obtaining a global consistent state estimate of the multi-sensor system.

[0065] In this step, a specific step for encrypting the local state estimate using a dynamic masking mechanism is provided.

[0066] In this embodiment, a specific implementation manner of a distributed multi-sensor tracking method based on privacy protection and anomaly suppression is provided.

[0067] In one embodiment, the step of using a pre-established robust Kalman filter to suppress random anomaly interference according to the original measurement information to obtain a local state estimate includes:

[0068] Determine the topological structure and system state space model of the sensor network system;

[0069] Initialize the state initial value of the sensor node and the initial covariance matrix Μ0;

[0070] Based on the pre-established robust Kalman filter, combine the system dynamics model to perform state prediction to obtain a priori state vector and a priori covariance matrix M i,k / k-1 ;

[0071] Obtain the measurement information y of each local sensor node i,k , fuse the information shared by neighbor nodes to obtain a local intermediate state vector and a local posterior covariance matrix Mi,k 。

[0072] In one embodiment, a connected multi-sensor network is modeled as an undirected graph G = (N, ε) with a node set N, where |N| represents the number of nodes in the multi-sensor network. The neighborhood of a node, i.e., the set of nodes connected to node i, is defined as N i 。

[0073] In this embodiment, a connected multi-sensor network is considered and modeled as an undirected graph G = (N, ε) with a node set N.

[0074] In one embodiment, the system state model and the measurement model are established as:

[0075] x k = Ax k-1 + w k (33)

[0076] y i,k = H i,k x i,k + v i,k (34)

[0077] where x k is the state vector at time k; y i,k is the measurement vector of the i-th sensor at time k; w k ∈(0, Q) represents system noise with a mean of 0 and a variance of Q; v i,k ∈(0, R) is measurement Gaussian noise with a mean of 0 and a variance of R; A is the system matrix; H is the measurement matrix.

[0078] In one embodiment, for sensor node i:

[0079] Perform state prediction:

[0080]

[0081] where, is the global posterior state estimate value of sensor node i at time k - 1;

[0082] Perform error covariance matrix prediction:

[0083] M i,k / k-1 = AM i,k-1 A T + Q (36)

[0084] where M i,k-1 is the global posterior state estimate value at time k - 1, and A T represents the transposed form of the system matrix A.

[0085] In one embodiment, the measurement information y of each local sensor node is obtained i,k , and the information shared by neighbor nodes is fused to obtain the local intermediate state vector and the local posterior covariance matrix M i,k .

[0086] Node i shares the information matrix Γ i,k with adjacent nodes, and the information matrices of neighbor nodes are fused to obtain the posterior covariance matrix M i,k , and the formula is as follows:

[0087]

[0088] Calculate the Kalman filter gain K i,k :

[0089]

[0090] Perform measurement update to obtain the local intermediate state vector and the local posterior covariance matrix M i,k , as follows:

[0091]

[0092] where sat(·) is the outlier suppression function.

[0093] In one embodiment, in the process of obtaining the measurement information y of each local sensor node i,k , and fusing the information shared by neighbor nodes to obtain the local intermediate state vector and the local posterior covariance matrix M i,k , the specific design process of the average consensus filter is as follows:

[0094]

[0095] where Γ i,k (m) is the information matrix of the i-th sensor after the m-th consensus iteration at time k, w i,l is the weight between the i-th sensor node and the l-th sensor node, is the inverse matrix of the local posterior covariance, Γ i,k (n) is the information matrix after n iterations.

[0096] In one embodiment, the specific design method of the outlier suppression mechanism is:

[0097]

[0098] where represents the innovation at time k; is the upper bound of the saturated innovation, which is usually preset by historical experience.

[0099] In one embodiment, a dynamic masking mechanism is used to encrypt the local estimated value, and an average consensus filter is used to fuse the encrypted estimated values transmitted between adjacent sensor nodes according to the topological structure of the network system to obtain the global consistent state estimation of the multi-sensor system. The specific process is as follows:

[0100] Use the dynamic masking function h(·) to encrypt the local intermediate state estimated value, so that:

[0101]

[0102] where r i,k (m - 1) is the encrypted intermediate state estimated value of the i-th sensor node after m - 1 iterations at time k; h i (·) is the masking function of sensor node i; m - 1 is the consensus iteration number; π i is the input parameter of the masking function, and then we get:

[0103]

[0104] where, is the intermediate state estimation vector of the i-th sensor after the m-th consensus iteration at time k, w i,l is the weight between the i-th sensor node and the j-th sensor node;

[0105] After n iterations, the global consistent state estimation of the multi-sensor network system is obtained

[0106] In one embodiment, the specific method for designing the dynamic mask is as follows:

[0107] (1) Define the dynamic masking function h(·):

[0108]

[0109] (2) Design the dynamic masking function as:

[0110]

[0111] where φ, σ, δ are positive scalars; γ is a vector, and the dimension of the vector is the same as the dimension of the state

[0112] Beneficial effects:

[0113] ​1. Compared with traditional robust filtering methods, the present invention adds saturation boundaries to measurement information to suppress the innovation mutation caused by anomalies, while suppressing the interference of outliers and maintaining higher estimation accuracy;

[0114] 2. Compared with differential privacy methods, the present invention applies dynamic masking technology to shared sensitive information, which has a faster convergence speed and can achieve accurate global consensus;

[0115] 3. The method adopted by the present invention is simple and easy to implement, requires few parameters to be adjusted, and has wide applicability.

Claims

1. A distributed multi-sensor tracking method based on privacy protection and anomaly suppression, characterized in that, The distributed multi-sensor tracking method based on privacy protection and anomaly suppression includes: Obtain the original measurement information of local sensors; Suppress random anomaly interference according to the original measurement information by using a pre-established robust Kalman filter to obtain local state estimation values; Encrypt the local state estimation values by using a dynamic masking mechanism, and fuse the encrypted estimation values transmitted between adjacent sensor nodes by using an average consensus filter according to the topological structure of the network system, so as to obtain the global consistent state estimation of the multi-sensor system.

2. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression according to claim 1, wherein The step of suppressing random anomaly interference according to the original measurement information by using a pre-established robust Kalman filter to obtain local state estimation values includes: Determine the topological structure and system state space model of the sensor network system; Initialize the state initial value of the sensor node and the initial covariance matrix Μ0; Based on a pre-established robust Kalman filter, combined with the system dynamics model for state prediction, a prior state vector is obtained and a prior covariance matrix M i,k / k-1 ; Obtain the measurement information y of each local sensor node i,k , and fuse the information shared by neighbor nodes to obtain the local intermediate state vector and the local posterior covariance matrix M i,k .

3. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression according to claim 2, characterized in that A connected multi-sensor network is modeled as an undirected graph G=(N, ε), where |N| represents the number of nodes in the multi-sensor network. The neighborhood of a node, i.e., the set of nodes connected to node i, is defined as N i .

4. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression according to claim 3, characterized in that Establish a system state model and a measurement model as follows: x k = Ax k-1 + w k (1) y i,k = H i,k x i,k + v i,k (2) where, x k is the state vector at time k; y i,k is the measurement vector of the i-th sensor at time k; w k ∈(0, Q) represents system noise with mean 0 and variance Q; v i,k ∈(0, R) is measurement Gaussian noise with mean 0 and variance R; A is the system matrix; H is the measurement matrix.

5. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression according to claim 4, wherein For sensor node i: Perform state prediction: Among them, is the global posterior state estimate value of the sensor node i at the (k - 1)th moment; Perform error covariance matrix prediction: M i,k / k-1 = AM i,k-1 A T + Q(4) Among them, M i,k-1 is the global posterior state estimation value at time k-1, and A T represents the transposed form of the system matrix A.

6. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression according to claim 5, characterized in that, Node i shares the information matrix Γ with adjacent nodes i,k and fuses the information matrices of neighbor nodes to obtain the posterior covariance matrix M i,k The formula is as follows: Calculate the Kalman filter gain K i,k : Perform measurement update to obtain the local intermediate state vector and the local posterior covariance matrix M i,k , as follows: where sat(·) is an outlier suppression function.

7. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression according to claim 6, wherein Obtaining the measurement information y of each local sensor node i,k , and fusing the information shared by neighbor nodes to obtain the local intermediate state vector and the local posterior covariance matrix M i,k In the process, the average consensus filter is designed as follows: Among them, Γ i,k (m) is the information matrix of the i-th sensor after the m-th consensus iteration at time k, w i,l is the weight between the i-th sensor node and the l-th sensor node, is the inverse matrix of the local posterior covariance, Γ i,k (n) is the information matrix after n iterations.

8. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression according to claim 7, characterized in that The specific design method of the outlier suppression mechanism is: Among them, represents the innovation at time k; is the upper bound of the saturated innovation, which is usually preset by historical experience.

9. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression according to claim 8, characterized in that Encrypt the local estimation values by using a dynamic masking mechanism, and fuse the encrypted estimation values transmitted between adjacent sensor nodes by using an average consensus filter according to the topological structure of the network system to obtain the global consistent state estimation of the multi-sensor system. The specific process is as follows: Encrypt the local intermediate state estimation values by using a dynamic masking function h(·), so that: where r i,k (m - 1) is the estimated value of the encrypted intermediate state after the (m - 1)-th iteration at the k-th moment of the i-th sensor node; h i (·) is the masking function of sensor node i; m - 1 is the number of consensus iterations; π i is the input parameter of the masking function, and then we get: Among them, is the intermediate state estimation vector of the i-th sensor after the m-th consensus iteration at time k, w i,l is the weight between the i-th sensor node and the j-th sensor node; Obtain the global consensus state estimate of the multi-sensor network system after n iterations 10. The distributed multi-sensor tracking method based on privacy protection and anomaly suppression according to claim 9, characterized in that The specific design method of the dynamic mask is as follows: (1) Define the dynamic masking function h(·): (2) Design the dynamic masking function as: where φ, σ, δ are positive scalars; γ is a vector, and the dimension of the vector is consistent with the dimension of the state vector.

Images