Methods, devices, and systems for controlling status parameters via communication equipment

By using a Kalman filter and a process noise model to iteratively estimate state parameters in communication devices, the problem of observation noise when communication devices only receive data from a subset of devices is solved, achieving fast convergence and efficient consensus algorithm performance.

CN117425861BActive Publication Date: 2026-11-17MITSUBISHI ELECTRIC CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202180099093.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2021-06-11
Filing Date
2021-12-16
Publication Date
2026-11-17
Estimated Expiration
2041-12-16

AI Technical Summary

Technical Problem

Existing cooperative control algorithms are susceptible to observation noise when communication devices receive data from only some devices, resulting in slower consensus algorithm convergence and increased communication data volume. Existing solutions are either unsuitable or uneconomical.

Method used

A Kalman filter is used to iteratively estimate the local values ​​of the state parameters. Data is exchanged between the processing equipment and other communication equipment. A process noise model is used for modeling, and the covariance matrix is ​​dynamically adjusted to reduce the influence of observation noise. The covariance matrix of the process noise is controlled to achieve convergence.

Benefits of technology

Without increasing the amount of communication data, the impact of observation noise on the consensus algorithm is effectively reduced, ensuring that the communication equipment can quickly converge to the consensus value and improving system performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117425861B_ABST
    Figure CN117425861B_ABST
Patent Text Reader

Abstract

This disclosure relates to a method (40) for controlling state parameters of a communication device (20) referred to as a processing device (20) in a set of communication devices (20), a subset of which is associated with the processing device, the subset including the processing device and the communication devices, wherein the processing device receives data from the communication devices for updating (S42) local values ​​of the state parameters, the processing device estimates (S41) aggregated state parameters of the local values ​​of the state parameters of the aggregated subset of communication devices by using a Kalman filter of an applied process model, wherein errors introduced by the process model are modeled as process noise, and the processing device updates the local values ​​of its state parameters based on the estimated aggregated state parameters; wherein the processing device (20) determines (S44) the convergence level of the local values ​​of the state parameters and modifies (S45) the covariance matrix of the process noise based on the determined convergence level.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to communication systems, and more specifically to the cooperative control of multiple communication devices, also known as a multi-agent system (MAS). Background Technology

[0002] Cooperative control algorithms aim to enable multiple communication devices to achieve a common goal.

[0003] For example, consensus algorithms aim to enable multiple communication devices to reach an agreement on state parameters. In practice, many communication devices exchange data with their neighbors to agree on the values ​​of state parameters, thereby enabling them to work together in a coordinated manner. When multiple communication devices agree on the same state parameter value, they can be said to have reached a consensus.

[0004] Cooperative control algorithms (such as consensus algorithms or model predictive control (MPC) algorithms) are commonly applied to formation control problems in mobile robots, unmanned aerial vehicles (UAVs), automated underwater vehicles (AUVs), satellites, aircraft, spacecraft, and automated road systems. Cooperative control algorithms can also be applied to problems unrelated to formation, such as task allocation, payload transfer, role assignment, air traffic control, and synchronization.

[0005] Typical applications of cooperative control algorithms, such as consensus algorithms, could be measuring the average temperature over a geographic area from a set of remote sensors, or controlling a convoy of vehicles on a highway in order to, for example, maintain the geometric arrangement of the convoy of vehicles substantially constant over time.

[0006] Typically, communication devices exchange local values ​​of their state parameters. Therefore, each communication device can update its own local values ​​of state parameters based on the local values ​​of state parameters received from neighboring communication devices.

[0007] One problem with traditional cooperative control algorithms (such as consensus algorithms) is that they assume the local values ​​of the state parameters are perfectly known. However, these local values ​​of the state parameters are often corrupted by additive noise, which can, for example, affect the convergence of the consensus algorithm.

[0008] In the scientific literature, solutions relying on Kalman filter-based state observers have been proposed. These observers are used to denoise the local values ​​of observed state parameters in the context of a consensus algorithm. However, most solutions depend on a global formula for the consensus algorithm, meaning they only apply if each communication device receives data from all other communication devices. In most practical systems, however, communication devices can only receive data from some other communication devices, making these solutions, which rely on a global formula for the consensus algorithm, inherently unsuitable. They can be applied, for example, using multi-hop communication, to enable each communication device to receive data from all other communication devices, as proposed in NPL 1. However, this solution implies that the consensus algorithm will converge more slowly and requires a significantly increased amount of data exchanged between communication devices.

[0009] NPL 2 proposes an alternative solution for denoising the local values ​​of state parameters using a Kalman filter. In NPL 2, the Kalman filter-based state observer can be locally implemented by each communication device under certain assumptions. However, the solution proposed in NPL 2 requires each communication device to send its local covariance matrix of state parameter values ​​to other communication devices at each iteration of the consensus algorithm. Therefore, each communication device sends its local state parameter values ​​and its local covariance matrix, leading to a significant increase in the amount of data exchanged between communication devices.

[0010] Therefore, existing solutions are not directly applicable to situations where each communication device only receives data from some communication devices and / or requests the transmission of additional data that leads to contradictions, because such additional data will also be corrupted by noise.

[0011] List of cited references

[0012] [Non-patent literature]

[0013] [NPL 1]

[0014] D. Haumann, V. Willert and A. Wahrburg, “Kalman filtering in mobile consensus networks”, in 2012 IEEE International Symposium on Intelligent Control, Dubrovnik, Croatia, October 2012, pp. 944-950.

[0015] [NPL 2]

[0016] Wei Ren, RW Beard, and DB Kingston, “Multi-agent Kalman consensus with relative uncertainty”, in Proceedings of the 2005 American Control Conference, Portland, Oregon, USA, pp. 1865-1870. Summary of the Invention

[0017] This disclosure aims to improve this situation by overcoming at least some of the limitations of the prior art described above. In particular, this disclosure aims to propose a solution for denoising observations of local values ​​of state parameters when each communication device receives data only from some other communication devices, without increasing the amount of data to be exchanged.

[0018] For this purpose, and according to a first aspect, this disclosure relates to a method for controlling state parameters by a communication device, referred to as a processing device, in a set of communication devices, the processing device storing local values ​​of the state parameters and exchanging data with other communication devices to enable all communication devices to converge toward the same value (referred to as a consensus value) of the state parameters, wherein a subset of communication devices is associated with the processing device, the subset including the processing device and communication devices from which the processing device receives data, the method comprising iterative steps of the following steps:

[0019] - Receive local values ​​of status parameters from other communication devices in the subset;

[0020] - Update the local values ​​of its state parameters based on the received local values ​​of the state parameters, wherein at least some of the received local values ​​are corrupted by observation noise;

[0021] - Send the local values ​​of its status parameters to other communication devices;

[0022] The processing device iteratively estimates aggregated state parameters based on the received local values ​​of state parameters using a Kalman filter. The aggregated state parameters are a subset of the local values ​​of the communication device's state parameters. The Kalman filter relies on a process model (also called a state-space model) that includes process noise to model the inaccuracies of the process model. The processing device updates the local values ​​of its state parameters based on the estimated aggregated state parameters.

[0023] The processing device iteratively determines the convergence level of the local values ​​of the state parameters and controls the covariance matrix of the process noise based on the determined convergence level.

[0024] In a particular implementation, the method may also include one or more of the following features, which are considered individually or in any technically possible combination.

[0025] In a particular implementation, the Kalman filter process model assumes that each communication device associated with a subset of the processing device receives data only from the communication devices of that subset.

[0026] In a particular implementation, modifying the covariance matrix of the process noise includes reducing the norm of the covariance matrix as the local values ​​of the state parameters converge.

[0027] In a particular implementation, the covariance matrix of the observation noise is further modified based on the covariance matrix of the process noise, and the modification of the process noise covariance matrix includes making the ratio between the norm of the process noise covariance matrix and the norm of the observation noise covariance matrix decrease as the local values ​​of the state parameters converge.

[0028] In a particular implementation, the covariance matrix of the process noise corresponds to the covariance matrix of the observation noise multiplied by a positive weighting factor, and modifying the covariance matrix of the process noise includes reducing the weighting factor as the local values ​​of the state parameters converge.

[0029] In a particular implementation, determining the convergence level includes estimating the consensus value and calculating the distance between the local values ​​of the state parameters and the estimated consensus value.

[0030] In a particular implementation, determining the convergence level at iteration k by the processing device includes calculating the distance between two multivariate Gaussian random variables.

[0031] In a particular implementation, the two multivariate Gaussian variables include:

[0032] - Random variables, having mean and covariance matrices calculated based on local values ​​of state parameters observed in successive iterations; and

[0033] - A random variable representing an estimate of the consensus value corrupted by observed noise.

[0034] In a particular implementation, the distance calculated is the Bhattacharyya distance.

[0035] In a particular implementation, the convergence level is determined based on the number of iterations in the update step.

[0036] In a particular implementation, the processing device dynamically adjusts the process model at each iteration to account for data packets containing updated local values ​​that were not received from other communication devices in the subset.

[0037] According to a second aspect, this disclosure relates to a computer program product including instructions that, when executed by a communication device including a communication unit and processing circuitry, configure the communication device to perform the method described in any one of the embodiments of this disclosure.

[0038] According to a third aspect, this disclosure relates to a computer-readable storage medium including instructions that, when executed by a communication device including a communication unit and processing circuitry, configure the communication device to perform the method described in any one of the embodiments of this disclosure.

[0039] According to a fourth aspect, this disclosure relates to a communication device including a processing circuit and a communication unit configured to perform the method described in any one of the embodiments of this disclosure.

[0040] According to a fifth aspect, this disclosure relates to a communication system comprising a plurality of communication devices according to any one of the embodiments of this disclosure.

[0041] The invention will be better understood by reading the following description and referring to the accompanying drawings, which are given by way of example and are not limiting. Attached Figure Description

[0042] [ Figure 1 ]

[0043] Figure 1 It is a schematic diagram representing the communication link between communication devices.

[0044] [ Figure 2 ]

[0045] Figure 2 It is a local representation of the consensus algorithm.

[0046] [ Figure 3 ]

[0047] Figure 3 It is the global representation of the consensus algorithm.

[0048] [ Figure 4 ]

[0049] Figure 4 It is a local representation of the noisy consensus algorithm.

[0050] [ Figure 5 ]

[0051] Figure 5 It is the global representation of the noisy consensus algorithm.

[0052] [ Figure 6 ]

[0053] Figure 6 It is the global representation of the noisy consensus algorithm.

[0054] [ Figure 7 ]

[0055] Figure 7 It is the global representation of the noisy consensus algorithm.

[0056] [ Figure 8 ]

[0057] Figure 8 It is the global representation of the noisy consensus algorithm.

[0058] [ Figure 9 ]

[0059] Figure 9 This is a representation of the Luhnberg observer.

[0060] [ Figure 10 ]

[0061] Figure 10 It represents the successive steps of a Kalman filter.

[0062] [ Figure 11 ]

[0063] Figure 11 It is a representation of a state observer based on a Kalman filter.

[0064] [ Figure 12 ]

[0065] Figure 12 It is a representation of a Kalman filter-based state observer used in noisy consensus algorithms.

[0066] [ Figure 13 ]

[0067] Figure 13 This is a diagram illustrating the main steps of an exemplary implementation of a method for controlling state parameters.

[0068] [ Figure 14 ]

[0069] Figure 14 This is a schematic diagram of an exemplary implementation of a communication device. Detailed Implementation

[0070] In these figures, the same reference numerals denote the same or similar elements from one figure to another. For clarity, unless otherwise explicitly stated, the elements shown are not displayed to scale.

[0071] As stated above, this disclosure relates to the collaborative control of multiple communication devices.

[0072] We first introduce the conventions used in the following description, and how to evaluate the performance of cooperative control algorithms.

[0073] A) Graph Theory

[0074] It is natural to model data exchange between communication devices using directed / undirected graphs. A graph (i.e., a directed graph) consists of a pair of... Composition, among which, It is a finite non-empty set of communication devices, and It is a set of ordered pairs of communication devices (called directed edges) (∈ ij It is a directed edge or communication link from communication device j to communication device i. For simplicity, we use "edge" instead of "directed edge" and "graph" instead of "directed graph" below (in context, i.e., directed or undirected is considered clear).

[0075] In contrast, the pairs of communication devices in an undirected graph are unordered. A directed path is an ordered sequence of edges in a directed graph, which takes the form ∈ ij ,∈ ki ,∈ lk ,...,in Similarly, an undirected path in an undirected graph is defined, where ∈ ij It means ∈ ji A directed graph is called "strongly connected" if there exists a directed path from every communication device to every other communication device. An undirected graph is called connected if there exists a path between any two different communication device pairs. A directed tree is a directed graph in which every communication device has a parent node in addition to the root. A spanning tree of a directed graph is a directed tree formed by the graph edges connecting all the communication devices in the graph. If there exists a spanning tree that is a subset of the graph, we say that the graph has (or contains) a spanning tree. Note that the condition for a directed graph to have a spanning tree is equivalent to the case where there exist communication devices with directed paths to all other communication devices.

[0076] For a total of n communication devices, if ∈ ij ∈ε, where i≠j, then the adjacency matrix A of the deterministic directed graph is n =[a ij ] 0≤i,j≤n-1 Defined as a ii =0 and a ij ==1. Set a ij =0 indicates that communication device i does not (i.e., cannot or does not want to) receive data from communication device j. The Laplace matrix of a directed graph is defined as L n =[l ij ] 0≤0≤i,j≤n-1 , where l ij=-a ij And l ii =Σ j a ij ,Right now:

[0077] L n =D n -A n

[0078] in, It is the in-degree matrix (the number of communication devices from which communication device i receives data). For an undirected graph, the Laplace matrix is ​​symmetric positive semi-definite.

[0079] Figure 1 Let represent an undirected graph with n = 3 communication devices 20. In this graph, to distinguish the communication devices 20, they are referred to as 20-k, where 1 ≤ k ≤ n = 3. The associated adjacency matrix and Laplace matrix are as follows:

[0080]

[0081]

[0082] B) Consensus Algorithm

[0083] set up Here are the local values ​​of the state parameters of the i-th communication device. State parameters represent parameters that need to be coordinated between communication devices using, for example, a consensus algorithm. State parameters can be, for example, the position, velocity, or oscillation phase of the communication device.

[0084] The continuous-time single integrator static consensus algorithm can be summarized as follows:

[0085]

[0086] Where, Θ i (t) represents the local value of its state parameter available for the i-th communication device at time t (defined by the adjacency matrix). i A subset of communication devices, and α ij (t) represents the potentially time-varying weighting coefficients. In other words, the local control information of each communication device is driven by the state of the local control information of its (potentially time-varying) neighbors. This is typically extended for weighted directed graphs. An adjacency matrix is ​​defined such that (j, i) ∈ ε, where i ≠ j. otherwise

[0087] Continuous-time consensus algorithms can be written in matrix form:

[0088]

[0089] Where x = [x0(t), x1(t), ..., x n-1 [t]t is a global aggregated state parameter that includes local values ​​of the state parameters of all communication devices. I m It is an identity matrix of size m. It is the Kronecker matrix product.

[0090] In the following description, we consider a discrete version of the single-integrator static consensus algorithm in a non-restrictive manner. Assume that communication devices repeatedly update the local values ​​of their state parameters over time and transmit the updated values ​​via the communication link at equal time intervals. Each communication device updates its local value of its state parameter according to the following:

[0091]

[0092] In the following description, we consider, in a non-restrictive manner, the case of applying a consensus algorithm with a single weighting coefficient α for all communication devices:

[0093]

[0094] This system can also be represented as:

[0095]

[0096] in:

[0097] - It is a subset Θ i n in (k) i The aggregated state parameters (vector) of a communication device, whose local values ​​are available at time k;

[0098] -

[0099] The corresponding systems, such as Figure 2 As shown, and can be considered as being caused by the command signal u i,k The first-order integrator driven by the command signal u i,k Constructed as a feedback loop -x i,k-1 The weighted superposition of the external reference signal and the external reference signal, which includes contributions from other communication devices.

[0100] For the global aggregate state parameters, we have:

[0101]

[0102] in:

[0103] Wn,k =I n -αL n,k L n,k =D n,k -A n,k

[0104] Among them, D n,k and L n,k It corresponds to an unweighted directed graph (a ij =0 or a ij =1) in-degree matrix and Laplace matrix.

[0105] For the global aggregate state parameters, the system can also be written as:

[0106]

[0107] The corresponding system for global aggregate state parameters is as follows: Figure 3 As shown. Logically, the entire system is endogenous; that is, it is constructed as a first-order integrator with a feedback loop driven by globally aggregated state parameters and no external reference signal. The controller here is a weighted Laplace matrix. The command signal is established based on observations of the integrator output.

[0108] In the case of deterministic graphs, it can be shown that as long as the Laplace operator is connected (λ2(L... n )>0, where λ2(L n This corresponds to the so-called Laplace matrix L. n The second smallest eigenvalue), if α ≤ 1 / deg max If the system converges, then deg max This is the maximum degree of the graph, that is, the maximum number of neighbors that a communication device can communicate with. Note that in this case, matrix W... n,k It is random, that is, it has rows with a total of 1 and has a diagonal.

[0109] To illustrate the behavior of the consensus algorithm, we first assume that each communication device always receives scalar information from all other communication devices. Therefore, we have:

[0110]

[0111] in, and It is a completely one vector. Finally, when the matrix... It is idempotent At that time, we obtained:

[0112]

[0113] The local values ​​of the state parameters of all communication devices converge to the same value, which is the average of the initial local values. In this case, we talk about average consensus. Note that the term average is sometimes used to describe all cases where consensus converges toward a weighted average of the initial local values. The convergence of discrete-time consensus algorithms can actually be generalized to other matrices W. n,k In fact, it satisfies Where v is a column vector.

[0114] It can be proven that for directed graphs, discrete-time consensus algorithms asymptotically achieve consensus if and only if matrix W n Having eigenvector 1 n The relevant single unitary eigenvalue, i.e., λ2(W n ) < 1, where λ2(W n ) corresponds to matrix W n The second largest eigenvalue (similarly, λ2(L) n ) > 0, where λ2(Ln) corresponds to the so-called Laplace matrix L with additional constraints on the weighting coefficients. n The second smallest eigenvalue). This result effectively means that the directed graph has a spanning tree (i.e., there are no two independent graphs in the topology). Then, the consensus algorithm converges towards the following:

[0115]

[0116] Where, v = [v i ] 0≤i≤n-1 ≥0 satisfies v t W n =v t and That is, v is W associated with the eigenvalue 1. n The left-row eigenvectors. This means that the global aggregate state parameters of the communication device converge toward the same linear combination of the initial global aggregate state parameters x0, rather than necessarily toward the average value. Since the left and right eigenvalues ​​of a square matrix are the same, each random matrix has at least one row of eigenvectors associated with the eigenvalue 1. To obtain the average consensus, i.e., Matrix W n It only needs to be doubly random. In fact, discrete-time consensus algorithms can asymptotically achieve average consensus if and only if the directed graph is strongly connected (each communication device has a spanning tree) and degree-balanced (each communication device has the same number of connections in both directions - in-degree equals out-degree).

[0117] For an undirected graph, if and only if matrix W n Having eigenvector 1 nDiscrete-time consensus algorithms can asymptotically achieve consensus when associated with a single unitary eigenvalue. From a topological perspective, this means the graph is connected. If an undirected graph has a spanning tree, then it must be strongly connected and balanced (W0). n It is symmetric, therefore doubly random. The consensus algorithm converges on average.

[0118] C) The influence of noise values ​​of state parameters

[0119] Without losing generality, we now describe the consensus algorithm for scalar state parameters (i.e., m=1).

[0120] Multi-agent and more general control systems are affected by various error sources, which can be modeled as additive noise in the control system. These error sources are mainly of two types: disturbance noise and observation (or measurement) noise.

[0121] Typical types of "real" disturbances are gusts of wind or friction in a motor used to move a vehicle. If these disturbances cannot be measured to be accounted for by the controller, the control system may obtain different state values ​​for the same command. Therefore, it is important to account for the effects of these disturbances in order to drive the control system to a reference state. This is precisely the advantage of closed-loop control systems in being able to handle disturbances. Disturbances can occur as either inputs or outputs of the control system. Disturbances can also be "virtual" when designing a controller that relies on a model that cannot accurately reflect the behavior of the actual system. In the general case of a first-order consensus algorithm, disturbances can be modeled as follows:

[0122] x i,k =x i,k-1 +u i,k-1 +q i,k-1

[0123] Where, q i,k It is an additive disturbance that affects the input of the control system. This disturbance is often referred to as process noise.

[0124] This disclosure aims to mitigate the impact of a second source of error (i.e., observation (or measurement) noise). To establish the command signal for a closed-loop system, the value of the system state needs to be known. However, this state is typically not directly observed but is combined with additive observation noise. In multi-agent systems, observation noise may originate from the transmission itself. This would be the case, for example, using signals transmitted with analog modulation. Analog signals are indeed observed to have additive degradation that may be caused by the propagation channel but primarily by the first stage (thermal noise) of the receiver. In the general case of a single integrator consensus algorithm, for the i-th communication device, we have:

[0125] x i,k =x i,k-1 +u i,k-1

[0126] Among them, the command signal u i,k It is based on the local value x of the status parameter of the i-th communication device. i,k Local values ​​of state parameters of neighbors The observations were constructed. Due to noise, the i-th communication device does not observe its neighbors. The actual local values ​​of the state parameters are not the observed noisy versions:

[0127]

[0128] in:

[0129] - It is a subset Θ i The nth in (k) i The aggregated state parameter observations (vector) of a communication device, where the local values ​​of the state parameters are available at time k, y j,k This is the local value of the (noisy) state parameter obtained by the j-th communication device;

[0130] - It is the influence of n from the i-th communication device i An aggregate vector of noise values ​​of the local values ​​of the state parameters of each neighbor, η ij,k It is the noise component added to the local value of the state parameter sent from the j-th communication device to the i-th communication device.

[0131] In this case, the noise is edge-dependent, meaning that the local values ​​of the state parameters from a given communication device are observed by its neighbors with different noise components.

[0132] Without applying any methods to reduce the impact of observation noise, a consensus algorithm is established by feeding back noisy observations. The discrete-time consensus algorithm applied by the i-th communication device can be expressed as:

[0133]

[0134] in, These are the local values ​​of the state parameters used in the noisy consensus algorithm for the i-th communication device. Note that the neighbor n i The number can change over time based on the transmission status, and it can also vary depending on the communication device. The corresponding system seen by the i-th communication device is as follows: Figure 4 As shown.

[0135] Modeling the impact of global system (i.e., global aggregate state parameters) on transmission and edge-related noise is not straightforward, because the noise components are different at each edge leaving the same communication device.

[0136] A global system can be interpreted as such Figure 5 As shown, aggregation state parameters are observed on n parallel lines, with different noise components on each line. In this figure, It is the adjacency matrix A n,k The i-th row, and It is a vector of size n, where the position of the neighbor connected to the i-th communication device is... The middle part is filled with noise components, while the rest is zero. Therefore, on every line, we have:

[0137]

[0138] in:

[0139]

[0140] n parallel branches can eventually be replaced by a single branch, where the state is determined by matrix αA. n,k The following was observed:

[0141] y k =αA n,k x k +ξ k Where, ξ k =[ξ 0,k ξ 1,k ,...,ξ n-1,k ] t The corresponding systems are as follows: Figure 6 As shown. The command signal is established based on noiseless observations of the local values ​​of the state parameters, while the second component (the contribution of neighbors) should be established based on the noisy observations described above. Consensus is established by feeding back noisy observations without applying any methods to reduce the impact of observation noise.

[0142]

[0143] The presence of additive transmission noise components is uncommon because multi-agent systems rely primarily on digital modulation, where transmission impairments result in information being received correctly or not at all (packet loss). Nevertheless, local values ​​of state parameters can still be reduced by noise, for example, from the quantization of local values ​​prior to transmission. For instance, when updating the position of a communication device driven by an integrator using positioning techniques (e.g., GNSS), noise can be used to observe the local values ​​themselves. In this case, the noise is device-dependent (and edge-independent), meaning the same noise local value is sent to all of the communication device's neighbors. This is similar to the transmission noise described earlier, where the noise only affects the values ​​sent by the communication device to its neighbors. Each communication device should use a local state value x. i,k and neighbor status values The command signal u is constructed from noisy observations. i,k :

[0144]

[0145] in:

[0146] - It is the influence of n from the i-th communication device i An aggregate vector of noise values ​​of the local values ​​of the state parameters of each neighbor, η j,k It is the noise component added to the local value of the state parameter received from the j-th communication device by the i-th communication device (and through all neighbors), where j∈Θ. i (k).

[0147] When the local state value is also affected by noise, the i-th communication device should establish a command signal based on the observation results:

[0148]

[0149] in:

[0150] - It is an aggregated state parameter observation result (vector) supplemented by local observation results of the state parameters of the i-th communication device. as well as:

[0151] - It is an aggregate vector of noise values ​​expanded by noise components that affect the local values ​​of the state parameters of the i-th communication device.

[0152] When the local state value of the i-th communication device is unaffected by noise, the noise component is reduced to...

[0153] When dealing with global systems, the first two cases differ only in how the command signal is constructed. In both cases, the integrator's output is directly observable with noise, as shown below:

[0154] y k =x k +η k

[0155] Where, η k =[η 0,k η 1,k , ..., η n-1,k ] t .like Figure 7 As shown, when the noise component only affects the local values ​​of the state parameters sent to neighbors, a portion of the command signal is built based on noisy observations, while a second portion, dependent on the local values ​​of the internal state parameters, uses noise-free observations. When the noise component also affects the local values ​​of the internal state parameters, the total command signal is calculated based on the noisy observations (see [link to documentation]). Figure 8 ).

[0156] At this stage, it is important to emphasize the misleading aspects of the global description of consensus algorithms. One might notice that, since the global aggregate state parameters are available from... Figure 6 The first feedback branch provides this information, which can then be used to calculate the contribution of the third branch, which avoids noise. It should not be forgotten that the consensus algorithm is a distributed algorithm. In fact, the first branch is a diagonal matrix corresponding to the local values ​​of each communicating device's own state parameters. The second branch (adjacency matrix) brings knowledge of the local values ​​of each communicating device's neighbors. Therefore, the second branch is the core of the consensus algorithm.

[0157] This disclosure aims to reduce the impact of observation noise on the performance of consensus algorithms. In all the cases described above regarding measurement or observation noise, noisy consensus algorithms can be modeled as follows:

[0158]

[0159] Among them, the situation is in the noise component ξ k The construction differs. It can be easily shown that the algorithm can be rewritten as follows:

[0160]

[0161] in:

[0162]

[0163] Therefore, the value obtained from the noise consensus is a superposition of the noise-free value (expected value) and correlated colored noise. As an autoregressive process of order 1, the component c... kThe behavior resembles a random walk. The results show that noisy consensus algorithms cannot converge to such an equilibrium value.

[0164] D) Denoise the consensus algorithm

[0165] One possible approach is to reduce the noise level of the observations (also known as denoising). The most common method relies on stochastic signal processing algorithms, where some or all variables are described as stochastic processes with some known statistics. As shown in the equation above, a noisy consensus algorithm can be represented as a superposition of noise-free consensus values ​​and colored, correlated noise. This can be viewed as a result of noisy observations... For process x k The estimation, where the process of interest is described by linear state-space equations:

[0166]

[0167] Assuming Gaussian observations, this problem falls within the well-known framework of Kalman filtering.

[0168] E) Kalman Filter

[0169] When the measurement and observation results are linear Gaussian, the Kalman filter is a closed-form solution of the more general Bayesian recursive filter. Typically, the model is described as follows:

[0170]

[0171] in:

[0172] -x k It is the state value to be estimated;

[0173] -F k It is an n×n state transition matrix;

[0174] -v k It is a deterministic control vector;

[0175] -B k It is an n×n command input matrix;

[0176] -q k It describes the process noise (also known as state or system noise) that describes the uncertainty of the model, assumed to originate from a source with a covariance matrix Q. k Zero-mean Gaussian process

[0177] -y k It is the state observation (or measurement) vector;

[0178] -H k It is an m×n observation matrix;

[0179] -r k It is assumed that the covariance matrix R is from k Zero-mean Gaussian process The observation noise extracted from it.

[0180] Assume process r k and q k They are unrelated to each other. The initial state and the noise vector at each step are assumed to be independent.

[0181] The Bayesian filter is based on the observations y0, y1, ..., y k Recursively calculate the state value x k The marginal distribution p(x) k |y 0:k In this specific case, we have a probabilistic state-space model, where

[0182]

[0183]

[0184] Due to the distribution p(x) k |y 0:k Since the distribution is Gaussian, filtering simplifies to estimating its mean and covariance matrices.

[0185]

[0186] in:

[0187]

[0188] The Kalman filter is a linear recursive estimator for unknown states. In each iteration, it produces estimates of the mean and covariance matrices. While the Kalman filter can be written as a single equation, it is typically described as a two-stage algorithm according to Bayesian methods. The prediction stage uses state estimates from previous iterations to generate an estimate of the state for the current iteration using knowledge from the model. This predicted state estimate is also called the prior state estimate because, although it is the state estimate for the current iteration, it does not include observational information from that iteration. In the update stage, the current prior prediction is combined with the current observational information to refine the state estimate. This improved estimate is called the posterior state estimate. The Kalman equations are summarized in Table I below.

[0189] Table I

[0190]

[0191] From the marginal probability density function p(x) k |y 0:kWith knowledge of the state, any estimator for the state can be built. The most common is the Minimal MSE (MMSE). In each iteration, the filter produces an estimate of the state. Based on the observation results y0, y1, ..., y k Minimize the mean square error:

[0192]

[0193] Also note that we have:

[0194]

[0195] The Kalman filter is the optimal linear filter in the following conditions: a) the model perfectly matches the real system, b) the input noise is white (uncorrelated), and c) the covariance of the noise is fully known. In these cases, the Kalman filter is unbiased and is the minimum unbiased variance estimator (MVUE) in this respect, and the best linear unbiased estimator (BLUE) in this specific case. Many variations exist, such as the information filter and the extended Kalman filter family designed to handle nonlinear models.

[0196] As mentioned earlier, the noisy consensus system is compatible with Kalman filtering for denoising potential consensus values. However, this is not strictly accurate because the noise process c k The coloring is not due to iid but rather to its autoregressive (AR) structure. Several methods have been devised to handle this situation, starting with state augmentation algorithms where colored noise is added to the state to be estimated. However, state augmentation methods are known to suffer from implementation instability.

[0197] F) State Observer

[0198] Subsequent notation being x k Instead, a simplified notation is used (the value of the noiseless consensus algorithm is no longer considered). Based on the state observer principle, Kalman filtering can also be used for denoising systems. Let's first remind ourselves of the state observer principle in the deterministic linear case. Assume the state of a linear, time-invariant physical discrete-time system satisfies:

[0199]

[0200] Where, x k It is the state of the system, v k It is the system input, y kThis is the system output. For control purposes, the system state is expressed through the diagonal gain matrix v. k =Gx k The input that is fed back to the system.

[0201] For simplicity, let's assume the reference is zero. As shown in the equations above, the system state is not observed in this way, but rather through the observation matrix H. Therefore, it is necessary to first estimate the state to close the loop. This can be achieved using, for example, a so-called Luhnberg observer defined by the following equation:

[0202]

[0203] The principle of the Luhnberg state observer is as follows: Figure 9 As shown. Essentially, the Luhnberg observer utilizes a system model driven by the error signal between the output of the real system and the output of the model. This is illustrated when matrix FK... L When H has all eigenvalues ​​within the unit circle, the Luhnberg observer of this discrete-time system is asymptotically stable (the state error converges to zero). Therefore, by appropriately choosing the Luhnberg matrix gain K... L At that time, it can be done through the gain matrix. The output of the observer system is fed back to the input of the observer and the system. Note the command signal u. k =Bv k Not only based on the state The state values ​​can be constructed from estimates, and can be constructed based on partial knowledge of the state as previously described for noisy consensus. However, it is always necessary to estimate the state values ​​based on noisy observations.

[0204] The principle of the Luhnberg observer can be extended to systems with additive disturbances. It can be noted that the Luhnberg observer is very similar to the Kalman filter. In the case of random additive noise, the Kalman filter itself is a dynamic state observer. By inserting a process and measuring the noise contribution, the system model is modified as follows:

[0205]

[0206] Where, q k It is process noise and r k This is observation noise. Process noise should not be confused with input noise. Process noise models the correlation of the model, while observation noise models the input and output noise added to the state before the observation. The state can then be estimated using the Kalman equation:

[0207]

[0208] Connecting the Kalman filter notation provided above, we have

[0209]

[0210] As can be seen from the preceding equations, the Kalman filter generates an estimate of the state at iteration k based on the measurement obtained at iteration k. This is made possible by the prediction phase of the Kalman filter.

[0211] In consensus algorithms, observations are available with a delay of one iteration because each neighbor's local value is updated first by the neighbor and then transmitted. To overcome this problem, one can rely on, for example, the output of the prediction phase of the Kalman filter used for the current iteration, or, in the case of distributed algorithms (where the system can be updated sequentially after the Kalman filter), the output of the update phase of the Kalman filter using observations from previous iterations.

[0212] Looking at the equations, we can see that the prediction step at iteration k depends only on the input at iteration k-1, and therefore can be computed at iteration k. On the other hand, the update step requires the observed input y. k And matrix H k This knowledge will be available with a delay of one iteration. To comply with the constraints, the equations of the Kalman filter are simply rearranged as follows: Figure 10 As shown. In Figure 10 In the diagram, the left side corresponds to the traditional Kalman filter, where iterations k-1, k, and k+1 correspond to the traditional iterations of the Kalman filter. The right side corresponds to the Kalman filter equations arranged according to the consensus algorithm's iterations k, k+1, and k+2. Essentially, the Kalman predictor is used during iteration k according to the following equation:

[0213] -Predictions based on previous observations:

[0214] -Estimation of the previous state:

[0215] -Prediction of the current state:

[0216] The principle of using a Kalman filter as a state observer in a closed-loop system is as follows: Figure 9 As shown, the system is a closed loop. However, it should be noted that other implementations of the state observer based on the Kalman filter can be used in this disclosure. For example, an alternative is to simply use the Kalman gain in the Luhnberg observer.

[0217] G) Application of State Observer to Noise Consensus

[0218] As mentioned above, consensus algorithms can be described in different ways depending on the nature of the noise, and this disclosure aims to mitigate the impact of measurement or observation noise. The different cases described above differ in whether the components of the command signal, which depend on the local values ​​of the state parameters, are established from noisy observations and whether the local values ​​of the state parameters are observed directly with noise or through an observation matrix. Without loss of generality, the application of the state observer is described here in two cases: one is communication equipment-related noise, where the command signal is established entirely from noise observations, and the other is transmission edge-related noise.

[0219] a) Structure (a) - Communication Equipment Related Noise

[0220] In this structure, the consensus algorithm is modeled as a single integrator whose output is directly observed to have additive noise components. As shown in Table II, the integrator should be controlled by a feedback loop constructed from the state values ​​applied to the Laplace matrix.

[0221] Table II

[0222]

[0223]

[0224] Therefore, the models given in Table III (such as...) can be used. Figure 12 The application shown in section a) is a state observer based on a Kalman filter.

[0225] Table III

[0226]

[0227] b) Structure (b) - Transmission and Edge-Related Noise

[0228] In this architecture, the consensus algorithm is modeled as a single integrator that observes the output through an adjacency matrix with additive noise components. As shown in Table IV, the integrator is controlled by a feedback loop constructed as the difference between the local state values ​​applied to the degree matrix and the neighbor state values ​​applied to the adjacency matrix.

[0229] Table IV

[0230]

[0231] Therefore, the model given in Table V (such as...) can be used. Figure 12 The application shown in section b) is a state observer based on a Kalman filter.

[0232] Table V

[0233]

[0234]

[0235] It is important to note that the Kalman filter needs to reflect the correlation (r) present in the observation noise. k =A n,k η k ).

[0236] In both cases, the gain matrix G in the feedback loop is the identity matrix.

[0237] Due to noise, it is impossible to observe the true local values ​​of the state parameters. Based on the equations above, we use a Kalman predictor or filter to establish an observer of the local values ​​of the state parameters. Then, we replace the true state values ​​with their estimates. (or ), to create command signals (or Ultimately, the communication device "simply" implements the standard consensus algorithm, where noisy local values ​​are replaced by their estimates in the command. The problem here is that the system is described by a global state-space equation, i.e., by considering globally aggregated state parameters. When implementing a denoising Kalman filter, a Kalman filter-based state observer is not allowed to rely solely on the subset Θ. i n i The local values ​​received from each of the n neighbors are used to predict the local values ​​of the state parameters of each communication device. In contrast, the Kalman filter-based state observer described above requires knowledge of the local values ​​from all n communication devices.

[0238] H) Methods for controlling state parameters

[0239] As described above, this disclosure relates to a method 40 for controlling state parameters by communication devices 20 in a set of communication devices 20. It should be emphasized that the control method 40 is implemented by each communication device 20 of the set under consideration, or by each communication device 20 except the leader if the cooperative control algorithm uses a leader-follower approach (where one communication device—the leader—does not update its local value to force other communication devices to follow its local value). In the following, it is assumed that the control method 40 is implemented by the i-th communication device, referred to as 20-i (1 ≤ i ≤ n, where n is the total number of communication devices 20). Each communication device 20 stores its own local values ​​of the state parameters, which are typically different, and the communication devices 20 implement the control method 40 so that the local values ​​of their respective state parameters converge to the same value (referred to as the "consensus value"). For this purpose, the control method 40 uses a cooperative control algorithm, such as the consensus algorithm discussed above.

[0240] Communication device 20-i is adapted to a subset Θ of other communication devices 20-j (j≠i) i Receive data. Assume a subset Θ. i Including n i 20-j other communication devices, of which n i <n-1. Therefore, communication device 20-i cyclically receives subset Θ. i n i Each communication device 20-j stores the local values ​​of its state parameters and uses these received local values ​​to update its own local values ​​according to, for example, a consensus algorithm. Once it has updated its own local values ​​of state parameters, it sends these updated local values ​​to other communication devices, for example, to a subset Θ from which it has already received local values ​​of state parameters. i The same communication device 20-j.

[0241] We use This represents a subset of communication devices, which corresponds to subset Θ. i The union with communication device 20-i. Therefore, the subset Including n i +1 < n communication devices. As mentioned above, control method 40 is implemented by each communication device 20, and n subsets are defined. 1 ≤ i ≤ n, each communication device has one subset. Of course, there are n subsets. Not exactly the same, because each communication device 20-i only receives information from some of the other n-1 communication devices 20 (n... i ) Receive data.

[0242] In the following text, the communication device 20-i that implements the control method 40 is referred to as "processing device 20-i" to avoid confusion with other communication devices 20.

[0243] Focusing on processing device 20-i, we assume in the following text, in a non-restrictive manner, that the associated subset Θ i It does not change over time (i.e., from one iteration to another), and the number n i It does not change over time. However, in some cases, it may have a subset Θ that changes over time, for example. i And the changing number n i Or, for example, it might cause processing device 20-i to be less than n in some iterations. i 20-j (j≠i) other communication devices experience packet loss in receiving data. In the case of packet loss, the solution involves setting the corresponding observation noise covariance to a very high value for that particular iteration so that lost observations replaced by appropriately chosen values ​​are not considered.

[0244] Figure 13 The main steps of the method 40 for controlling state parameters by processing device 20-i are schematically illustrated. For example... Figure 13 As shown, the control method 40 includes several steps that are executed and iterated by the processing device 20-i.

[0245] like Figure 13 As shown, the control method 40 includes conventional steps of a cooperative algorithm, such as a consensus algorithm, executed by the processing device 20-i, which include:

[0246] - Step S40, from the subset associated with processing device 20-i Other communication devices 20 receive local values;

[0247] - Step S42, update the local value of its state parameter based on the received local value of the state parameter, wherein at least some of the received local values ​​are corrupted by observation noise;

[0248] - Step S43: Send the local value of its status parameter to other communication devices 20.

[0249] At this stage, it is important to emphasize that, Figure 13 As shown, the order of steps in each iteration is given for illustrative purposes only, and the steps may be performed in a different order. For example, when the other communication devices 20 perform the sending step S43, they perform the update step S42 after they have already performed the subset step. Other communication devices in (i.e., from subset Θ) i The communication device 20 in the middle) receives the local value in step S40. In addition, the receiving step S40 and the sending step S43 can also be executed immediately before the updating step S42, in which case the received / sent local value corresponds to the local value updated during the previous iteration, and so on.

[0250] During step S42, the processing device 20 is based on the subset Θ i The communication device 20 receives the updated local value and updates its own local value of state parameters according to the consensus algorithm.

[0251] However, since all or part of the received local values ​​are observed using observation noise, the processing device 20 does not directly use the received local values ​​and implements a Kalman filter-based state observer to denoise the local values ​​of the observed state parameters. Therefore, as Figure 13 As shown, control method 40 includes step S41, estimating a subset based on the observations of the local values ​​using a Kalman filter-based state observer. The local value. More specifically, a Kalman filter-based state observer is used to estimate aggregated state parameters, which aggregate a subset of state parameters. The local values ​​of the communication device 20 (i.e., including the local values ​​of the state parameters of the processing device 20-i). Then, during update step S42, instead of directly using the (noisy) local values ​​of the received state parameters, the processing device 20-i updates the local values ​​of its own state parameters based on the estimated aggregated state parameters. As described above, due to the subset Θ i The observations of the communication device 20 are not immediately available, but typically have a delay of one iteration. The estimated aggregate state parameters may correspond to, for example, a prior estimate (i.e., a prediction) of the aggregate state parameters (e.g., for the current iteration) or a posterior estimate of the aggregate state parameters (e.g., for previous iterations).

[0252] The state observer based on the Kalman filter applies a process model that assumes a subset of... Each communication device 20 only from the subset The communication device 20 receives data. This assumption applies to processing device 20-i, but not to the subset Θ. i All communication devices 20. In fact, subset Θ i Each communication device 20-j (j≠i) is based on its own subset Θ j The communication device in the middle receives local values ​​to update the local values ​​of its own state parameters, its own subset Θ. j Typically different from subset Θ i Therefore, the process model assumed by the Kalman filter-based state observer is incorrect; at least in this respect, it ignores the subset Θ. i The communication devices in the subset are based on never being in the subset The fact that communication devices 20 receive local values ​​to update the local values ​​of their own state parameters. The error introduced by this assumption is modeled as process noise in the process model.

[0253] In a non-restrictive manner, we hereby assume that the system is affected by transmission edge-dependent noise. Using the notation introduced above, we can assume that processing device 20-i aims to update the local values ​​of its own state parameters using a consensus algorithm:

[0254] x i,k =x i,k-1 +u i,k-1

[0255] in:

[0256]

[0257]

[0258] The problem is that the local values ​​of the received state parameters are corrupted by observation noise:

[0259]

[0260] In the following text, we will use typical Kalman notation. Note that we are only focusing on processing device 20-i here. Therefore, the local values ​​of the state parameters published by the neighbors are observed directly through noise rather than through the observation matrix as described above.

[0261] Processing device 20-i can be assumed (incorrectly) to be in addition to subsets those n in i There are no other communication devices 20 besides +1 communication device 20 to apply the previously proposed Kalman filter-based state observer. Therefore, the Kalman filter used is limited to a subset. The communication device 20 in the process. Therefore, the Kalman filter and the assumed process model are local to the processing device 20-i. Therefore, u i,k It can be modified as follows:

[0262]

[0263] in, It is a subset Aggregation state parameters:

[0264]

[0265] To establish the process model, the update equation of the consensus algorithm is extended to aggregate state parameters.

[0266]

[0267] Ideally, the system should use Closed. In the current case, assume that aggregated state parameters with observation noise are observed.

[0268]

[0269] In this expression, the local values ​​of the state parameters of processing device 20-i are also assumed in a non-restrictive manner to be observed in the presence of observation noise, which represents the noise when a subset of... Other communication devices 20 in the system affect the observation noise of their local values ​​during observation. Therefore, a state observer based on a Kalman filter can be implemented to feed back noisy observation results, but denoising estimation of state parameters can also be achieved.

[0270] As stated above, this equation assumes that in a subset Besides those n i There are no other communication devices 20 besides +1 communication device 20. This equation applies only to the local value x of the state parameter of processing device 20-i. i,k It is accurate because the other local values ​​are actually generated by communication device 20 in subset Θ. i Updated in the middle. Therefore, the introduced error is modeled as process noise:

[0271]

[0272] in, This represents the process noise resulting from the error executed in the update equation of the consensus algorithm. Therefore, the process model described above can be used to implement a Kalman filter-based state observer, as shown below:

[0273]

[0274] and:

[0275]

[0276] For example, process noise can be assumed. and observation noise The unbiased Gaussian random variables have corresponding covariance matrices.

[0277] like Figure 13 As shown, the control method 40 further includes a step S44 where the processing device 20-i determines the convergence level of the local values ​​of the state parameters. The convergence level represents a subset. Whether the local value of communication device 20 is close to the consensus value. A non-restrictive example for calculating the convergence level is provided below.

[0278] like Figure 13 As shown, the control method 40 further includes step S45, wherein the processing device 20 controls the covariance matrix in the Kalman filter based on the estimated convergence level. In fact, although the process model is inaccurate because it ignores cases not in the subset. The communication device 20 in the middle, but it can be assumed that when the subset When the communication devices 20 in the middle are far from reaching a consensus, the resulting error will be higher than the error generated when they are close to reaching a consensus. In other words, if the convergence level indicator (subset) If the local values ​​of the state parameters have not converged, then the covariance matrix... It should have a higher norm than when the convergence level indicates that the local value is nearly convergent. This is achieved by controlling the covariance matrix. As the local values ​​of the state parameters gradually converge to the consensus values, the influence of the process model on the estimation of the aggregated state parameters can be gradually increased.

[0279] For example, in some implementations, the covariance matrix of the process noise is modified. This includes making the norm of the covariance matrix gradually decrease as the local values ​​of the state parameters gradually converge. For example, the covariance matrix can be modified in the following way. The coefficients of the covariance matrix are such that the covariance matrix... The norm (e.g., the Frobenius norm) decreases as the local values ​​of the state parameters converge.

[0280] Alternatively, or in combination, the covariance matrix relative to the observation noise can be used. To modify the covariance matrix of process noise For example, the covariance matrix of process noise Modified to make the covariance matrix of process noise Norm and covariance matrix of observation noise The ratio between the norms gradually decreases as the local values ​​of the state parameters converge. Preferably, the ratio should be less than 1 (<1) when the local values ​​are almost converged, in which case the Kalman filter will benefit the process model rather than the observations used to estimate the aggregate state parameters.

[0281] For example, the covariance matrix of process noise It can be modified as follows:

[0282]

[0283] in, This corresponds to the estimated convergence level, and f corresponds to a predetermined function. Corresponding to the weighting factors, and for example, positive scalar values, in this case, the covariance matrix With covariance matrix Proportional. If we assume a convergence level If the value decreases as consensus approaches, then the function f makes if but In this case, the covariance matrix and The ratio between the Frobenius norms is equal to

[0284] However, the covariance matrix of the process noise can be modified similarly. Without using the covariance matrix of observation noise For example, as follows:

[0285]

[0286] In some implementations, a convergence level is assumed. If the function decreases as it approaches consensus, then the function f can be defined as:

[0287]

[0288] Here, K1 is a predefined positive scalar value, and K2 > 0 is a predefined positive scalar value. For example, K1 = 1 and K2 = 2. However, in other implementations, other definitions of the function f can be used, provided that the function f decreases as the local values ​​of the state parameters converge.

[0289] In some implementations, convergence level The convergence level can be determined based on the number of iterations of the update steps performed from the initial local values ​​of the state parameters. In this case, it is assumed that the consensus algorithm converges with increasing iteration count. This includes obtaining the number of iterations k, and the convergence level is, for example, determined by... Given, it produces in the previous example of function f.

[0290] Alternatively, it can be based on a subset. The convergence level is estimated by using local values ​​of the state parameters of the communication device 20. For example, a consensus value can be estimated, and the distance between the local values ​​of the state parameters and the estimated consensus value can be calculated.

[0291] In some cases, consensus values ​​can be based on (based on) Observation results of aggregation state parameters To estimate. For example, the consensus value estimate μ. i,k It can be estimated as based on (or The average local value is calculated from the local values ​​in the range, and the distance can be calculated as... (or The variance of the local values ​​of (in the middle). Taking into account the observed subset Θ i Local value of communication device 20 in Then we can have:

[0292]

[0293]

[0294] In a preferred embodiment, the convergence level This can be determined by calculating the distance between two multivariate Gaussian random variables, which can be defined as:

[0295]

[0296]

[0297] in:

[0298] - It is based on the local values ​​of the aggregated state parameters observed in successive iterations. The calculated mean;

[0299] - It is based on the local values ​​of the aggregated state parameters observed in successive iterations. Calculate the covariance matrix;

[0300] - Where x i,k The estimate μ used as the consensus value i,k ;

[0301] - This is considering only subset Θ i The covariance matrix of the observed noise in the case of communication equipment 20.

[0302] Therefore, we measure the distance between current observations hypothesized to be Gaussian with known mean and covariance matrices and between current consensus values ​​hypothesized to be measured in observational noise. The mean and covariance matrices are estimated using any suitable means, such as an infinite impulse response (IIR) filter. The distance between the two Gaussian random variables can then be estimated using Bhattacharyya distance.

[0303]

[0304] in

[0305] Alternative expressions for calculating the distance between multivariate Gaussian random variables can be used, for example:

[0306]

[0307] Then, the covariance matrix It can be controlled as described above, for example, controlled as Or be controlled as

[0308] I) Detailed Examples

[0309] Below we provide a non-restrictive example of a detailed implementation of control method 40, assuming in a non-restrictive manner that the convergence level is determined using the Bach distance, and in,

[0310] We use the same notation as described above:

[0311] - in,

[0312] - in,

[0313] - in,

[0314] -Θ i (k) is a subset of communication device 20, and processing device 20-i receives data from communication device 20 at iteration k.

[0315] a) Initialization (iteration k=0)

[0316] Initialization can be performed as follows:

[0317] -x i,0 =x i During step S43, it is sent to other communication devices 20;

[0318] -

[0319] -

[0320] -

[0321] - in It is the covariance of the initial local values ​​of the state parameters of the communication device;

[0322] -β < 1;

[0323] - It is assumed that the covariance matrix of the observed noise is known using "typical" values ​​or by any appropriate means of measurement;

[0324] -α is a control parameter for the consensus algorithm, set to a value that ensures the convergence of the noise-free consensus algorithm with respect to the Laplace matrix of the system. For example, α = 1 / (2deg) max ).

[0325] b) Iteration k≥1

[0326] The observations from previous iterations can be built based on local values ​​received from other communication devices 20 during step S40:

[0327]

[0328] Then, during step S41, the state observer based on the Kalman filter can be updated as follows:

[0329]

[0330]

[0331]

[0332]

[0333]

[0334]

[0335]

[0336] The aggregate state parameter corresponding to the current iteration k The prior estimate, and Corresponding to the aggregation state parameters for the previous iteration k-1 The posterior estimate uses the observations available at iteration k.

[0337] The consensus algorithm update step S42 can be executed as follows:

[0338]

[0339] in or

[0340] Convergence level During step S44, the following is estimated:

[0341]

[0342]

[0343]

[0344]

[0345]

[0346]

[0347] And the covariance matrix of process noise It can be updated during step S45 as follows

[0348]

[0349] During step S43, the updated local value x i,k It was sent to other communication devices 20.

[0350] J) Communication equipment

[0351] Figure 14 An exemplary embodiment of a communication device 20 suitable for implementing control method 40 is illustrated schematically.

[0352] In this exemplary embodiment, the communication device 20 includes processing circuitry 21 and a communication unit 22 coupled to the processing circuitry 21. For example, the processing circuitry 21 includes one or more processors and storage devices (hard disks, solid-state drives, optical disks, electronic storage devices, etc.), wherein computer program products are stored for execution by one or more processors. Alternatively, or in combination thereof, the processing circuitry 21 may include one or more programmable logic circuits (FPGAs, PLDs, etc.) and / or one or more application-specific integrated circuits (ASICs) and / or a set of discrete electronic components, etc. The communication unit 22 includes, for example, transceiver circuitry, which includes components well known to those skilled in the art (antennas, amplifiers, local oscillators, mixers, analog and / or digital filters, etc.). The communication unit 22 is compatible with one or more wired and / or wireless communication protocols. Typically, any communication protocol can be used. In fleet applications (i.e., where the communication device 20 is embedded in individual vehicles of a fleet), one or more communication protocols may include one or more V2V communication protocols, such as those defined by ITS-G5, IEEE WAVE, or the 3GPPCellular-V2X standard.

[0353] K) Other implementation methods

[0354] It should be emphasized that this disclosure is not limited to the exemplary embodiments described above. Variations of the exemplary embodiments described above are also within the scope of this disclosure.

[0355] For example, so far, it has been assumed that the subset Θ i The number of communication devices 20 in the system is fixed at n. i This means the propagation channel is perfect. In practice, propagation channel impairment can lead to packet loss. Control method 40 can be extended to a switching topology by replacing the matrix with equivalents set to 0, which are swapped with entries corresponding to communication links that have suffered packet loss. and The deterministic version. Therefore, processing device 20-i can dynamically adjust the process model at each iteration to account for the absence of subset Θ. i The communication device 20 of (k) receives a data packet containing the updated local value. Similarly, the subset Θ i (k) can change from one iteration to another, and the processing device 20-i can update the process model accordingly. Another solution is to set the corresponding observation noise covariance to a very high value for a specific iteration where the grouping has been lost, so that lost observations replaced by arbitrary values ​​are not considered.

[0356] Furthermore, in some cases, other error sources may be included in the process noise, such as inherent process noise. The global covariance matrix of the process noise can be set as the covariance matrix of the inherent process noise and the covariance matrix of the process noise based on convergence level control. The sum. Furthermore, in some cases, it can be found in the covariance matrix. It includes small diagonal components to address situations where, for example, the estimated convergence level would represent near-consensus that is too fast relative to the actual convergence of the consensus algorithm.

Claims

1. A method for controlling state parameters by communication devices, referred to as processing devices, in a set of communication devices, wherein the processing devices store local values ​​of the state parameters and exchange data with other communication devices to enable all communication devices to converge toward the same value of the state parameters having a consensus value, wherein, A subset of communication devices is associated with the processing device, the subset including the processing device and communication devices from which the processing device receives data, the method comprising iterative steps of the following steps: - Receive local values ​​of the status parameters from the other communication devices associated with the processing device in the subset; - Update the local values ​​of its state parameters based on the local values ​​of the received state parameters, wherein at least some of the received local values ​​are corrupted by observation noise; - Send the local values ​​of its status parameters to other communication devices; The processing device iteratively estimates aggregated state parameters based on the received local values ​​of the state parameters by using a Kalman filter of the applied process model. The aggregated state parameters aggregate the local values ​​of the state parameters of the subset of the communication devices. The Kalman filter uses a covariance matrix of process noise that models the error introduced by the process model. The processing device updates the local values ​​of its state parameters based on the estimated aggregated state parameters. The processing device iteratively determines the convergence level of the local values ​​of the state parameters and controls the covariance matrix of the process noise based on the determined convergence level.

2. The method according to claim 1, wherein, Modifying the covariance matrix of the process noise includes making the norm of the covariance matrix of the process noise decrease as the local values ​​of the state parameters converge.

3. The method according to claim 1, wherein, The process noise covariance matrix is ​​further modified based on the covariance matrix of the observed noise, wherein modifying the process noise covariance matrix includes making the ratio between the norm of the process noise covariance matrix and the norm of the observed noise covariance matrix decrease as the local values ​​of the state parameters converge.

4. The method according to claim 3, wherein, The covariance matrix of the process noise corresponds to the covariance matrix of the observation noise multiplied by a positive weighting factor, and modifying the covariance matrix of the process noise includes reducing the weighting factor as the local values ​​of the state parameters converge.

5. The method according to claim 1, wherein, Determining the convergence level includes estimating the consensus value and calculating the distance between the local value of the state parameter and the estimated consensus value.

6. The method according to claim 2, wherein, Determining the convergence level includes estimating the consensus value and calculating the distance between the local value of the state parameter and the estimated consensus value.

7. The method according to claim 3, wherein, Determining the convergence level includes estimating the consensus value and calculating the distance between the local value of the state parameter and the estimated consensus value.

8. The method according to claim 4, wherein, Determining the convergence level includes estimating the consensus value and calculating the distance between the local value of the state parameter and the estimated consensus value.

9. The method according to any one of claims 1 to 8, wherein, Determining the convergence level at iteration k by the processing device includes calculating the distance between two multivariate Gaussian random variables.

10. The method according to claim 9, wherein, The two multivariate Gaussian variables include: - A random variable having a mean and covariance matrix calculated based on local values ​​of the state parameters observed in successive iterations; and - A random variable representing the estimate of the consensus value that is corrupted by the observation noise.

11. The method according to claim 9, wherein, The calculated distance is the Bach distance.

12. The method according to claim 10, wherein, The calculated distance is the Bach distance.

13. The method according to any one of claims 1 to 3, wherein, The convergence level is determined based on the number of iterations in the update step.

14. The method according to any one of claims 1 to 8, wherein, The processing device dynamically adjusts the process model at each iteration to take into account data packets containing updated local values ​​received from other communication devices not from the subset.

15. The method according to claim 9, wherein, The processing device dynamically adjusts the process model at each iteration to take into account data packets containing updated local values ​​received from other communication devices not from the subset.

16. The method of claim 10, wherein, The processing device dynamically adjusts the process model at each iteration to take into account data packets containing updated local values ​​received from other communication devices not from the subset.

17. The method according to claim 11, wherein, The processing device dynamically adjusts the process model at each iteration to take into account data packets containing updated local values ​​received from other communication devices not from the subset.

18. The method according to claim 12, wherein, The processing device dynamically adjusts the process model at each iteration to take into account data packets containing updated local values ​​received from other communication devices not from the subset.

19. The method according to claim 13, wherein, The processing device dynamically adjusts the process model at each iteration to take into account data packets containing updated local values ​​received from other communication devices not from the subset.

20. The method according to any one of claims 1 to 8, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

21. The method according to claim 9, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

22. The method according to claim 10, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

23. The method according to claim 11, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

24. The method according to claim 12, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

25. The method according to claim 13, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

26. The method according to claim 14, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

27. The method according to claim 15, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

28. The method according to claim 16, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

29. The method according to claim 17, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

30. The method according to claim 18, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

31. The method according to claim 19, wherein, The process model of the Kalman filter assumes that each communication device associated with the processing device in the subset receives data only from the communication devices in the subset.

32. A computer program product comprising instructions that, when executed by a communication device including a communication unit and processing circuitry, configure the communication device to perform the method according to any one of claims 1 to 31.

33. A computer-readable storage medium comprising instructions that, when executed by a communication device including a communication unit and processing circuitry, configure the communication device to perform the method according to any one of claims 1 to 31.

34. A communication device comprising processing circuitry and a communication unit configured to perform the method according to any one of claims 1 to 31.

35. A communication system comprising a plurality of communication devices according to claim 34.