Patent Application
20240403384
This disclosure relates to a data processing device, a control system, a data processing method and a program.
A controllability Gramian is known as an indicator for controllability of a system, see, for example, Non-Patent Document 1-6. Whether a system is controllable can be determined by whether the controllability Gramian is regular. Furthermore, the magnitude of the eigenvalues of the controllability Gramian quantitatively indicates the degree of influence of the input on the state.
A Controllability Gramian can be easily calculated if a mathematical model of a system is known. Hereinafter, calculation methods based on the mathematical model of a system are called “model-based methods”. However, the amount of data necessary for modeling the system is not always available. In such cases, a mathematical model of the system cannot be obtained.
In contrast to the model-based method, calculation methods that calculate based on data of a state trajectory of a system without using a mathematical model are called “data-driven methods”. Data-driven methods have the advantage of requiring fewer decision variables than model-based methods because they do not need to identify a mathematical model of the system.
Methods for estimating controllability Gramians, see, for example, Non-Patent Document 7, and maximizing controllability Gramians, see, for example, Non-Patent Document 8, using data-driven methods are known. Both of these methods are discrete-time models formulated on the basis of discrete time. However, the physical information contained in the data is represented clearly in the continuous-time model, which is formulated on the basis of continuous time, whereas the discrete-time model has a problem of unclarity. In other words, conventional discrete-time models have the problem that it is difficult to utilize prior knowledge about the characteristics of continuous-time systems.
The purpose of this disclosure is to estimate the controllability Gramian for a system of which mathematical model is unknown, using a data-driven method for continuous-time systems.
In order to solve the above problem, a data processing device according to an embodiment of this disclosure estimates the limit:
of a controllability Gramian:
defined by:
in:
when:
holds, where:
is an n-dimensional vector representing the state of a control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix.
This data processing device comprises:
for the following q time intervals:
when:
holds;
expressed as:
based on the sets of state data:
acquired by the data acquisition unit 10, calculates:
and estimates:
by numerically obtaining the solution X of the following linear equation:
with respect to:
and
In one embodiment, the set of time-series state data acquired by the data acquisition unit may include noise and the controllability Gramian calculation unit estimates:
by numerically obtaining the solution X of the following liner equation:
instead of the liner equation:
In one embodiment, the controllability Gramian calculation unit performs numerical calculations using prior knowledge about the signs of some or all of the matrix components of the solution X.
Another embodiment of this disclosure is also a data processing device. This data processing device estimates the matrix B which maximizes the trace:
of the limit:
of a controllability Gramian:
defined by:
in:
when:
holds, where:
is an n-dimensional vector representing the state of the control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix. This data processing device comprises:
for the following q time intervals:
when:
holds;
by numerically obtaining the solution:
of the following linear equation:
when:
holds;
and
In one embodiment, the maximization condition calculation unit may obtain a first input matrix:
by calculating the unit eigenvector corresponding to the maximum eigenvalue of:
when:
holds.
In one embodiment, the maximization condition calculation unit may obtain a second input matrix:
by calculating the n×n matrix in which the (k, k) component is 1 and the other components are 0, where the (k, k) component is the one that is the maximum among the diagonal components of:
when:
holds.
Yet another embodiment of this disclosure is a control system. This control system controls external control objects. This control system comprises:
Yet another embodiment of this disclosure is a data processing method. This method estimates the limit:
of a controllability Gramian:
defined by:
in:
when:
holds, where:
is an n-dimensional vector representing the state of a control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix. This method comprises:
for the following q time intervals:
when:
holds;
expressed as:
based on the sets of state data:
acquired by the data acquisition unit 10, calculating:
and estimating:
by numerically obtaining the solution X of the following linear equation:
with respect to:
and
Yet another embodiment of this disclosure is also a data processing method. This method estimates the matrix B which maximizes the trace:
of the limit:
of a controllability Gramian:
defined by:
in:
when:
holds, where:
is an n-dimensional vector representing the state of the control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix. This method comprises:
for the following q time intervals:
when:
holds;
by numerically obtaining the solution:
of the following linear equation:
when:
holds;
and
Another embodiment of this disclosure is a program. This program estimates the limit:
of a controllability Gramian:
defined by:
in:
when:
holds, where:
is an n-dimensional vector representing the state of a control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix. This program causes the computer to perform the method, comprising:
acquiring a set of time-series state data:
for the following q time intervals:
when:
holds;
defining:
expressed as:
based on the sets of state data:
acquired by the data acquisition unit 10, calculating:
and estimating:
by numerically obtaining the solution X of the following linear equation:
with respect to:
and
outputting the estimated controllability Gramian.
Yet another embodiment of this disclosure is also a program. This program estimates the matrix B which maximizes the trace:
of the limit:
of a controllability Gramian:
defined by:
in:
when:
holds, where:
is an n-dimensional vector representing the state of the control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix. This program causes the computer to perform the method, comprising:
for the following q time intervals:
when:
holds;
by numerically obtaining the solution:
of the following linear equation:
when:
holds;
and
Yet another embodiment of this disclosure is a data processing device. This data processing device estimates:
when:
holds and the limit of a controllability Gramian of a matrix A, which is defined by:
is defined by:
where:
is an n-dimensional vector representing the state of a control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix, is a known n×m matrix and 4 the amount of change in A. This data processing device comprises:
for the following q time intervals:
when:
holds;
expressed as:
based on the sets of state data:
acquired by the data acquisition unit 10, calculates:
and estimates:
by numerically obtaining the solution X of the following linear equation.
with respect to:
and
an output unit that outputs the estimated controllability Gramian.
Any combination of the above components, and any transformation of the expression of the invention among devices, methods, systems, recording media, computer programs, etc., is also valid as a form of the invention.
According to the present disclosure, the controllability Gramian for a system of which mathematical model is unknown can be estimated, using a data-driven method for continuous-time systems.
and
The disclosure will be explained below with reference to drawings based on suitable embodiments. The embodiments are examples rather than limitations of the disclosure. All features or combinations of features described in the embodiments are not necessarily essential to the disclosure. Identical or equivalent components, parts, and processes shown in each drawing shall be given the same symbol, and redundant explanations will be omitted where appropriate. The scale and shape of each part shown in each drawing are set for convenience in order to facilitate explanation, and are not to be construed as limiting unless otherwise noted. When terms such as “first,” “second,” etc. are used in this specification or in the claims, unless otherwise mentioned, these terms do not indicate any order or degree of importance, but are intended only to distinguish one configuration from another. In addition, in each drawing, some parts of the components that are not important in explaining the embodiments are omitted.
Before describing the specific embodiments, let us first explain the basic findings.
The following are definitions of mathematical symbols used herein.
R: real number.
positive real number.
non-negative real number.
n×n symmetric matrix.
n×n diagonal matrix.
0: zero matrix.
I: unit matrix.
the i-th standard basis of the n-dimensional real vector space.
Moore-Penrose pseudo-inverse of matrix A.
the maximum eigenvalue of real symmetric matrix of which all eigenvalues are real numbers.
trace of square matrix A.
Note that for any square matrices A and B:
holds.
Frobenius norm of matrix A, i.e.:
holds.
Furthermore, for an n×n symmetric matrix P of which (i, j) components are:
i.e.:
column vector:
and row vector:
are defined as follows:
For example:
holds for:
respectively.
[Linear Systems and controllability Gramians]
We consider the following linear system.
is the state of the system:
is the control input:
is an n×n constant matrix and:
is an n×m constant matrix.
In this case, the controllability Gramian:
is defined as follows.
holds.
The controllability Gramian represents the set of states that can be reached at time t, starting from the initial state given a control input:
satisfying the following.
is regular when the system is controllable, i.e.:
holds.
The set of states that can be reached in this case is expressed as below.
In particular, the direction that can be reached can be quantified by the limit of controllability Gramians:
in the following.
is the sum of the eigenvalues of matrix below.
is a measure of the strength of controllability of the system.
The following lemma is the basis for the computation of the controllability Gramian, see, for example, Non-Patent Literature 9.
For the system shown in (1), we consider the Lyapunov equation as below.
are the unknown variables in equation (1). If matrix A is a Hurwitz matrix, i.e., the real part of all eigenvalues of matrix A is negative, there exists a unique solution X, which is equal to the following.
Once the mathematical model of (1) is obtained, the controllability Gramian:
is computable and can be optimized by (2) or (3). However, another approach is needed if a mathematical model is not available. One possible approach is to use measured data on the behavior of the system. This is formulated as follows.
We denote the state (1):
by:
with respect to:
under the conditions of:
Also, we denote the segment of the state trajectory in the time interval under the same boundary conditions and inputs as above:
by the following.
Therefore, the problem of estimating and maximizing:
based on given data can be organized as follows.
For the system shown in (1), let matrix A be a Hurwitz matrix and unknown. Let a set of data:
consisting of segments of q state trajectories is given.
holds.
(i) We estimate:
when a known input matrix B is given.
(ii) We obtain the input matrix:
that maximizes:
under the conditions of the following.
is a given pair of possible input matrices.
Note the following four points. First, in Problem 1(i), the input matrix B is assumed to be known. On the other hand, the main objective of Problem 1(ii) is to find the optimal input channel for controlling the system. From this point of view, Problem 1(i) corresponds to a performance analysis of the input matrix B. Namely, in this case, the input matrix B is evaluated to measure the performance of the following.
Second, in general:
increases with the norm of B. Therefore, the condition:
is required for the maximization of:
in Problem 1(ii).
Third, although B can be chosen flexibly, the following two settings are essential.
The former is the case where the inputs to the system are completely arbitrary. The latter is useful when the system is a network of n nodes, with one input added to each node. As for the other settings of B, approximate solutions can be obtained from the above settings. This may be obtained, for example, by truncating less influential part of B, as in the low-rank matrix approximation based on singular value decomposition.
Fourth, it is assumed that the resulting set of data is a continuous-time signal and contains no noise. However, even if this assumption does not hold, the method described herein is still applicable to the above problem. This point is discussed below.
We consider the system (1) with B=0.
In this system, the Lyapunov equation is given by the following.
is a given symmetric matrix and:
is the unknown variable in this equation. The unique solution:
exists if A is a Hurwitz matrix. Note that (5) is different from (3) even though:
holds.
We now consider solving (5) when A is unknown and a set of state trajectory data is given. The problem is formulated as follows.
We assume that the matrix A is a Hurwitz matrix and is unknown with respect to the system shown in (4). We assume that a symmetric matrix:
and a set of data consisting of segments of q state orbits:
is given. Here:
holds. We obtain the solution of (5):
under this condition.
The solution to Problem 2 is given as follows. We consider the state trajectory x in (4) from the initial state x(0).
is obtained by making a quadratic form of x(t) with respect to both sides of (5). Furthermore:
is obtained by integrating both sides of this equation over the time interval as below.
The left side of (7) is equal to:
because:
holds from (4). From the above, (7) can be expressed as below.
The following q equations are obtained by applying the sets of data in Problem 2 to (9).
Here, (10) can be considered to be a linear equation for an unknown matrix Y because:
and the right side are known constants. In the standard form, (10) is expressed as follows.
are unknown variables. Also:
is the matrix given by the following.
is a vector formed by arranging the values on the right side of (10) with respect to the following.
The solution to Problem 2 is thus obtained as follows.
We consider Problem 2. If:
holds, there exists a unique solution Y to (10), which is equal to the solution of Problem 2 (End of Lemma 2).
Lemma 2 shows that the solution to Problem 2 can be obtained by solving the linear equation (10), namely (11), using a set of state orbit data for the system shown in (4).
[Estimation of controllability Gramians].
The solution of Problem 1(i) is described below.
According to the aforementioned Lemma 1, Problem 1(i) comes down to solving the Lyapunov equation (3) by a data-driven method. Below, we obtain the solution of (3) based on data using the same method that yielded (10). Similar to (6), we make the quadratic form of the state x(t) when B=0 on the left side of (3) and (1).
However, in this case, unlike in (8), we cannot replace:
by:
in general. This is because x(t) is a state of (1) and:
holds, in general.
Therefore first, we introduce the following system defined by the transpose matrix of matrix A:
where:
is the state. In this system, we use the symbol:
to express the state:
when:
holds.
Here, we make the quadratic form:
which satisfies the following.
On the other hand:
holds by (3).
Therefore, we obtain:
by (14) and (15). This has the same form as (19).
However, in this case, the set of data for the state trajectory in (13) in Problem 1 is not obtained. This difficulty can be resolved by the following result.
We consider:
in Problem 1. We assume:
holds and choose n arbitrary data from the set of data. Here, n is the dimension of:
in (1).
It can be considered that the chosen data are:
without loss of generality. We assume the following.
[equation 84]
E(t):=[x(t+t11,t11,x11) x(t+t21,t21,x21) . . . x(t+tn1,tn1,xn1)] (17)
is regular:
holds with respect to the following.
holds (End of Lemma 3).
We consider data points:
and state trajectory:
with respect to the following which are given.
The product of:
for any:
is given by the following.
By obtaining this relation for j=1, 2, . . . , n, the following is obtained for t.
This means that:
holds if:
is regular (End of proof).
We can transform the set of data in Problem 1 into the set of data in the state trajectory in (13) by this lemma. Note that the assumption of:
is not restrictive. This is because the number of data can be increased by dividing the state trajectory into multiple segments.
The linear equation:
is obtained by (16), (19) and the following.
This is expressed as:
with respect to:
and the following vector.
As a result, the solution to Problem 1 is obtained as follows.
We consider Problem 1(i). We assume the following.
holds, there exists a unique solution X to (20), which is equal to the solution of Problem 1 (End of Theorem 1).
According to Theorem 1, the solution of Problem 1(i) is given as the solution of the linear equation (20).
In the above discussion, it was assumed that the set of data obtained is a continuous-time signal and contains no noise. However, this is not limited to this, and as shown below, the aforementioned results can also be applied to data containing noise.
We assume that data:
is assumed to contain noise. In order to suppress the effect of noise, it is preferable to use more data to create the matrix:
are obtained with respect to:
Then, from (19) and the properties of the Moore-Penrose pseudo-inverse, the relationship:
is obtained. As a result, an approximate solution to Problem 1(i) is obtained from a variant of (20):
are defined with respect to the p data and the following.
Next, we assume that the sets of data are given as acyclic sample data of state orbitals. The left side of (20) consists of the endpoints of each state trajectory and can be computed from the sample data of the state trajectories. In contrast, the right side of (20) is the integral of the quadratic form of the state orbit. This can be computed approximately from sample data, for example, using the trapezoidal formula. By applying these principles, an approximate solution to the above problem can be obtained.
The solution of Problem 1(ii) is described below.
The following shows the maximization of the controllability Gramian:
for the input matrix B.
For the system shown in (1), we assume that matrix A is a Hurwitz matrix. We consider the following maximization problem:
where:
is a given pair of possible input matrices.
We define Y* as follows.
which is finite positive definite since matrix A is a Hurwitz matrix. The following proposition holds.
(i)
We define the unit eigenvector:
with respect to the following largest eigenvalue.
The matrix:
is the solution of (24) for:
holds.
(ii)
We assume that:
is a diagonal matrix of which (k, k) component is 1 and the other components are 0. Here:
is the row number, also the column number, when the diagonal component of the following matrix.
is the solution of (24) for:
and is
holds. Here:
is the i-th diagonal component of the following.
(i) See, e.g., Non-Patent Document 10.
(ii) We define the i-th diagonal component of matrix B as the following.
is obtained from (2) and the properties of the trace.
is a diagonal matrix. Therefore, from (28):
holds, for any:
which satisfies:
On the other hand, from the definitions of (29) and:
holds (End of proof).
Lemma 4 means that, the input matrix that maximizes:
is characterized by:
in (25). On the other hand:
is equal to the unique solution of the Lyapunov equation (5) for:
Therefore, from Lemma 2 and Lemma 4, the following holds.
We consider Problem 1(ii). We assume that (12) hold with respect to the set of data given in Problem 1. We assume that:
is the solution to:
with respect to:
Furthermore, in Lemma 4, we assume that:
are the input matrices with respect to:
as we defined:
with respect to:
In this case, the following proposition holds.
(i) The input matrix:
is the solution of Problem 1(ii) with respect to the following:
(ii) The input matrix:
is the solution of Problem 1(ii) with respect to the following:
Theorem 2 shows that the solution of Problem 1(ii) can be obtained by solving the linear equation (10) and forming the input matrix by the method of Lemma 4.
of the controllability Gramian:
defined by:
when:
holds, where:
is an n-dimensional vector representing the state of the control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix.
The data acquisition unit 10 acquires a set of time-series state data in a plurality of time intervals. In the following description, the data acquisition unit 10 acquires a set of time-series state data:
for the following q time intervals.
The controllability Gramian calculation unit 12 defines:
expressed as:
based on the sets of state data:
acquired by the data acquisition unit 10, calculates:
and estimates:
by numerically obtaining the solution X of the following linear equation.
The output unit 15 outputs the controllability Gramian estimated by the controllability Gramian calculation unit 12.
Here, the control objects are, for example, the following.
(Example 1) Power system. The time series data include the amount of electricity generated, the amount of electricity used, temperature and humidity, etc. at each power plant.
(Example 2) Robots. Time-series data include load, posture, external force and motion state, etc.
(Example 3) Human body. Time-series data include blood pressure, pulse rate, body temperature and blood composition, etc.
(Example 4) Chemical plant. Time-series data include temperature, humidity, atmospheric pressure and amount of dust in the air, etc.
(Example 5) Equipment to be maintained. Time-series data include failure alarm frequency, electrical resistance, metal fatigue and heat generation, etc.
These control objects are configured as network systems. The mathematical model describing the time-series data output from these systems are complex, and generally such a mathematical model itself is unknown. In contrast, by estimating and outputting controllability Gramians using this implementation, it is possible to determine which node of the control objects configured as network systems should be controlled as the input channel, while the mathematical model is still unknown. The above features are also common to all of the following forms.
According to this embodiment, the controllability Gramian can be estimated using a data-driven method based on a set of acquired continuous-driven state data for a system of which mathematical model is unknown.
The set of time-series state data acquired by the data acquisition unit may include noise. In this case, the controllability Gramian calculation unit 12 estimates:
by computing an approximate solution to Problem 1(i) from:
instead of (20). Here:
are defined with respect to the p data and the following.
According to this embodiment, controllability Gramians can be estimated using a data-driven method even when the state data contains noise.
When the state data contains noise, knowledge about the signs of some or all of the matrix components of the solution X of (23) may be given as prior knowledge. In this case, the controllability Gramian calculation unit 12 performs numerical calculations using such prior knowledge.
According to this method, controllability Gramians can be estimated more accurately using a data-driven method, even when the state data contains noise.
of the limit:
of a controllability Gramian:
defined by:
when:
holds, where:
is an n-dimensional vector representing the state of the control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix.
The data acquisition unit 10 acquires sets of time-series state data in a plurality of time intervals. In the following description, the data acquisition unit 10 acquires a set of time-series state data:
for the following q time intervals.
The maximization condition calculation unit 14 estimates the matrix B which maximizes:
by numerically obtaining the solution:
of the following linear equation:
based on the sets of state data:
acquired by the data acquisition unit 10.
The output unit 15 outputs the input matrix when the controllability Gramian is maximized based on the maximization condition estimated by the maximization condition calculation unit 14.
According to this embodiment, for a system of which mathematical model is unknown, the input matrix when the controllability Gramian is maximized can be estimated using a data-driven method based on a set of acquired continuous-time state data.
In the fifth embodiment, the maximization condition calculation unit 14 obtains a first input matrix:
by calculating the unit eigenvector corresponding to the maximum eigenvalue of:
when:
holds.
According to this embodiment, effective input matrix can be obtained when:
holds.
In the sixth embodiment, the maximization condition calculation unit obtains a second input matrix:
by calculating the n×n matrix in which the (k, k) component is 1 and the other components are 0, where the (k, k) component is the one that is the maximum among the diagonal components of:
when:
holds.
According to this embodiment, effective input matrix can be obtained when:
holds.
The sensor 16 detects a set of time-series state data from the control objects 100. The sensor 16 then transmits the detected set of state data to the data acquisition unit 10 of the data processing device 1. The data processing device 1 estimates:
and transmits it to the control unit 18. The control unit 18 generates control signals based on:
and controls the control objects 100 using them.
According to this embodiment, it is possible to detect time-series state data from an external control object, to estimate the controllability Gramian using a data-driven method based on the detected set of state data and appropriately to control said control object based on the estimated controllability Gramian.
The control input value determination unit 191 is sent the set of state data detected by the sensor 16. Based on this set of state data, the control input value determination unit 191 determines what control should be performed on the control object determined by the input channel determination unit 192. The control input value determination unit 191 sends the determined control contents to the input channel determination unit 192.
The input channel determination unit 192 is sent the estimated controllability Gramian. The input channel determination unit 192 determines which of the control objects 100 should be controlled based on this controllability Gramian. The input channel determination unit 192 selects the determined control object and executes the control determined by the control input value determination unit 191 on the selected control object.
According to this embodiment, it is possible to detect time-series state data from external control objects, to estimate controllability Gramians using a data-driven method based on a set of detected state data, to select an appropriate control object based on the estimated controllability Gramians and appropriately to control the selected control object.
The sensor 16 detects a set of time-series state data from the control objects 100. The sensor 16 transmits the detected set of state data to the data acquisition unit 10 of the data processing device 1. The data processing device 2 estimates the matrix B which maximizes:
and sends it to the control unit 18. The control unit 18 generates control signals based on the matrix B and uses them to control the control objects 100.
According to this embodiment, it is possible to detect time-series state data from external control objects, to estimate the input matrix when the controllability Gramian is maximum using a data-driven method based on the detected set of state data, and appropriately to control the control objects based on the estimated input matrix.
of the controllability Gramian:
defined by:
when:
holds, where:
is an n-dimensional vector representing the state of the control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix.
In step S1, the method obtains the time-series state data set:
in the q time intervals:
when:
holds.
In step S2, the method defines:
expressed as:
calculates:
and estimates:
with respect to:
by numerically obtaining the solution X of the following linear equation.
In step S3, the method outputs the controllability Gramian estimated in step S2.
According to this embodiment, the controllability Gramian can be estimated by a computer using a data-driven method based on a set of acquired continuous-time state data for a system of which mathematical model is unknown.
This method estimates the matrix B which maximizes the trace:
of the limit:
of the controllability Gramian:
defined by:
when:
holds, where:
is an n-dimensional vector representing the state of the control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix.
In step S1, the method obtains the time-series state data set:
in the q time intervals:
when:
holds.
In step S4, the method estimates the matrix B which maximizes:
by numerically obtaining the solution:
of the following linear equation:
when:
holds.
In step S5, the method outputs the input matrix when the controllability Gramian is maximized based on the maximization condition estimated in step S4.
According to this embodiment, the input matrix when the controllability Gramian is maximized can be estimated by a computer using a data-driven method based on a set of acquired continuous-time state data for a system of which mathematical model is unknown.
The eleventh embodiment is a program. This program estimates the limit:
of the controllability Gramian:
defined by:
in:
when:
holds, where:
is an n-dimensional vector representing the state of the control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix. This program make a computer perform
the step of obtaining the time-series state data set:
in the q time intervals:
when:
holds,
the step of defining:
expressed as:
calculating:
and estimating:
by numerically obtaining the solution X of the following linear equation:
with respect to:
and
the step of outputting the estimated controllability Gramian.
According to this embodiment, software can be implemented as a program to estimate the controllability Gramian using a data-driven method based on a set of acquired continuous-time state data for a system of which the mathematical model is unknown.
The twelfth embodiment is a program. This program estimates the matrix B which maximizes the trace:
of the limit:
of the controllability Gramian:
defined by:
in:
when:
holds, where:
is an n-dimensional vector representing the state of the control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix.
This program make a computer perform
the step of obtaining the time-series state data set:
in the q time intervals:
when:
holds,
the step of estimating the matrix B which maximizes:
by numerically obtaining the solution:
of the following linear equation:
when:
holds and
the step of outputting the input matrix when the controllability Gramian is maximized based on the estimated maximization condition.
According to this embodiment, software can be implemented as a program to estimate the input matrix when the controllability Gramian is maximized by a computer using a data-driven method based on a set of acquired continuous-time state data for a system of which mathematical model is unknown.
The following shows the results by the first embodiment with respect to a set of noiseless time-series state data are presented. We obtain the solution of Problem 1(i) in the system shown in (1) when:
holds. Here, we assume that the following hold.
For comparison, the true values of controllability Gramians:
which is obtained using (3) and “lyap” of Matlab™, are as follows.
The data acquisition unit 10 of the data processing device 1 in
in the six time intervals shown in
are determined by (17) and (18), respectively. Then, the following is obtained.
This means the following.
Therefore, (22) holds and there exists a unique solution X to the linear equation (22).
The controllability Gramian calculation unit 12 solves (22) numerically to obtain the following values.
This is consistent with the true value (30), indicating the usefulness of this embodiment.
The following shows the results by the second embodiment with respect to a set of time-series state data including noise.
The signals in
The controllability Gramian calculation unit 12 solves (23) numerically to obtain the following values.
This result is largely consistent with the true value (30), although it is worse than (31).
The following is the actual result obtained using the third embodiment with respect to sets of time-series state data including noise. In this case, the sets of the time-series state data are also shown in
Here, we assume that we have prior knowledge that:
has the following pattern of sign.
Here, * indicates that the sign is unknown. Then, the problem of estimating:
boils down to the following optimization problem.
are related to (21) and used in (23).
corresponds to:
of (21).
is the i-th component of:
is obtained by numerically computing (33). This approximate solution applying the above prior knowledge is closer to the true value (30) than the approximate solution (32).
The following describes the results obtained from the fourth embodiment using the data in
is obtained. (12) holds since:
holds. Namely, there exists a unique solution Y for (10) with respect to:
Numerical calculation in the embodiment yields the following.
The following is the results obtained from the fifth embodiment using the data in
is obtained with respect to:
by the numerical calculation according to the embodiment when:
holds.
takes the maximum value of:
by this first input matrix. In contrast to this result:
takes the value of:
when the input matrix:
with all components ⅓ is used as an example for comparison. This is clearly smaller than the value:
obtained in the embodiment.
The following are the results obtained from the sixth embodiment using the data in
is obtained with respect to:
by the numerical calculation according to the embodiment when:
holds.
The values of:
in each case of:
are shown in Table 1. From this, it is understood that:
is the most effective input matrix, indicating the usefulness of this embodiment.
In the embodiments described above, one of the goals was to estimate the “ease of control” index when the input channel to each node of the network is fixed (Problem 1(i)). In other words, the following problems were discussed.
For the system shown in Equation (1), we assume that the matrix A is a Hurwitz matrix and is unknown. We assume that a set of data:
consisting of segments of q state trajectories is given.
Here:
holds.
(i):
is estimated when a known input matrix B:
is given.
In other words, in the embodiments explained above, it was assumed that the amount of action between each node constituting the network, i.e., the strength of the connection between nodes and an indicator of the amount of transmitted data, are all fixed values. In this case, it was possible to know how to select input channels to improve ease of control. In contrast, the following describes the estimation of ease of control when the amount of action is adjusted, i.e., estimation of controllability Gramian when the amount of action changes. The goal is to know how to adjust the action amount in order to improve the ease of control.
In the following, we consider the estimation of the controllability Gramian when a change A is added to an unknown matrix A. This problem can be formulated as follows.
We consider the system in Equation (1). Here, we assume that:
is known, while:
is unknown. Also, we assume that the data of N state trajectories are given to the system as follows:
holds. Also, we assume that:
such that:
is a stable matrix is arbitrarily given. At this point, we obtain:
Problem 3 is to estimate the value of the controllability Gramian from the data when the A matrix is varied by A for the system in Equation (1). This problem appears, for example, when applying state feedback to the system or when adjusting the connection strength between nodes in a network system.
In the following, we describe a method for solving the Lyapunov equation using the state trajectories of the system as a preparation for Problem 3. We consider the following system:
Here:
is a stability matrix. Also, the state trajectory of the system in equation (35) can be expressed in the same way as described above as:
For this system, we consider the following Lyapunov equation.
Here:
is a matrix such that:
is stable, and:
is a symmetric matrix. In this case, it is known that equation (36) has a unique solution.
At this point, we consider the following problem (Problem 4)
We consider the system in Equation (35). Here:
is assumed to be stable but unknown. For this system, we assume that the data D of N state trajectories are given as in Problem 3. Also, we assume that:
such that:
is a stable matrix are arbitrarily given. In this case, we obtain the solution of Equation (36).
The solution to Problem 4 is given as follows. First, for both sides of equation (36), we consider the quadratic form with respect to the state trajectory:
then:
holds. Here:
is obtained since:
holds for the first term on the left side.
is obtained by integrating both sides of this equation over the time interval:
Here, N linear equations for the unknown number Y:
is obtained by applying the state trajectory data:
to equation (39).
Next, we consider rewriting equation (40) in a matrix-based form. First, equation (39) can be expressed as:
using lvec and rvec. If this is noted, Equation (40) can be rewritten as:
equivalently. Here:
are matrices determined by the state trajectory data of the system in Equation (35) and their i-th row:
are given by:
respectively.
The solution to Problem 4 is then obtained as follows.
We consider Problem 4. If:
holds, the solution of equation (40) is unique in the range of symmetric matrices and is equal to the solution of Problem 4. (End of Theorem 3).
Next, we consider the solution to Problem 3.
can be computed using Theorem 3 if state trajectory data of:
which is a dual system of Equation (35) are available. This proposition can be expressed as follows. From Lemma 1, the controllability:
is obtained as a unique solution of where equation of:
Similarly to the above discussion:
is obtained if we note that equation (43) corresponds to equation (36) to which:
are applied. Thus, by applying the state trajectory of the system in equation (13) to equation (44), a linear equation with the solution of:
can be constructed from the state trajectory data only.
However, in Problem 3, the state trajectory of the system in Equation (13) is not generally available except in the case that
holds. Therefore, the above ideas cannot be used directly for Problem 3. On the other hand, if the state trajectory of the system in Equation (35) can be transformed into the state trajectory of the system in Equation (13), a solution to Problem 3 can be constructed. Such a data transformation is accomplished by the following Lemma 5.
We express the state trajectory of the system of Equation (13) by:
in the same way as aforementioned
We assume that the state trajectory data D in Problem 3 is arbitrarily given. Here, we assume that N≥n. We assume that, for this D:
are defined by:
respectively. In this case, if E is regular:
holds for arbitrary:
We give the solution to problem 3 using the above results. First, we assume that the function:
is defined by the following.
Then N linear equations:
are obtained by applying:
to equation (44). Furthermore:
is obtained as an equivalent expression to equation (47). is obtained. Here:
are matrices determined by the state trajectory data D in Problem 3 and their i-th row:
are given by:
respectively.
Thus, the solution to Problem 3 can be obtained as follows.
We consider Problem 3. Here, we assume that N≥n. If E0 in Equation (45) is regular and:
holds, the solution of Equation (47) is unique in the range of symmetric matrices and is equal to the solution of Problem 3.
As in the first embodiment,
The data processing device 1 estimates:
when:
holds and the limit of controllability Gramian of matrix A, which is defined by:
is defined by:
where:
is an n-dimensional vector representing the state of the control object:
is an m-dimensional vector representing the control input, A is an unknown n×n matrix and B is a known n×m matrix.
The data acquisition unit 10 acquires a set of time-series state data in a plurality of time intervals. In the following description, the data acquisition unit 10 acquires a set of time-series state data:
for the following q time intervals.
The controllability Gramian calculation unit 12 defines:
expressed as:
based on the sets of state data:
acquired by the data acquisition unit 10, calculates:
and estimates:
by numerically obtaining the solution X of the following linear equation.
The output unit 15 outputs the controllability Gramian estimated by the controllability Gramian calculation unit 12.
According to this embodiment, the controllability Gramian can be estimated when the amount of action between each node constituting the network changes.
The present disclosure has been described above based on the embodiments. It is understood by those skilled in the art that these embodiments are examples, that various variations are possible in the combination of each component and each processing process and that such variations are also within the scope of the disclosure.
Any combination of the above mentioned embodiments and variations is also useful as an embodiment of the disclosure. The new embodiment resulting from the combination will have the respective effects of each of the embodiments and variations that are combined.
The principle of the disclosure can be applied to the control of systems in various fields as follows.
(Application example 1) When the control object is an electric power system, optimal power supply control can be performed based on time-series data of power generation, power consumption, temperature and humidity, etc. at each power plant.
(Application example 2) When the control object is a robot, optimal posture control can be performed based on time-series data such as load, posture, external force and motion state, etc.
(Application example 3) When the control object is a human body, optimal medication control can be performed based on time-series data such as blood pressure, pulse rate, body temperature and blood composition, etc.
(Application example 4) When the control object is a chemical plant, the system can perform optimal operation control based on time-series data such as temperature, humidity, air pressure and the amount of dust in the air, etc.
(Application example 5) When the control object is equipment maintenance, optimal maintenance control can be performed based on time-series data such as failure alarm frequency, electrical resistance, metal fatigue and heat generation, etc.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2021-143668 | Sep 2021 | JP | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2022/033059 | 9/2/2022 | WO |
