Tag Archives: System Identification

Rank Considerations for the Observer-Kalman System Identification Procedure

Reading Time: 4 minutes

Last time, we looked in detail at the derivation of the OKID procedure for finding the impulse response of a system from arbitrary input-output data. However, there are some specificities to consider when collecting data, and we can derive these by looking at the rank of the matrices involved.

In this article, we’ll reconsider how the quality and uniqueness of the solution for an ordinary least-squares problem is affected by the ranks of the matrices involved. Based on that, we’ll derive a formula for determining the number of measurements we need to find a good estimate of the Markov parameter matrix.

The Result

Let’s jump to the result real quick here and then look at how it is derived. Assume that we want to determine the Markov parameters

(1)   \begin{equation*} \mathbf{M} =   \begin{bmatrix}     \mathbf{C} \mathbf{A}^{l-1} \mathbf{B} &     \ldots     \mathbf{C} \mathbf{B} &     \mathbf{D}   \end{bmatrix} \end{equation*}

of order l for a system with m inputs and p outputs.

Schematic of a Multiple-Input, Multiple-Output (MIMO) System

The number of measurement points required is given by

(2)   \begin{equation*}   n = om + \left[o\left(m+p\right)+1\right] l \end{equation*}

Here, o is the oversampling factor, which gives the number of data points we have per entry in the Markov parameter matrix. We need this oversampling to average out measurement error due to noise, and thus the larger the oversampling factor, the higher the quality of our estimate for the Markov parameter matrix.

On the other hand, a higher oversampling factor also means that we can either

  • increase the number of measurements while keeping the order of our Markov parameter matrix constant, or
  • decrease the order of our Markov parameter matrix while keeping the number of measurements.

A lower order of the Markov parameter matrix may decrease the quality of the estimate we get from the Eigensystem Realisation Algorithm. A higher number of measurements possibly increases the measurement effort and time.

Rank Considerations for OKID

Let’s review the central equation of the OKID with observer:

(3)   \begin{equation*}   \underbrace{\begin{bmatrix}     \mathbf{y}_{l} &     \mathbf{y}_{l+1} &     \cdots &     \mathbf{y}_{n-1}   \end{bmatrix}}_{=:\mathbf{Y}} \\   \approx \\   \mathbf{\tilde{M}}_l   \underbrace{\begin{bmatrix}     \mathbf{u}_0 & \cdots &\mathbf{u}_{n-l-1} \\     \mathbf{y}_0 & \cdots &\mathbf{y}_{n-l-1} \\     \vdots & \ddots & \vdots \\     \mathbf{u}_{l-1} & \cdots &\mathbf{u}_{n-2} \\     \mathbf{y}_{l-1} & \cdots &\mathbf{y}_{n-2} \\     \mathbf{u}_{l} & \cdots &\mathbf{u}_{n}   \end{bmatrix}}_{=:\mathbf{\tilde{U}}} \end{equation*}

Now have a look at the dimensions of these matrices:

  • \mathbf{Y} \in \mathbb{R}^{p \times \left(n-l\right)}
  • \mathbf{\tilde{M}} \in \mathbb{R}^{p \times \left[m+\left(m+p\right)l\right]}
  • \mathbf{\tilde{U}} \in \mathbb{R}^{\left[m+\left(m+p\right)l\right] \times \left(n-l\right)}

Thus, we have p\left[m+\left(m+p\right)l\right] unknowns and p\left(n-l\right) equations. Looking back at basic linear algebra, thus we know that we need

(4)   \begin{equation*} n-l = m+\left(m+p\right)l \end{equation*}

to hold for the solution to be uniquely determined. We might even need more than that, as this assumes that the measurement data is rich enough so that Equation 3 does not contain linearly dependent rows. However, in any case, the number of our measurements must be at least as large as given by Equation 4.

What happens if we have less measurements than this? Well, in this case the equation is not sufficiently determined, and there are arbitrarily many solutions \mathbf{\tilde{M}} to the equation. The ordinary least-squares approach will deliver a solution, but it is not clear whether that solution accurately describes the system we want to identify — although it will describe a system that will provide the same output given the inputs.

Eliminating Noise by Oversampling

However, if we follow Equation 4 exactly, i.e. if we have exactly as many measurements as specified by this equation, we will also have an exact solution. This means that we will exactly incorporate all the measurement noise into our Markov parameters. Usually, we do not want that. Instead, we want to average out that measurement noise by having multiple samples.

That means that we will need more than the number of measurements given by Equation 4 — a lot more. To quantify this number, we’ll consider an oversampling factor o. This factor gives us the number of samples per parameter we want to have. It also gives the factor by which we can diminish the influence of the measurement noise onto our estimate.

Now, if we want to oversample each parameter by factor o, we need o equations for each parameter. Thus, the following equation must hold

(5)   \begin{equation*} n-l = o\left[m+\left(m+p\right)l\right] \end{equation*}

Solving that equation for n, we get our final result, given in Equation 1.

Impact of Using an Observer

Clearly, this is the result when we use an observer for identifying the Markov parameters. If we have a system which already is sufficiently stable, we do not need the observer approach. In that case, we have a much simpler equation:

(6)   \begin{equation*}   \underbrace{\begin{bmatrix}     \mathbf{y}_{l} &     \mathbf{y}_{l+1} &     \cdots &     \mathbf{y}_{n-1}   \end{bmatrix}}_{=:\mathbf{Y}} \\   \approx \\   \mathbf{M}_l   \underbrace{\begin{bmatrix}     \mathbf{u}_0 & \cdots &\mathbf{u}_{n-l-1} \\     \vdots & \ddots & \vdots \\     \mathbf{u}_{l} & \cdots &\mathbf{u}_{n}   \end{bmatrix}}_{=:\mathbf{U}} \end{equation*}

Here we have m\left(l+1\right) rows in \mathbf{U}, so that the following relationship must hold for an oversampling factor o:

(7)   \begin{equation*} n = o m \left(l+1\right) + l \end{equation*}

This is smaller than the number required for the observer approach, as we do not have to determine many less entries of the Markov parameters. Thus, we may safe some measurement effort if we have a stable system.


It is quite important that we have a sufficient number of measurements

  • to get a uniquely determined set of Markov parameters, and
  • to properly average out our measurement noise.

We can get the number of measurements we need by rank considerations of our basic equations and a simple oversampling approach. If we have a sufficiently stable system, we can avoid using the observer approach and thus reduce the number of measurements required while keeping the order of the Markov parameter set constant.

The Observer/Kalman System Identification Procedure Explained

Reading Time: 10 minutes

Previously, we have had a look at the Eigensystem Realisation Algorithm (ERA) by Joung and Pappa [1]. This algorithm allows us to find the parameters for a linear, time-invariant (LTI) system in discrete-time from measurements of the impulse response.

What, however, do we do when we don’t have measurements of the impulse response? We might not want to hit our system too hard, or have measurements with non-trivial inputs. Maybe we cannot wait for our system to return to a zero state before applying the impulse.

Again, Professor Steve Brunton came to the rescue with two videos on the Observer/Kalman System Identification procedure (or OKID for short) in his series on data-driven control:

Aktivieren Sie JavaScript um das Video zu sehen.

Unfortunately, Prof. Brunton skips over the details and only gives a very short overview. Yet, I always try to understand the details of the methods I apply — if not for being able to assess when they may be used, then to learn about the principles being applied.

So I had a look into the papers by a group around Jer-Nan Juang, Minh Phan, Lucas G. Horta and Richard W. Longman, in which the OKID was introduced [2][3][4]. And again, there are some clever ideas in there that I found quite enlightening regarding handling the descriptions of dynamic systems:

  • The basic approach of the OKID uses the fact that the response of an LTI to any input is related to the output by convolution of the impulse response.
  • By reshaping the measurement data, a linear equation between said impulse response and the measurement data is established.
  • This linear equation is solved using an ordinary least-squares approach.

The basic approach only works for asymptotically stable systems which approach zero state fast enough. However, there is a extension for non-stable or too slow systems: Instead of identifying the system itself, the OKID uses a modified system which is made asymptotically stable by a Luenberger observer [5], and then reconstructs the impulse response of the original system from the modified system.

Finally, this observer extension also leads to the “Kalman” part in the name for this procedure: The observer that is identified “on the go” this way turns out to be an optimal Kalman filter [6] given the amount of noise seen on the measurements.

So, all in all, the whole approach is quite ingenious. However, it’s not as complicated as Professor Brunton makes it sound, and we’ll briefly go over it and its basic ideas in this article.

Review: Impulse Response

As for the ERA, we are dealing with discrete-time LTI systems. But different from previously, we’ll directly consider a Multiple Input, Multiple Output (MIMO) system.

Schematic of a Multiple-Input, Multiple-Output (MIMO) System

The dynamics of this system shall be defined by the following recurrence equation:

(1)   \begin{align*}   \mathbf{x}_{k+1} &=     \mathbf{A} \mathbf{x}_k     + \mathbf{B} \mathbf{u}_k \\   \label{eq:system-output}   \mathbf{y}_k &=     \mathbf{C} \mathbf{x}_k     + \mathbf{D} \mathbf{u}_k \end{align*}

Note that this time we also consider the case where the output is directly influenced by the output via the matrix \mathbf{D}. Now, let’s assume that we have an input sequence \mathbf{u}_k and observe the system starting at time step k_0. We can determine the state \mathbf{x}_{k_0+r} (with r \geq 0) to be

(2)   \begin{equation*} \mathbf{x}_{k_0+r} = \mathbf{A}^r \mathbf{x}_{k_0} + \sum_{j=1}^r \mathbf{A}^{r-j} \mathbf{B} \mathbf{u}_{k_0+j-1} \end{equation*}

This can be verified by complete induction. For the case r=0 we get \mathbf{x}_{k_0} = \mathbf{A}^0 \mathbf{x}_{k_0} = \mathbf{x}_{k_0}.

For r \geq 0 we get

(3)   \begin{align*} \mathbf{x}_{k_0+r+1} &= \mathbf{A} \mathbf{x}_{k_0+r} + \mathbf{B} \mathbf{u}_{k_0+r} \\ &= \mathbf{A} \mathbf{A}^r \mathbf{x}_{k_0} + \mathbf{A} \sum_{j=1}^r \mathbf{A}^{r-j} \mathbf{B} \mathbf{u}_{k_0+j-1} + \mathbf{B} \mathbf{u}_{k_0+r} \\ &= \mathbf{A}^{r+1} \mathbf{x}_{k_0} + \sum_{j=1}^r \mathbf{A}^{r+1-j} \mathbf{B} \mathbf{u}_{k_0+j-1} + \mathbf{B} \mathbf{u}_{k_0+r} \\ &= \mathbf{A}^{r+1} \mathbf{x}_{k_0} + \sum_{j=1}^{r+1} \mathbf{A}^{r+1-j} \mathbf{B} \mathbf{u}_{k_0+j-1} \end{align*}

By extension, we can give the following expression for the output \mathbf{y}_{k_0+r}:

(4)   \begin{align*} \mathbf{y}_{k_0+r}   &= \mathbf{C} \mathbf{A}^r \mathbf{x}_{k_0}    + \sum_{j=1}^r \mathbf{C} \mathbf{A}^{r-j} \mathbf{B} \mathbf{u}_{k_0+j-1}    + \mathbf{D} \mathbf{u}_{k_0} \\   &= \mathbf{C} \mathbf{A}^r \mathbf{x}_{k_0} +      \underbrace{\begin{bmatrix}        \mathbf{C} \mathbf{A}^{r-1} \mathbf{B} &        \cdots &        \mathbf{C} \mathbf{B} &        \mathbf{D}      \end{bmatrix}}_{=:\mathbf{M}_r}      \bar{\mathbf{u}}^{\left(r\right)}_{k_0} \end{align*}

Here, we define the column vector \bar{\mathbf{u}}^{\left(r\right)}_{k_0} by stacking the inputs at time steps k_0,\ldots,k_0+r above each other:

(5)   \begin{equation*} \bar{\mathbf{u}}^{\left(r\right)}_{k_0} =   \begin{bmatrix}     \mathbf{u}_{k_0} \\     \vdots \\     \mathbf{u}_{k_0+r}   \end{bmatrix} \end{equation*}

The matrix \mathbf{M}_r gives us the so-called Markov-Parameters of length r, and when reversed, gives the impulse response of the system. If we had these Markov-Parameters, we could extract the system matrices with the help of the Eigensystem Realisation Algorithm (ERA).

Handling Asymptotically Stable Systems

For an asymptotically stable system, the term \mathbf{A}^r approaches zero for increasing values of r. Thus, for some sufficiently large p we may assume that \mathbf{A}^p \approx \mathbf{0}.

This helps us in simplifying Equation 4

(6)   \begin{align*} \mathbf{y}_{k_0+p}   &= \mathbf{C} \underbrace{\mathbf{A}^p}_{\approx \mathbf{0}} \mathbf{x}_{k_0} +      \mathbf{M}_p      \bar{\mathbf{u}}^{\left(p\right)}_{k_0} \\   &\approx \mathbf{M}_p      \bar{\mathbf{u}}^{\left(p\right)}_{k_0} \end{align*}

Now, we can aggregate this into a larger equation:

(7)   \begin{equation*}   \begin{bmatrix}     \mathbf{y}_{k_0+p} &     \mathbf{y}_{k_0+p+1} &     \cdots &     \mathbf{y}_{k_0+n-1}   \end{bmatrix} \\   \approx \\   \mathbf{M}_p   \begin{bmatrix}     \bar{\mathbf{u}}^{\left(p\right)}_{k_0} &     \bar{\mathbf{u}}^{\left(p\right)}_{k_0+1} &     \cdots &     \bar{\mathbf{u}}^{\left(p\right)}_{k_0+n-p}   \end{bmatrix} \end{equation*}

With n measurements (ideally, we have n \gg p) we can thus find \mathbf{M} by solving the equation above. Usually, this will be done using an ordinary least-squarey approach.

Handling Insufficient Stability

In the previous section, we assumed that the system we measure is asymptotically stable and is sufficiently damped so that we can have our measurement count n sufficiently large relative to the length of our impulse response p. So how do we handle systems where this is not the case?

Stabilisation Using a Luenberger Observer

The idea presented by Juang, Phan, Horta and Longman is quite simple: To arbitrarily choose the eigenvalues of the observed system, they construct a Luenberger observer. In general application, such an observer aims to reconstruct the internal state of a system from knowledge about the internal dynamics, about the inputs to the system and the measurements obtained from the system. It has the following structure:

(8)   \begin{align*}   \hat{\mathbf{x}}_{k+1} &= \mathbf{A} \hat{\mathbf{x}}_k + \mathbf{B} \mathbf{u}_k - \mathbf{K} \left(\mathbf{y}_k-\hat{\mathbf{y}}_k\right) \\   \hat{\mathbf{y}}_k &= \mathbf{C} \hat{\mathbf{x}}_k + \mathbf{D} \mathbf{u}_k \end{align*}

In this case, \hat{\mathbf{x}}_k and \hat{\mathbf{y}}_k represent the state of the observer, which according to the construction of the observer shall follow the state of the original system. The observer state is adjusted using the correction term -\mathbf{K} \left(\mathbf{y}_k-\hat{\mathbf{y}}_k\right).

Usually, when constructing a Luenberger observer, we choose \mathbf{K} given our knowledge about the original system in such a way that the observation error \mathbf{e}_k := \hat{\mathbf{x}}_k-\mathbf{x}_k is asymptotically stable, i.e. approaches zero. The dynamics of this error are described as follows:

(9)   \begin{equation*}   \mathbf{e}_{k+1} = \left(\mathbf{A} + \mathbf{K} \mathbf{C}\right) \mathbf{e}_k \end{equation*}

Given a completely observable system, we can select \mathbf{K} in such a way that the eigenvalues of these error dynamics can be set arbitrarily. However, these eigenvalues also become the eigenvalues of the observer system, allowing us to arbitrarily stabilise it.

In our case, we want the observer state to be equal to the actual state, so we set \hat{\mathbf{x}}_k=\mathbf{x}_k, and arrive at the new state equations:

(10)   \begin{align*}      \mathbf{x}_{k+1} &=     \underbrace{      \left(\mathbf{A}+\mathbf{K} \mathbf{C}\right)}_{=:\mathbf{\hat{A}}}     \mathbf{x}_k    +    \underbrace{     \begin{bmatrix}      \mathbf{B}+\mathbf{K} \mathbf{D} & -\mathbf{K}     \end{bmatrix}}_{=:\mathbf{\hat{B}}}    \underbrace{     \begin{bmatrix}      \mathbf{u}_k \\      \mathbf{y}_k     \end{bmatrix}}_{=:\mathbf{\hat{u}}} \\   \mathbf{y}_k &=      \mathbf{C} \mathbf{x}_k      +    \underbrace{     \begin{bmatrix}      \mathbf{D} & \mathbf{0}     \end{bmatrix}}_{=:\mathbf{\hat{D}}}    \mathbf{\hat{u}} \end{align*}

These equations describe the exact same system as in Equations 1 and ??, but with the current output of the system \mathbf{y}_l fed back into the system in such a way that the observed state always equals the actual system state. With this little trick, we have created a stable system to identify.

To find the Markov-Parameters \mathbf{\hat{M}} of the new system, we simply apply the algorithm we used for sufficiently stable systems. This time, we use \mathbf{\hat{u}}_l as input, which combines both the actual input and our measurement.

However, there’s one thing we need to be careful about: \mathbf{\hat{D}} is not completely unconstrained in this problem. Indeed, its second part must be \mathbf{0}. If we were to allow non-zero values there, the best fit would be \mathbf{\hat{D}} = \begin{bmatrix} \mathbf{0} & \mathbf{I} \end{bmatrix}, with all other system matrices being zero. This system would just forward the input \mathbf{y}_k into its output. This would of course be a perfect fit — but it would not properly represent our system.

So instead, we have to find the best fit for

(11)   \begin{equation*}     \begin{bmatrix}     \mathbf{y}_{k_0+p} &     \mathbf{y}_{k_0+p+1} &     \cdots &     \mathbf{y}_{k_0+n-1}   \end{bmatrix} \\   \approx \\   \mathbf{\tilde{M}}_p   \begin{bmatrix}     \bar{\mathbf{\hat{u}}}^{\left(p-1\right)}_{k_0} &     \bar{\mathbf{\hat{u}}}^{\left(p-1\right)}_{k_0+1} &     \cdots &     \bar{\mathbf{\hat{u}}}^{\left(p\right)}_{k_0+n-p} \\     {\mathbf{u}}_{k_0+p} &     {\mathbf{u}}_{k_0+p+1} &     \cdots &     {\mathbf{u}}_{k_0+n}   \end{bmatrix} \end{equation*}

Note how all rows consist of pairs of input and output, but the last row consists only of the input? This way, the identified system cannot feed through the measured output to its output, and instead the equation identifies the dynamics of the system we want to identify.

The result, however, does not yet represent the Markov-Parameters of our observer. Instead, it has the form

(12)   \begin{equation*}   \mathbf{\tilde{M}}   =   \begin{bmatrix}     \mathbf{C} \mathbf{\hat{A}}^{r-1} \mathbf{\hat{B}}     &     \cdots     &     \mathbf{C} \mathbf{\hat{B}}     &     \mathbf{D}   \end{bmatrix} \end{equation*}

For recovery of the Markov-Parameters of the original system, this is quite inconsequential (we just have to be careful in the implementation). If we wanted to actually have the Markov-Parameters of the observer — which we might, as we will see when we look at the specific properties of that observer — then we would have to restructure the response as follows by appending an appropriately sized zero-matrix at the end:

(13)   \begin{equation*}   \mathbf{\hat{M}}   =   \begin{bmatrix}     \mathbf{C} \mathbf{\hat{A}}^{r-1} \mathbf{\hat{B}}     &     \cdots     &     \mathbf{C} \mathbf{\hat{B}}     &     \mathbf{D} &     \mathbf{0}   \end{bmatrix} \end{equation*}

Recovery of Original Markov-Parameters

But how do we get the Markov-Parameters of the original system? Again, Juang et al provide us with a little trick. Let’s first look at the structure of our stabilized Markov-Parameters:

(14)   \begin{align*}   \mathbf{\hat{M}}     &=       \begin{bmatrix}        \mathbf{C} \mathbf{\hat{A}}^{r-1} \mathbf{\hat{B}} &        \cdots &        \mathbf{C} \mathbf{\hat{B}} &        \mathbf{\hat{D}}       \end{bmatrix} \\     &=       \begin{bmatrix}        \mathbf{\hat{M}}_{r} & \ldots & \mathbf{\hat{M}}_1 & \mathbf{\hat{M}}_{0}       \end{bmatrix} \end{align*}


(15)   \begin{align*}   \mathbf{\hat{M}}_{0} &=     \begin{bmatrix}       \mathbf{D} &       \mathbf{0}     \end{bmatrix} \\   \mathbf{\hat{M}}_r &=     \mathbf{C}     \left(\mathbf{A}+\mathbf{K} \mathbf{C}\right)^{r-1}     \begin{bmatrix}        \mathbf{B}+\mathbf{K} \mathbf{D} &        -\mathbf{K}     \end{bmatrix} \\     &=:       \begin{bmatrix}         \mathbf{\hat{M}}^{\left(1\right)}_r &         \mathbf{\hat{M}}^{\left(2\right)}_r       \end{bmatrix} \end{align*}

Now, it is quite straightforward to extract \mathbf{D} from \mathbf{\hat{M}}_0. For the others, we find that

(16)   \begin{equation*}   \mathbf{\hat{M}}^{\left(1\right)}_r     + \mathbf{\hat{M}}^{\left(2\right)}_r \mathbf{D}   =   \mathbf{C}   \left(\mathbf{A}+\mathbf{K} \mathbf{C}\right)^{r-1} \mathbf{B} \end{equation*}

For r=1, we can immediately recover \mathbf{M}_1:

(17)   \begin{equation*}   \mathbf{M}_1 := \mathbf{C} \mathbf{B}   =     \mathbf{\hat{M}}^{\left(1\right)}_1     + \mathbf{\hat{M}}^{\left(2\right)}_1 \mathbf{D} \end{equation*}

For r>1 we first expand the last element of the power:

(18)   \begin{align*}   \mathbf{\hat{M}}^{\left(1\right)}_r     + \mathbf{\hat{M}}^{\left(2\right)}_r \mathbf{D}   &=     \mathbf{C}     \left(\mathbf{A}+\mathbf{K} \mathbf{C}\right)^{r-1}     \mathbf{B} \\   &=      \mathbf{C}     \left(\mathbf{A}+\mathbf{K} \mathbf{C}\right)^{r-2}     \mathbf{A}     \mathbf{B} \nonumber\\     &+     \underbrace{       \mathbf{C}       \left(\mathbf{A}+\mathbf{K} \mathbf{C}\right)^{r-2}       \mathbf{K}     }_{=-\mathbf{\hat{M}}^{\left(2\right)}_{r-1}}     \underbrace{       \mathbf{C} \mathbf{B}     }_{=\mathbf{M}_1} \end{align*}

Looking further, we can identify a pattern:

(19)   \begin{equation*}      \mathbf{M}_r   =     \mathbf{\hat{M}}^{\left(1\right)}_r     + \sum_{k=-1}^{r-1}          \mathbf{\hat{M}}^{\left(2\right)}_{r-k-1}          \mathbf{M}_k \end{equation*}

The reader is encouraged to prove this identity by complete induction over \mathbf{r}. Hint: First prove for s,t \geq 0 the following, more general identity:

(20)   \begin{equation*}   \mathbf{C} \left(\mathbf{A} + \mathbf{K} \mathbf{C}\right)^s \mathbf{A}^t \mathbf{B}   = \\   \mathbf{C} \mathbf{A}^{s+t} \mathbf{B}   + \sum_{k=0}^{s-1} \mathbf{C} \left(\mathbf{A} + \mathbf{K} \mathbf{C}\right)^k \mathbf{K} \mathbf{C} \mathbf{A}^{s+t-k-1} \mathbf{B} \end{equation*}

With a few more steps Equation 19 follows for t=0.

Thus, the Markov-Parameters \mathbf{M}_r:=\mathbf{C}\mathbf{A}^{r-1}\mathbf{B} can be easily retrieved from Equation 19.

Why Kalman?

Now we know how the term “observer” got into the name. But how did “Kalman” end up there?

When we described the observer that we would use as model, we actually described a very specific observer, namely one that would immediately provide the same output as the actual system. The authors of the OKID paper [2] use the term “deadbeat observer” for this. This is the fastest possible observer for a discrete time system. Any faster, and the observer would predict the output of the system, violating causality.

Indeed, later in the paper, the authors show that this deadbeat observer is identical with the Kalman filter one would find given the variance of noise seen in the measurement. A more detailed explanation of their proof might follow in a later article.

Thus, extracting not only the original Markov Parameters, but also the parameter \mathbf{K} for the observer may be very beneficial if one wanted to design a Kalman filter for the system.

How do we get the system parameters for this Kalman filter? Well, we already have its Markov parameters, given by \mathbf{\hat{M}}, and can use the ERA on them. That gives us dynamics matrices for a Kalman filter.


Let’s apply this method to our spring pendulum example from the ERA article. Remember: We had a horizontal spring pendulum with the following setup:

Spring Pendulum

Different from the case with ERA, we now use an arbitrary input signal (which is the force F applied to the mass), and determine the output signal. Again, we simulate the system and add noise to the output. The result is shown in the following figure.

Measurement of Pendulum Output (red) with Random Input (blue)

The input is a pseudorandom binary sequence (PRBS). Such sequences can be generated, e.g., using the max_len_seq function from the SciPy signal processing library.

As we know from the ERA experiments, our pendulum is quite stable, so we can use the method for asymptotically stable systems. The result is shown in the following figure:

Comparison of impulse responses – Actual vs. Response estimated by OKID

It seems that the estimate fits the actual impulse response (this time without noise) quite well. However, we’re in a dilemma:

  • If we increase the length of our estimated impulse response, we reduce the amount of data we have available to counter all the noise, and thereby also the quality of our estimate.
  • If we decrease the length of our estimated impulse response, we get a better estimate, but we have less data for our ERA procedure.

So we need to find a good compromise there by playing with the length of our estimated impulse response.


We have seen that the idea of the OKID procedure is not as involved as one might believe. The basis are three very simple ideas:

  • The Markov Parameters determine the impulse response, and thereby define the input-output relationship — together with the initial state.
  • For asymptotically stable systems, the influence of the initial state diminishes over time, eventually becoming so small that we can ignore it.
  • Not sufficiently asymptotically stable systems can be made stable by considering a Luenberger observer in their place. The problem is then reduced to identifying the Markov Parameters for this observer, and retrieving the Markov Parameters for the original system from this intermediate result.
  • The observer such identified actually is a Kalman filter for the system, given the noise seen in the measurement.

It seems useful to always use the observer-enhanced procedure by default. Even for sufficiently asymptotically stable systems the use of the observer would reduce p and thus the amount of measurements required.

Now, finally, this opens up a lot of new possibilities for system identification: We can use the response to any set of inputs to determine the system model. This also allows us to identify a system that already is subject to control and where it would be impractical or impossible to remove that control. One simple example are flight-tests with an already controlled multicopter to improve the controller, but another would be to identify the progression of a disease without having to stop treatment (which of course would be unethical).

To take the design of a multicopter as an example, we could simply record in-flight data from sensors as well as the control inputs sent to our engines, and use that to improve our already existing model. We do not even have to provide specific inputs, but just a general recording of flight data will help, as long as it was rich enough.

[1] [doi] J. Juang and R. S. Pappa, “An Eigensystem Realization Algorithm for Modal Parameter Identification and Model Reduction,” Journal of guidance control and dynamics, vol. 8, iss. 5, 1985.
author = {Juang, Jer-Nan and Pappa, Richard S.},
title = {{A}n {E}igensystem {R}ealization {A}lgorithm for {M}odal {P}arameter {I}dentification and {M}odel {R}eduction},
journal = {Journal of Guidance Control and Dynamics},
year = {1985},
volume = {8},
number = {5},
month = nov,
abstract = {A method called the eigensystem realization algorithm is developed for modal parameter identification and model reduction of dynamic systems from test data. A new approach is introduced in conjunction with the singular-value decomposition technique to derive the basic formulation of minimum order realization which is an extended version of the Ho-Kalman algorithm. The basic formulation is then transformed into modal space for modal parameter identification. Two accuracy indicators are developed to quantitatively identify the system and noise modes. For illustration of the algorithm, an example is shown using experimental data from the Galileo spacecraft.},
doi = {10.2514/3.20031},
file = {:Juang.Pappa1985 - An Eigensystem Realization Algorithm for Modal Parameter Identification and Model Reduction.pdf:PDF},
[2] [doi] J. Juang, M. Phan, L. G. Horta, and R. W. Longman, “Identification of Observer/Kalman Filter Markov Parameters: Theory and Experiments,” , Hampton, Virginia 23665, techreport 104069, 1991.
author = {Jer-Nan Juang and Minh Phan and Lucas G. Horta and Richard W. Longman},
title = {{Identification of Observer/Kalman Filter Markov Parameters: Theory and Experiments}},
year = {1991},
type = {techreport},
number = {104069},
address = {Hampton, Virginia 23665},
month = jun,
abstract = {An algorithm to compute Markov parameters of an observer or Kalman filter from experimental input and output data is discussed. The Markov parameters can then be used for identification of a state space representation, with associated Kalman gain or observer gain, for the purpose of controller design. The algorithm is a non-recursive matrix version of two recursive algorithms developed in previous works for different purposes. The relationship between these other algorithms is developed. The new matrix formulation here gives insight into the existence and uniqueness of solutions of certain equations and gives bounds on the proper choice of observer order. It is shown that if one uses data containing noise, and seeks the fastest possible deterministic observer, the deadbeat observer, one instead obtains the Kalman filter, which is the fastest possible observer in the stochastic environment. Results are demonstrated in numerical studies and in experiments on an ten-bay truss structure.},
doi = {10.2514/6.1991-2735},
file = {:Juang.Phan.ea1991 - Identification of Observer_Kalman Filter Markov Parameters_ Theory and Experiments.pdf:PDF},
owner = {ralfg},
school = {NASA Langley Research Center},
timestamp = {2019-09-11},
url = {https://ntrs.nasa.gov/search.jsp?R=19910016123},
[3] [doi] M. Phan, J. N. Juang, and R. W. Longman, “Identification of Linear Multivariable Systems by Identification of Observers with Assigned Real Eigenvalues,” Journal of the astronautical sciences, vol. 40, iss. 2, p. 261–279, 1992.
author = {M. Phan and J.N. Juang and R.W. Longman},
title = {{Identification of Linear Multivariable Systems by Identification of Observers with Assigned Real Eigenvalues}},
journal = {Journal of the Astronautical Sciences},
year = {1992},
volume = {40},
number = {2},
pages = {261--279},
month = jun,
doi = {https://doi.org/10.2514/6.1991-949},
file = {:Phan.Juang.ea1992 - Identification of Linear Multivariable Systems by Identification of Observers with Assigned Real Eigenvalues.pdf:PDF},
[4] [doi] M. Phan, L. G. Horta, J. N. Juang, and R. W. Longman, “Linear system identification via an asymptotically stable observer,” Journal of optimization theory and applications, vol. 79, iss. 1, p. 59–86, 1993.
author = {M. Phan and L.G. Horta and J.N. Juang and R.W. Longman},
title = {{Linear system identification via an asymptotically stable observer}},
journal = {Journal of Optimization Theory and Applications},
year = {1993},
volume = {79},
number = {1},
pages = {59--86},
month = oct,
issn = {0022-3239},
abstract = {This paper presents a formulation for identification of linear multivariable systems from single or multiple sets of input-output data. The system input-output relationship is expressed in terms of an observer, which is made asymptotically stable by an embedded eigenvalue assignment procedure. The prescribed eigenvalues for the observer may be real, complex, mixed real and complex, or zero corresponding to a deadbeat observer. In this formulation, the Markov parameters of the observer are first identified from input-output data. The Markov parameters of the actual system are then recovered from those of the observer and used to realize a state space model of the system. The basic mathematical formulation is derived, and numerical examples are presented to illustrate the proposed method.},
doi = {10.1007/BF00941887},
file = {:Phan.Horta.ea1993 - Linear System Identification Via an Asymptotically Stable Observer.pdf:PDF},
owner = {ralfg},
timestamp = {2019-09-11},
[5] [doi] D. Luenberger, “Observing the State of a Linear System,” Ieee transactions on military electronics, vol. MIL8, p. 74–80, 1964.
author = {Luenberger, David},
title = {{Observing the State of a Linear System}},
journal = {IEEE Transactions on Military Electronics},
year = {1964},
volume = {MIL8},
pages = {74--80},
month = may,
abstract = {In much of modern control theory designs are based on the assumption that the state vector of the system to be controlled is available for measurement. In many practical situations only a few output quantities are available. Application of theories which assume that the state vector is known is severely limited in these cases. In this paper it is shown that the state vector of a linear system can be reconstructed from observations of the system inputs and outputs. It is shown that the observer, which reconstructs the state vector, is itself a linear system whose complexity decreases as the number of output quantities available increases. The observer may be incorporated in the control of a system which does not have its state vector available for measurement. The observer supplies the state vector, but at the expense of adding poles to the over-all system.},
doi = {10.1109/TME.1964.4323124},
file = {:Luenberger1964 - Observing the State of a Linear System.pdf:PDF},
[6] [doi] R. E. Kalman, “A New Approach to Linear Filtering and Prediction Problems,” Transactions of the asme – journal of basic engineering, vol. 82, p. 35–45, 1960.
author = {Kalman, R. E.},
title = {{A New Approach to Linear Filtering and Prediction Problems}},
journal = {Transactions of the ASME - Journal of Basic Engineering},
year = {1960},
volume = {82},
pages = {35--45},
doi = {10.1115/1.3662552},
file = {:Kalman1960 - A New Approach to Linear Filtering and Prediction Problems.pdf:PDF},
owner = {ralfg},
timestamp = {2015.08.09},

A deeper look into the Eigensystem Realisation Algorithm

Reading Time: 12 minutes

For my quadcopter control project, I am currently in the process of system identification, trying to find a model describing the dynamics of my motor-propeller-combination. Finding the theoretical model is quite simple, but then you’re left with finding the parameters — and not all of them are directly available from the motor vendors. So I have no other choice than to determine these parameters from measurements.

In my endeavour I came across this series of videos on data-driven control by machine learning by Professor Steve Brunton of the University of Washington, Seattle. There, basic principles of modern methods for system identification are explained very well. If you are new to control systems, I can also recommend his control boot camp.

In part of the series, Professor Brunton explains the Eigensystem Realisation Algorithm (ERA) by Juang and Pappa [1]. The ERA is a procedure by which a reduced-order model for a linear and time-invariant system (LTI) can be derived from data of an impulse-response measurement.

Aktivieren Sie JavaScript um das Video zu sehen.
Aktivieren Sie JavaScript um das Video zu sehen.
Aktivieren Sie JavaScript um das Video zu sehen.

However, while the videos do a pretty solid job of explaining the concepts and give the final formulae, I wanted to know more about their derivation. For this, I went to the original paper where the approaches were developed, and tried to reproduce the steps that led to the final results.

I follow this approach quite a lot when acquiring new skills. It helps not only better understanding — and memorizing — the final solution, but I also tend to pick up neat tricks along the way that often prove helpful elsewhere — even completely outside the domain of the original paper.

Here’s some interesting ideas you may pick up by reading this article:

  • When doing system identification, we mostly work with discrete-time systems, and in some aspects they are much easier to handle than continuous-time systems.
  • The discrete-time impulse response, which uniquely characterises the behaviour of LTI systems, can be easily expressed in closed form using the system matrices.
  • By reorganising our data, we can express seemingly non-linear problems in linear form.
  • The singular value decomposition allows us to examine the composition of a matrix and simplify it.
  • There are a few tricks for transforming matrix power expressions that may reduce the order of our problem.

Our Example

Juang and Pappa worked on large, flexible space structure such as the Galileo spacecraft (which should not be mixed up with the European Galileo satellite navigation system), and aimed to identify the oscillation modes of such structures, so that these could be taken into account in control of the spacecraft.

We’ll look at something similar, although much more simple: A spring-loaded, dampened horizontal pendulum. This pendulum consists of a mass m connected to a spring and a linear dampener. The mass can move horizontally, and we measure the position x of its centre of gravity, with x=0 being the position when the system is at rest.

Spring Pendulum

This system can be modelled as a Single-Input, Single-Output system: We have a single input u=F and a single output y=x.

Schematic of a Single-Input, Single-Output (SISO) System

Normally, we could easily provide at least the structure of a model for this system from first principles — namely, the conservation of momentum:

(1)   \begin{equation*}   m \ddot{x} + d \dot{x} + k x = F \end{equation*}

Introducing state variables x_1 = x and x_2 = \dot{x}, we can define the dynamics of this system in state space notation:

(2)   \begin{align*}   \frac{d}{dt} \begin{bmatrix} x_1 \\ x_2 \end{bmatrix}     &=     \begin{bmatrix}       0 & 1 \\       -\frac{k}{m} & -\frac{d}{m}     \end{bmatrix}     & \begin{bmatrix} x_1 \\ x_2 \end{bmatrix}     &+ \begin{bmatrix} 0 \\ \frac{1}{m} \end{bmatrix} F \\   y     &=     \begin{bmatrix} 1 & 0 \end{bmatrix}     & \begin{bmatrix} x_1 \\ x_2 \end{bmatrix} \end{align*}

Instead of actually measuring our system, we will be using the continuous-time dynamics to simulate the system and thereby perform “measurements”. We will add white noise to emulate measurement noise.

Impulse-Response Measurement

Now, what happens if we whack this system with a hammer? What if we give it a short kick and then measure the position of the mass? Let’s find out!

Measured Impulse Response of a Dampened Spring-Mass System

The figure shows the position of the mass over time. We see that the system oscillates around the neutral position, but the amplitude of the oscillation falls exponentially with time, until measurement noise exceeds the movement of the system. In this example, we whacked the mass with a short force of 1 Newton, and then let go. So, our input would look like this:

Discrete-Time Impulse Input – Spring Pendulum

This function is called the Kronecker Delta Function, and when used in analysis of signals and systems, it is referred to as the (discrete-time) unit impulse. It is one for exactly one tick of our (discrete) time-base and zero everywhere else. When given as an input to a discrete-time system, the output observed afterwards is called the discrete-time impulse response. Specifically, LTI systems are completely characterised by their impulse response.

As we are working on data that we would acquire by measurement, considering the continuous-time formulation is not useful, as we cannot get continuous-time measurement data. Instead, we’ll have some kind of quantisation in both time and value. This actually simplifies the maths of our problem a bit, as we will see in a moment.

We now want to have a look at how we can mathematically and generally describe the impulse response. This will be the basis of our model to which we will then fit our measurement data. In discrete time, we describe the dynamics of an LTI system using a recurrence relation in state space:

(3)   \begin{equation*}   \mathbf{x}_{k+1} = \mathbf{A} \mathbf{x}_k + \mathbf{b} u_k \end{equation*}

Let us assume that for our initial state \mathbf{x}\left(0\right)=0 holds. This is certainly true for dampened structures if we wait long enough for all oscillations to die down before we whack the structure. Now, if our input is 1 for time step k=0 and 0 for all other time steps, we can see the progression of our state by repeatedly applying Equation 3:

Time Step kState \mathbf{x}\left(k\right)
1\mathbf{A} \underbrace{\mathbf{x}\left(0\right)}_{=0} + \mathbf{b} \underbrace{u\left(0\right)}_{=1}=\mathbf{b}
2 \mathbf{A} \underbrace{\mathbf{x}\left(1\right)}_{=\mathbf{b}} + \mathbf{b} \underbrace{u\left(1\right)}_{=0} = \mathbf{A}\mathbf{b}
3 \mathbf{A} \underbrace{\mathbf{x}\left(2\right)}_{=\mathbf{A} \mathbf{b}} + \mathbf{b} \underbrace{u\left(2\right)}_{=0} = \mathbf{A}^2\mathbf{b}
n \mathbf{A}^{n-1}\mathbf{b}

So, in general we have \mathbf{x}_k = \mathbf{A}^{k-1} \mathbf{b} for k>0, and with the output relation

(4)   \begin{equation*}   y_k = \mathbf{c}^T \mathbf{x}_k \end{equation*}

we get

(5)   \begin{equation*}   y_k = \mathbf{c}^T \mathbf{A}^{k-1} \mathbf{b} \end{equation*}

Note that this impulse-response is completely independent of what our state variables \mathbf{x}_k represent. You can try it yourself: Substitute \mathbf{x}_k = \mathbf{T} \tilde{\mathbf{x}}_k with some invertible square matrix \mathbf{T} and determine the impulse response of that system. However, this actually works in our favour, as it allows us to arbitrarily select at least part of the parameters later on.

In general, the terms \mathbf{c}^T \mathbf{A}^{k-1} \mathbf{b} — or, in their Multiple-Input, Multiple-Output form \mathbf{C} \mathbf{A}^{k-1} \mathbf{B} — are called the Markov Parameters of the system.

Finding an Exact Model of Order n

What Juang and Pappa aim to do is to find parameter matrices \mathbf{A}, \mathbf{c}^T and \mathbf{b} so that the impulse response so described fits the measured data exactly. One way of doing so would be to explicitly write down the parametrised impulse-response — as we did in the last chapter — and try to solve this for the parameters directly.

However, that approach is rather ugly, and looking at Equation 5, this is a non-linear problem due to the powers of \mathbf{A} occurring. Instead, Juang and Pappa use quite a nice trick: They restructure the data in such a way that the problem becomes linear. This allows us to use our toolkit from linear algebra to find a solution.

In a first step, Juang and Pappa construct the vectors \bar{\mathbf{y}}_k, which are the column vectors of n observations starting at time step k:

(6)   \begin{equation*}   \bar{\mathbf{y}}_k := \begin{bmatrix} y_k \\ y_{k+1} \\ \vdots \\ y_{k+n-1} \end{bmatrix} \end{equation*}

We will see later that n is the order of the system we will build. For the time being, we shall assume that we have an arbitrary amount of measurements, so there are no limits on n. From Equation 5, we can see that

(7)   \begin{equation*}   \bar{\mathbf{y}}_k = \mathcal{O} \mathbf{A}^{k-1} \mathbf{b} \end{equation*}

with the observability matrix

(8)   \begin{equation*}   \mathcal{O} = \begin{bmatrix} \mathbf{c}^T \\ \mathbf{c}^T  \mathbf{A} \\ \mathbf{c}^T  \mathbf{A}^2 \\ \vdots \\  \mathbf{c}^T  \mathbf{A} ^{n-1} \end{bmatrix} \end{equation*}

Now, remember the function of the observability matrix: If the state has dimension n and the observability matrix has at least rank n, then we can use it to uniquely reconstruct the state from the outputs at time steps k,\ldots,k+n-1. Thus, if our system is observable, we can use the inverse of the observability matrix \mathcal{O} to find the state \mathbf{x}_k:

(9)   \begin{equation*}  \mathcal{O}^{-1} \bar{\mathbf{y}}_k = \mathbf{A}^{k-1} \mathbf{b} = \mathbf{x}_k \end{equation*}

However, from that we can construct any state at any time due to:

(10)   \begin{equation*}   \mathbf{x}_{k+1}=\mathbf{A} \mathbf{x}_k =  \mathbf{A}\mathcal{O}^{-1} \bar{\mathbf{y}}_k \end{equation*}

Applying this to y_k we get the recurrence

(11)   \begin{equation*}   \bar{\mathbf{y}}_{k+1} = \mathcal{O} \mathbf{x}_{k+1} = \mathcal{O} \mathbf{A}\mathcal{O}^{-1} \bar{\mathbf{y}}_k \end{equation*}

Now, let’s look at what we have done: We are able to describe the output at time steps k+1,\ldots,k+n from the output at time steps k,\ldots,k+n-1 using a linear operator! How do we find \mathcal{O} \mathbf{A}\mathcal{O}^{-1}? Equation 11 does not sufficiently specify their value. To increase the rank of the equation system, Joung and Pappa proceed to build the Hankel-matrices

(12)   \begin{align*}   \mathbf{H}_k   &=     \begin{bmatrix}       \bar{\mathbf{y}}_k & \bar{\mathbf{y}}_{k+1} & \cdots &\bar{\mathbf{y}}_{k+n-1}     \end{bmatrix} \\   &=     \begin{bmatrix}       y_k & y_{k+1} & \cdots & y_{k+n-1} \\       y_{k+1} & y_2 & \cdots & y_{n+1} \\       \vdots & \vdots & \ddots & \vdots \\       y_{k+n-1} & y_{k+n} & \cdots & y_{k+2n-2}     \end{bmatrix} \end{align*}

These matrices are constructed by listing n measurements starting at time step k in the first column, then listing n measurements starting at time step k+1 in the second column, and so on. Using Equation 11, we can find the following recurrence for \mathbf{H}_{k+1}:

(13)   \begin{equation*}      \mathbf{H}_{k+1} = \mathcal{O} \mathbf{A} \mathcal{O}^{-1} \mathbf{H}_k \end{equation*}

Thus, if we have any matrices \mathbf{H}_k and \mathbf{H}_{k+1}, with the former being a regular, invertible matrix, we can find

(14)   \begin{equation*}   \mathcal{O} \mathbf{A}^{-1} \mathcal{O}^{-1} = \mathbf{H}_{k+1} {\mathbf{H}_k}^{-1} \end{equation*}

Thus, we can express y_k using

(15)   \begin{align*}   y_k     &= \mathbf{e}^T \mathbf{H}_k \mathbf{e} \\     &= \mathbf{e}^T \left(\mathcal{O} \mathbf{A}^{-1} \mathcal{O}^{-1}\right)^{k-1} \mathbf{H}_1 \mathbf{e} \\     &=        \underbrace{\mathbf{e}^T}_{=:\hat{\mathbf{c}}^T}       {\underbrace{\left(\mathbf{H}_{2} {\mathbf{H}_1}^{-1}\right)}_{=:\hat{\mathbf{A}}}}^{k-1}       \underbrace{\mathbf{H}_1 \mathbf{e}}_{=:\hat{\mathbf{b}}} \\     &:= \hat{\mathbf{c}}^T \hat{\mathbf{A}}^{k-1} \hat{\mathbf{b}} \end{align*}

where \mathbf{e}^T=\begin{bmatrix} 1 & 0 & \ldots & 0 \end{bmatrix}. This exactly describes our measured system response:

Comparison of Impulse Responses: Measured vs. Estimated from Full Order System

The Problem of Noise

However, we also see that our estimated system exactly fits all the noise from our measurement. In our case we know from first-principles considerations that the exact model would only be of order n=2, while the model we developed here is of much larger order. Besides being quite a misuse of resources, this may also lead to considerable estimation errors, as we can see when plotting the step responses:

Comparison of Step Responses: Measured vs. Estimated from Full Order System

Our estimated system clearly deviates from the system we measured. This is no wonder, as the estimated system incorporates all the noise we measured.

Further, our Hankel-matrices are pretty badly conditioned. This means that small rounding errors during calculations may become large errors in the result. We know that — in the absence of measurement error — the rank of the Hankel-matrices cannot be larger than the order of the underlying system. If they are indeed regular, they only are due to measurement noise, which should ideally be much smaller in magnitude than the actual data. Thus, simple inversion is prone to large numerical errors.

In their original paper, Juang and Pappa describe a method of using only a subset of the measurements to reduce the size of the Hankel-matrices, and thereby to improve the conditioning of the matrices. However, today, the method of ERA is almost universally presented (e.g. on Wikipedia) as being based on Singular Value Decomposition (SVD), using the whole set of measurement data.

Singular Value Decomposition

Any matrix \mathbf{M} \in \mathbb{R}^{m \times n} can be represented in the form of its Singular Value Decomposition (SVD):

(16)   \begin{equation*}   \mathbf{M} = \mathbf{P} \mathbf{D} \mathbf{Q}^{T} \end{equation*}

The matrices all have special forms:

  • The matrices \mathbf{P} \in \mathbb{R}^{m \times m} and \mathbf{Q} \in \mathbb{R}^{n \times n} are square, orthogonal matrices, i.e. \mathbf{P}\mathbf{P}^{T}=\mathbf{I}_m and \mathbf{Q}\mathbf{Q}^{T}=\mathbf{I}_n. Note that the identity matrices may have different dimensions if \mathbf{M} is not square.
  • The matrix \mathbf{D} \in \mathbb{R}^{m \times n} is a diagonal matrix, with all diagonal elements being non-negative. As a common convention, the diagonal elements \sigma_k of \mathbf{D} — called the singular values of \mathbf{M} — are listed in decreasing order.

Essentially, the SVD expresses the contents of the matrix as the sum of matrices with decreasing amount of impact:

(17)   \begin{equation*}   \mathbf{M}_{ij} = \sum_{r=1}^{\min\left\{m,n\right\}} \sigma_r \mathbf{P}_{ir} \mathbf{Q}^T_{rj} \end{equation*}

As the matrices \mathbf{P} and \mathbf{Q} are orthogonal, the impact of the individual element is represented by \sigma_r. If we look at the Pareto plot of the singular values, we see that there are some dominant elements in there:

Pareto Plot of Singular Values of the Measurement Hankel Matrix

The plot shows the individual singular values in orange, ordered in decreasing order from left to right, and the cumulative proportion of the singular values, added up from left to right. We can see that the first two singular values are the most prominent, making up for about 55% of the total data values, and the remaining singular values are pretty small in comparison to that — although they still add up to representing 45% of the data. We also see that our matrix is awfully close to being singular, which it would be if any of the singular values were zero.

This is somewhat consistent with our original considerations, where we found out that our system is actually of second order. Thus, we may assume that the data represented by the major two singular values is our actual data, while the remaining singular values represent noise.

Now, what do we do with that information? We can use it to remove what we consider to be noise from our data. In essence, the vectors of \mathbf{P} and \mathbf{Q} for bases of vector spaces, and the singular values determine the magnitude by which the base vectors are represented in the data in the matrix. Thus, by removing some singular values and base vectors, we can project the data in the matrix onto the directions represented by the remaining base vectors, and thereby approximate the original matrix:

(18)   \begin{align*}   \mathbf{M}     &= \mathbf{P} \mathbf{D} \mathbf{Q}^{T} \\     &=       \begin{bmatrix} \tilde{\mathbf{P}} & \mathbf{P}_R \end{bmatrix}       \begin{bmatrix}         \tilde{\mathbf{D}} & \mathbf{0} \\         \mathbf{0} & \mathbf{D}_R       \end{bmatrix}       \begin{bmatrix}          \tilde{\mathbf{Q}}^T \\          {\mathbf{Q}_R}^T       \end{bmatrix} \\     &\approx \tilde{\mathbf{P}} \tilde{\mathbf{D}} \tilde{\mathbf{Q}}^{T} \\     &=: \tilde{\mathbf{M}} \end{align*}

By appropriately selecting \tilde{\mathbf{D}}, we can approximate \mathbf{M} arbitrarily close. Usually, one would determine the size of \tilde{\mathbf{D}} from first-principles considerations, or by defining a cut-off value in relative error.

With that, we can determine the pseudo-inverse of \tilde{\mathbf{M}}:

(19)   \begin{equation*}   \tilde{\mathbf{M}}^{*} = \tilde{\mathbf{Q}} \tilde{\mathbf{D}}^{-1} \tilde{\mathbf{P}}^{T} \end{equation*}

I encourage you to personally verify that indeed \tilde{\mathbf{M}} \tilde{\mathbf{M}}^{*} = \tilde{\mathbf{M}}^{*} \tilde{\mathbf{M}} = \mathbf{I}_{n \times m} holds, where \mathbf{I}_{n \times m} is the n \times m identity matrix (ones on the diagonal, zeroes everywhere else).

Simplified Model

With the simplified Hankel-Matrix

(20)   \begin{equation*}   \tilde{\mathbf{H}}_1 = \tilde{\mathbf{P}} \tilde{\mathbf{D}} \tilde{\mathbf{Q}}^{T} \end{equation*}

and its pseudo-inverse {\tilde{\mathbf{H}}_1}^{*}, we can build a simplified, cleaned-up model:

(21)   \begin{align*}   \tilde{y}_k     &=        \underbrace{\mathbf{e}^T}_{=:\tilde{\mathbf{c}}^T}       {\underbrace{\left(\mathbf{H}_{2} {\tilde{\mathbf{H}}_1}^{*}\right)}_{=:\tilde{\mathbf{A}}}}^{k-1}       \underbrace{\tilde{\mathbf{H}}_1 \mathbf{e}}_{=:\tilde{\mathbf{b}}} \\     &:= \tilde{\mathbf{c}}^T \tilde{\mathbf{A}}^{k-1} \tilde{\mathbf{b}} \end{align*}

Determining the impulse response of that model, we get a pretty clean fit, and it seems that the noise is actually being ignored.

Comparison of Impulse Responses: Measured vs. Estimated from Simplified, Full Order System

Also, our step response fits much better:

Comparison of Step Responses: Measured vs. Estimated from Full Order, Simplified System

Reduced-Order Model

But still, our matrices \tilde{\mathbf{c}}^T, \tilde{\mathbf{A}} and \tilde{\mathbf{b}} are of order \mathbf{n}, while they should be of second order for our system. To reduce that order, we’ll have to make use of some trickery. We’ll re-use Equation 21 and add an identity matrix in there:

(22)   \begin{align*}   \tilde{y}_k   &= \tilde{\mathbf{c}}^T \tilde{\mathbf{A}}^{k-1} \tilde{\mathbf{b}} \\   &= \underbrace{\tilde{\mathbf{c}}^T}_{=:\mathbf{e}^T} \underbrace{\mathbf{I}_n}_{=\tilde{\mathbf{H}}_1 {\tilde{\mathbf{H}}_1}^{*}} {\underbrace{\tilde{\mathbf{A}}}_{=\mathbf{H}_{2} {\tilde{\mathbf{H}}_1}^{*}}}^{k-1} \underbrace{\tilde{\mathbf{b}}}_{=:\tilde{\mathbf{H}}_1 \mathbf{e}} \\   &= \mathbf{e}^T \tilde{\mathbf{H}}_1 {\tilde{\mathbf{H}}_1}^{*} \left(\mathbf{H}_{2} {\tilde{\mathbf{H}}_1}^{*} \right)^{k-1} \tilde{\mathbf{H}}_1 \mathbf{e} \\  &= \mathbf{e}^T \tilde{\mathbf{P}} \tilde{\mathbf{D}} \tilde{\mathbf{Q}}^{T} \tilde{\mathbf{Q}} \tilde{\mathbf{D}}^{-1} \tilde{\mathbf{P}}^{T} \left(\mathbf{H}_{2} \tilde{\mathbf{Q}} \tilde{\mathbf{D}}^{-1} \tilde{\mathbf{P}}^{T} \right)^{k-1} \tilde{\mathbf{P}} \tilde{\mathbf{D}} \tilde{\mathbf{Q}}^{T} \mathbf{e} \end{align*}

Keep in mind that the values in \tilde{\mathbf{D}} are non-negative and all non-zero values are on the diagonal, so that we can take the square-root of both \tilde{\mathbf{D}} and its inverse. We’ll use that to regroup a bit:

(23)   \begin{align*}   \tilde{y}_k   &= \mathbf{e}^T \tilde{\mathbf{P}} \tilde{\mathbf{D}} \underbrace{\tilde{\mathbf{Q}}^{T} \tilde{\mathbf{Q}}}_{\mathbf{I}_n} \tilde{\mathbf{D}}^{-1} \tilde{\mathbf{P}}^{T} \left(\mathbf{H}_{2} \tilde{\mathbf{Q}} \tilde{\mathbf{D}}^{-1} \tilde{\mathbf{P}}^{T} \right)^{k-1} \tilde{\mathbf{P}} \tilde{\mathbf{D}} \tilde{\mathbf{Q}}^{T} \mathbf{e} \\   &= \mathbf{e}^T \tilde{\mathbf{P}} \tilde{\mathbf{D}} \tilde{\mathbf{D}}^{-1} \left(\tilde{\mathbf{P}}^{T} \mathbf{H}_{2} \tilde{\mathbf{Q}} \tilde{\mathbf{D}}^{-1} \right)^{k-1} \underbrace{\tilde{\mathbf{P}}^{T} \tilde{\mathbf{P}}}_{=\mathbf{I}_n} \tilde{\mathbf{D}} \tilde{\mathbf{Q}}^{T} \mathbf{e} \\   &= \mathbf{e}^T \tilde{\mathbf{P}} \underbrace{\tilde{\mathbf{D}} \tilde{\mathbf{D}}^{-\frac{1}{2}}}_{=\tilde{\mathbf{D}}^{\frac{1}{2}}} \left(\tilde{\mathbf{D}}^{-\frac{1}{2}} \tilde{\mathbf{P}}^{T} \mathbf{H}_{2} \tilde{\mathbf{Q}} \tilde{\mathbf{D}}^{-\frac{1}{2}} \right)^{k-1} \underbrace{\tilde{\mathbf{D}}^{-\frac{1}{2}} \tilde{\mathbf{D}}}_{=\tilde{\mathbf{D}}^{\frac{1}{2}}} \tilde{\mathbf{Q}}^{T} \mathbf{e} \\   &= \underbrace{\mathbf{e}^T \tilde{\mathbf{P}} \tilde{\mathbf{D}}^{\frac{1}{2}}}_{=:\hat{\mathbf{c}}^T}{\underbrace{\left(\tilde{\mathbf{D}}^{-\frac{1}{2}} \tilde{\mathbf{P}}^{T} \mathbf{H}_{2} \tilde{\mathbf{Q}} \tilde{\mathbf{D}}^{-\frac{1}{2}} \right)}_{=:\hat{\mathbf{A}}}}^{k-1} \underbrace{\tilde{\mathbf{D}}^{\frac{1}{2}} \tilde{\mathbf{Q}}^{T} \mathbf{e}}_{=:\hat{\mathbf{b}}} \\   &= \hat{\mathbf{c}}^T \hat{\mathbf{A}}^{k-1} \hat{\mathbf{b}} \end{align*}

Notice that in the first step, we have pushed \tilde{\mathbf{P}}^{T} from the left into the parenthesised, inner expression, and out the right side again. This does not change the value of the expression. We have then done the same with \tilde{\mathbf{D}}^{-\frac{1}{2}}, and then simplified the remaining expression. Now, if we check the dimensions of \hat{\mathbf{c}}^T, \hat{\mathbf{A}} and \hat{\mathbf{b}}, they indeed match the reduced order of our system.

In fact, nothing should have changed except for the order of the system, so looking at our impulse response, we should not see any difference:

Comparison of Impulse Responses: Measured vs. Estimated from Reduced Order System

Similarly, our step response should still look the same:

Comparison of Step Responses: Measured vs. Estimated from Reduced Order System


Our expedition into the details of the ERA was quite fruitful: We have learned a few techniques for restructuring, analysis and reduction of measurement data.

Many of these will come handy when we examine the Observer/Kalman System Identification Algorithm (OKID). The OKID is used to extract the impulse response from measurement data that was acquired using arbitrary control input.

Further, some of these ideas also will allow us to identify non-linear systems — or at least approximate them. As the engine/propeller combination includes possibly non-linear elements due to aerodynamic drag, this is specifically relevant for the identification of multicopter propulsion systems.

[1] [doi] J. Juang and R. S. Pappa, “An Eigensystem Realization Algorithm for Modal Parameter Identification and Model Reduction,” Journal of guidance control and dynamics, vol. 8, iss. 5, 1985.
author = {Juang, Jer-Nan and Pappa, Richard S.},
title = {{A}n {E}igensystem {R}ealization {A}lgorithm for {M}odal {P}arameter {I}dentification and {M}odel {R}eduction},
journal = {Journal of Guidance Control and Dynamics},
year = {1985},
volume = {8},
number = {5},
month = nov,
abstract = {A method called the eigensystem realization algorithm is developed for modal parameter identification and model reduction of dynamic systems from test data. A new approach is introduced in conjunction with the singular-value decomposition technique to derive the basic formulation of minimum order realization which is an extended version of the Ho-Kalman algorithm. The basic formulation is then transformed into modal space for modal parameter identification. Two accuracy indicators are developed to quantitatively identify the system and noise modes. For illustration of the algorithm, an example is shown using experimental data from the Galileo spacecraft.},
doi = {10.2514/3.20031},
file = {:Juang.Pappa1985 - An Eigensystem Realization Algorithm for Modal Parameter Identification and Model Reduction.pdf:PDF},