⚡ electrical engineering

Incremental Data Driven Transfer Identification

This paper introduces a geometric, incremental method for online transfer identification of deterministic linear time-invariant systems that leverages abundant data from a similar source system to select the most consistent model at each step, ensuring convergence to the true system as new data becomes informative.

Original authors: N. Naveen Mukesh, Debraj Chakraborty

Published 2026-02-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: N. Naveen Mukesh, Debraj Chakraborty

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to learn how to drive a brand-new, slightly different model of a car. You've never driven this specific car before, but you have spent thousands of hours driving a very similar model (let's call it the "Old Car").

The Problem:
Usually, to learn how to drive the new car perfectly, you need to drive it for a long time, making many mistakes and collecting lots of data. Only after you've driven it enough can you be sure exactly how the brakes and steering respond. But what if you need to drive this new car right now? You don't have time to collect years of data. If you try to guess the controls based on just a few seconds of driving, you might crash because your guess is too vague.

The Solution (The Paper's Idea):
This paper proposes a clever "transfer learning" trick. It says: "Start by assuming the new car behaves exactly like the Old Car. But, as you drive the new car and collect a tiny bit of data, slowly adjust your assumption to match the new reality."

Here is how the authors' method works, broken down into simple concepts:

1. The "Geometric" Map

Instead of thinking about complex math equations, imagine the behavior of a car as a shape (a geometric space).

The Old Car's Shape: You have a huge, detailed map of the Old Car's behavior because you've driven it so much. Let's call this a "Solid Blue Sphere."
The New Car's Reality: The new car has its own true shape, but you don't know it yet. Let's call this a "Hidden Red Star."
The Data: Every time you drive the new car for a few seconds, you get a tiny clue. This clue is like a small dot on a map. At first, you only have one or two dots. These dots could fit inside many different shapes. You don't know which shape is the real one.

2. The "Closest Fit" Strategy

The authors' method is like a smart navigator.

Step 1: You have your "Solid Blue Sphere" (the Old Car) and a few "dots" (new data).
Step 2: The navigator looks for a new shape that must contain your new dots (so it fits the new car's current behavior) but is also as close as possible to the Blue Sphere (so it doesn't stray too far from what you already know).
The Result: You get a "Hybrid Shape." It's not the Old Car, and it's not the True New Car yet, but it's the best possible guess given what you know right now.

3. The "Sliding" Process

As you drive the new car longer, you collect more dots.

Early Stage: You have very few dots. Your "Hybrid Shape" looks almost exactly like the Old Car (the Blue Sphere) because you need that extra information to fill in the gaps.
Middle Stage: You have more dots. The "Hybrid Shape" starts to stretch and shift away from the Blue Sphere, moving toward the Hidden Red Star.
Late Stage: You have enough dots to define the shape perfectly. The "Hybrid Shape" has now completely merged with the Hidden Red Star. You have perfectly identified the new car.

Why is this special?

Most other methods try to do one of two things:

Ignore the Old Car: They wait until you have enough data to ignore the Old Car entirely. This takes too long.
Mix them blindly: They just mash the data from the Old Car and New Car together. This is dangerous because if the cars are slightly different, the math might break, or you might end up with a model that fits neither car well.

This paper's method is unique because:

It's Automatic: It doesn't require you to guess "how much" to trust the Old Car vs. the New Car. The math naturally handles the balance.
It's Safe: It guarantees that at every single step, there is one and only one "best guess" model. You never get stuck with confusion.
It's Fast: It allows you to control the new car (like placing its "poles" or stability points) almost immediately, even with very little data, by leaning on the knowledge of the similar system.

The Real-World Analogy: Learning a New Language

Imagine you know Spanish perfectly (the Similar System). You move to a country where they speak a dialect that is 90% Spanish but has some unique slang (the True System).

Traditional Method: You refuse to speak until you have studied the dialect for 5 years.
This Paper's Method: You start speaking immediately. You assume everyone speaks perfect Spanish. When you hear a new slang word (new data), you gently adjust your mental dictionary to include it, but you keep your Spanish grammar as the foundation. As you hear more slang, your "mental dictionary" shifts closer and closer to the true dialect, eventually becoming perfect.

Conclusion

The authors have created a mathematical "bridge" that lets us use old knowledge to understand new systems instantly. It ensures that as we gather more evidence, our understanding smoothly and safely evolves from "what we think it is" to "what it actually is," allowing us to control complex machines (like nuclear reactors or unstable systems) much faster and safer than before.

1. Problem Statement

The paper addresses the challenge of online system identification for deterministic Linear Time-Invariant (LTI) systems when data is scarce.

Context: Traditional model-based control requires large datasets to satisfy rank conditions (informativity) before a unique model can be identified. Data-driven control methods also typically require fully informative datasets to guarantee stability and performance.
The Gap: In many practical scenarios (e.g., petrochemical processes with large time constants), collecting sufficient data is time-consuming, costly, or impossible in the early stages of operation.
The Setup: The authors assume access to:
1. Abundant data from a "similar" system ( $S$ ) which is close to, but not identical to, the "true" system ( $T$ ).
2. Sparse, incremental data collected online from the true system ( $T$ ).
Objective: To construct a unique, consistent system model at every time step $i$ using the sparse online data and the abundant similar data, such that the model converges to the true system as more online data becomes available.

2. Methodology: Geometric Transfer Identification

The core contribution is a geometric framework that treats system identification as a problem of subspace manipulation in the Grassmannian manifold.

A. Subspace Representation

Systems are represented by the subspaces spanned by their input-state-output data matrices.
Let $S$ be the subspace spanned by the abundant data from the similar system.
Let $T_i$ be the subspace spanned by the sparse data from the true system collected up to time $i$ .
The goal is to find a subspace $H_i$ $H_{i}$ that:
1. Contains the true system data ( $T_i \subseteq H_i$ ).
2. Has the correct dimension for system identification ( $\dim(H_i) = \Lambda$ , where $\Lambda = n + m$ ).
3. Is closest to the similar system subspace $S$ (minimizing the distance between subspaces).

B. The Incremental Construction Algorithm

The authors propose a closed-form formula to construct the optimal subspace $H_i$ at each step:
$H_i = T_i + (T_i^\perp \cap S)$
Where:

$T_i^\perp$ is the orthogonal complement of the current true system data subspace.
$T_i^\perp \cap S$ represents the component of the similar system's information that is orthogonal to the current true system data.
Intuition: The method "fills in" the missing dimensions of the sparse true system data using the most relevant information from the similar system, specifically the parts that do not contradict the new online data.

C. Model Extraction

Once $H_i$ is constructed, a unique state-space model $(A_i, B_i)$ is extracted using the Moore-Penrose pseudoinverse:
$[A_i \quad B_i] = X_i^+ \begin{bmatrix} X_i^- \\ U_i^- \end{bmatrix}^\dagger$
where $X_i^+, X_i^-, U_i^-$ are partitions of the basis matrix of $H_i$ .

3. Key Theoretical Contributions

The paper provides rigorous mathematical proofs for the following properties:

Existence and Uniqueness: Under the assumption that the similar and true system subspaces are not "partially orthogonal" (a generic property), the constructed subspace $H_i$ has the correct dimension $\Lambda$ and yields a unique state-space model at every time step.
Monotonic Divergence from Similar System: As more true system data is collected ( $i$ increases), the distance between the estimated model and the similar system ( $d(S, H_i)$ ) increases monotonically. The model naturally "drifts" away from the prior as evidence from the true system accumulates.
Monotonic Convergence to True System: Conversely, the distance between the estimated model and the true system subspace ( $d(T_\Omega, H_i)$ ) decreases monotonically.
Convergence Guarantees: Once the online data becomes informative (i.e., $\text{rank}(T_i) = \Lambda$ ), the constructed subspace $H_i$ converges exactly to the true system subspace $T_\Omega$ , and the identified parameters converge to the true parameters.
Error Bounds: The authors derive bounds on the Frobenius norm error of the identified parameters ( $\|[A_i, B_i] - [A_T, B_T]\|_F$ ) as a function of the subspace distance $d(H_i, T_\Omega)$ . This quantifies how estimation error decreases as the subspace distance shrinks.
No Tuning Required: Unlike Bayesian or optimization-based transfer learning methods, this approach requires no weighting parameters or online optimization solvers (like LMIs), making it computationally efficient.

4. Numerical Results and Case Studies

The authors validate the theory with two numerical examples:

Case 1: High-Dimensional Nuclear Reactor Model:
- A pressurized heavy water reactor model ( $n=56, m=15$ ) is used.
- A perturbed version serves as the similar system.
- Result: The identified model error ( $\|[A_i, B_i] - [A_T, B_T]\|_F$ ) and subspace distance decrease smoothly as data is collected, converging to zero once sufficient data is gathered.
Case 2: Pole Placement (Control Application):
- An unstable system is controlled using a "certainty equivalence" approach (designing a controller based on the current identified model).
- Comparison:
  - Using the Similar System directly resulted in an unstable closed-loop system.
  - Using the Incremental Transfer Identified Model (with minimal true data) resulted in a stable closed-loop system with poles close to the desired locations.
- Significance: This demonstrates that the method enables effective control design even when the true system data is insufficient for standalone identification.

5. Significance and Impact

Bridging the Data Gap: The method allows for the deployment of controllers and accurate models in the "early stages" of system operation where data is traditionally too sparse for identification.
Computational Efficiency: By avoiding iterative optimization and providing a closed-form geometric solution, the algorithm is suitable for real-time, online implementation.
Theoretical Rigor: Unlike many data-driven control papers that rely on heuristics or deep learning, this work provides formal guarantees on existence, uniqueness, and convergence rates based on subspace geometry.
Practical Applicability: The approach is particularly relevant for systems with slow dynamics, high costs of experimentation, or systems requiring frequent re-identification (e.g., aging infrastructure or time-varying processes).

In summary, this paper introduces a robust, mathematically grounded framework for incremental transfer learning in control systems, enabling the transition from a "similar" prior model to a "true" system model without a gap in control capability.