Neural Luenberger state observer for nonautonomous nonlinear systems

1. Problem Statement

The paper addresses the challenge of state observation for nonautonomous nonlinear systems (systems with external manipulated inputs) where the underlying mathematical model is unknown or too complex to derive analytically.

• Context: Traditional observers (e.g., Extended Kalman Filter, High-Gain Observers) rely heavily on accurate first-principles models. When models are unavailable, data-driven methods are preferred.
• Gap: Existing data-driven observers often focus on autonomous (unforced) systems or require known system dynamics to design the observer structure. There is a lack of fully model-free frameworks that can simultaneously learn the nonlinear state transformation and the specific input-affine terms required to handle external inputs for nonautonomous systems.
• Goal: Develop a model-free state observer that learns directly from historical input-output-state data, capable of handling manipulated inputs without requiring explicit knowledge of the system dynamics ($f$ and $g$).

2. Methodology: Neural Luenberger Observer (NLOX)

The authors propose a Neural Luenberger Observer for Nonlinear systems with external inputs (NLOX). This approach extends the Kazantzis-Kravaris/Luenberger (KKL) observer framework using deep learning.

Theoretical Foundation

• KKL Framework: The KKL observer maps the original nonlinear system states ($x$) to an observer state space ($z$) via an injective nonlinear transformation $T$. In the observer space, the dynamics are linear time-invariant (LTI).
• Nonautonomous Extension: For systems with inputs ($u \neq 0$), the standard KKL structure fails to converge. The authors adopt the extension by Bernard and Andrieu (2018), which augments the LTI observer dynamics with an input-affine term $\omega(z)u$.
  • Observer Dynamics: $\dot{z} = Az + By + \omega(z)u$
  • State Estimation: $\hat{x} = T^\dagger(z)$, where $T^\dagger$ is the left inverse of the transformation $T$.

Neural Network Architecture

Instead of solving the complex Partial Differential Equations (PDEs) required to find $T$ and $\omega$ analytically, the authors use Fully Connected Feedforward Multilayer Perceptrons (MLPs) to learn these mappings from data.

1. Input-Affine Term Network ($\omega$): A neural network that takes the observer state $z$ and learns the function $\omega(z)$ to compensate for the manipulated input $u$.
2. Inverse Mapping Network ($T^\dagger$): A second neural network that maps the observer state $z$ back to the estimated system state $\hat{x}$.
3. Training Process:
   • Data: Offline collection of trajectories containing inputs ($u$), outputs ($y$), and true states ($x$).
   • Discretization: The continuous observer dynamics are discretized (Forward Euler) for simulation.
   • Loss Function: The networks are trained to minimize the squared error between the true state $x$ and the estimated state $\hat{x}$:
     $$\mathcal{L} = \frac{1}{MN} \sum_{i,k} \| x_k^{(i)} - T^\dagger(z_k^{(i)} | \psi) \|_2^2$$
   • Optimization: Parameters are updated via backpropagation using the RMSprop optimizer.

Theoretical Guarantees

The paper provides a rigorous theoretical analysis of the observer's generalization error:

• Stability Assumptions: The system is assumed to be Incrementally Input-to-State Stable ($\delta$ISS).
• Lipschitz Continuity: The authors prove that the loss function is Lipschitz continuous with respect to the initial conditions and input sequences, given bounded neural network weights.
• Error Bound: A probabilistic upper bound on the generalization error is derived using Rademacher complexity. The bound scales as $\mathcal{O}(1/\sqrt{MN})$, where $M$ is the number of trajectories and $N$ is the trajectory length, proving statistical consistency. It also accounts for discretization errors.

3. Key Contributions

1. Model-Free Synthesis: A fully data-driven framework for synthesizing observers for nonautonomous nonlinear systems without requiring a first-principles model.
2. Joint Learning of Structure: Simultaneous learning of both the input-affine correction term ($\omega$) and the nonlinear state reconstruction ($T^\dagger$) within the KKL structure.
3. Theoretical Error Bounds: Derivation of a finite-sample generalization error bound that quantifies the trade-off between data richness (number of samples) and numerical resolution (sampling time).
4. High-Dimensional Applicability: Demonstration that the method scales effectively to higher-dimensional systems (6-state Williams-Otto reactor), addressing a limitation in existing data-driven literature which often focuses on low-dimensional (2-state) systems.

4. Numerical Results

The method was validated on two chemical process case studies and compared against model-based benchmarks (Analytical KKL, Extended Kalman Filter - EKF, and Sliding Mode Observer - SMO).

Case Study 1: Bioreactor (2 States)

• Setup: Constant-volume bioreactor with Contois growth kinetics.
• Performance:
  • NLOX achieved RMSE comparable to the analytical KKL solution (the "ground truth"), despite being model-free.
  • NLOX outperformed the EKF and SMO in terms of overall accuracy and robustness.
  • Robustness: Under stochastic input disturbances, NLOX maintained competitive performance, showing lower error for the microorganism concentration ($x_1$) compared to benchmarks.

Case Study 2: Williams-Otto Reactor (6 States)

• Setup: Non-isothermal CSTR with 3 reactions and 6 state variables (complex nonlinear dynamics).
• Implementation Detail: Due to the high dimensionality and non-convexity of the loss landscape, the authors trained six separate networks for the inverse mapping $T^\dagger$ (one for each state) rather than a single network, to avoid gradient competition and local minima.
• Performance:
  • NLOX achieved a 3.96-fold improvement in accuracy over EKF and a 1.37-fold improvement over SMO.
  • Crucially, NLOX demonstrated superior performance during the initialization phase (transient response) where the EKF struggled due to poor linearization assumptions.
  • The method successfully handled noise and disturbances, maintaining stable estimation across all 6 state variables.

5. Significance and Future Work

• Significance: This work bridges the gap between theoretical observer design (KKL) and modern data-driven techniques (Deep Learning). It proves that complex, high-dimensional, nonautonomous systems can be observed accurately without explicit modeling, which is vital for industrial processes where models are uncertain or unavailable.
• Limitations: The current approach relies on supervised learning, requiring access to true state data ($x$) for offline training.
• Future Directions: The authors plan to explore unsupervised learning methods. The goal is to formulate a joint optimization problem where the observer learns a "latent state" representation consistent with input-output data, eliminating the need for ground-truth state measurements during the training phase.

In summary, the paper presents a robust, theoretically grounded, and practically effective method for state estimation in complex chemical processes, demonstrating that neural networks can successfully learn the intricate geometric structures required for nonlinear observer design.