TL;DR
This paper introduces a robust framework for detecting and isolating sensor faults in general nonlinear systems by leveraging neural network-based Kazantzis-Kravaris/Luenberger (KKL) observers. By combining theoretical residual bounds for fault detection with a novel differentiated residual approach for fault isolation, the method effectively identifies both complete sensor failures and gradual degradations even in the presence of measurement noise and model uncertainty.
What Problem Does This Paper Address?
In complex autonomous systems, sensors are the primary windows into the state of the process. When a sensor fails completely or begins to drift (degradation), it can lead to catastrophic system failure if not caught early. While "hardware redundancy" (using multiple physical sensors for the same measurement) is a solution, it is often too expensive or bulky.
"Analytical redundancy"—using mathematical models to predict what a sensor should be reading and comparing it to the actual measurement—is a preferred alternative. However, designing these observers for nonlinear systems is notoriously difficult. Most existing methods require the system to have a very specific mathematical structure (like a triangular form), which limits their real-world applicability. The authors aim to create a general-purpose sensor Fault Detection and Isolation (s-FDI) method that works for a broad class of nonlinear systems without restrictive structural assumptions.
Methods
The core of the approach is the Neural Network-based KKL (NN-KKL) Observer.
- KKL Lifting: The KKL framework "lifts" a complex nonlinear system into a higher-dimensional state space where the dynamics become linear and stable.
- Neural Network Approximation: Instead of solving the grueling partial differential equations usually required to find the KKL transformation, the authors use neural networks to learn the mapping ($T$) and its inverse ($T^*$).
- Residual Generation: The observer predicts the system output ($\hat{y}$). The difference between the actual measurement ($y$) and this prediction is the residual ($r$).
- Detection: The authors derive a theoretical threshold ($\tau$). If a residual exceeds this threshold, a fault is detected.
- Isolation via Differentiation: A major challenge in nonlinear observers is that a fault in one sensor often causes residuals to spike across all sensors because the observer's internal states are coupled. To solve this, the authors propose using differentiated residuals—the numerical derivative of the residual. They found that abrupt faults create a distinct "spike" in the derivative that is highly localized to the specific faulty sensor, allowing for accurate isolation.
Datasets
The method was evaluated using numerical simulations of a network of 10 Kuramoto oscillators, a standard model for studying synchronization in systems like power grids and multi-agent robotics.
- Training Data: Generated using Latin hypercube sampling within the state space $X \subset [-2, 2]^{10}$.
- Test Data: 100 initial conditions were used to generate noise-free, fault-free trajectories to establish empirical thresholds for the isolation task.
Key Insights and Ideas
- Decoupling via Dynamics: The most significant "aha moment" is the use of differentiated residuals for isolation. Because the neural network transformation is non-diagonal, a fault in sensor A naturally "leaks" into the residual of sensor B. However, the instantaneous change (the derivative) at the moment of failure is much stronger in the faulty sensor's signal, providing a clear signature for isolation.
- Theoretical Rigor: Unlike many purely data-driven deep learning approaches, this method provides theoretical upper bounds on residuals under fault-free conditions, giving engineers a principled way to set detection thresholds.
- Generalizability: By using the KKL framework, the method avoids the "curse of structure." It doesn't care if your system is in a specific canonical form, making it applicable to a much wider variety of nonlinear dynamics.
Key Results
The observer demonstrated high accuracy in state estimation even under process and measurement noise, as shown in the tracking of unmeasured states $\theta_7$ and $\theta_8$:
In fault scenarios (detailed in Figure 2 below), the method successfully:
- Detected and isolated complete failures (sensor shutdowns).
- Identified multiple simultaneous faults occurring at different times.
- Handled gradual degradations, such as increasing white noise or sinusoidal drift, though isolation in these cases often occurs with a slight delay compared to abrupt faults.
Limitations and Future Work
- Smooth Faults: The isolation method relies on the "spike" in the differentiated residual. If a fault is extremely smooth and low-magnitude (like a very slow sigmoidal drift), it may be detected but remain difficult to isolate to a specific sensor.
- Autonomous Systems: The current work focuses on autonomous systems (those without external control inputs). Extending this to non-autonomous systems is a primary goal for future research.
- Real-world Validation: While the Kuramoto model is a robust benchmark, testing the framework on physical experimental data or high-fidelity industrial simulators is the next logical step.
Summary updated