TLDR: A new unscented Kalman filter (IPS-UKF) is introduced for real-time estimation of unknown inputs, system parameters, and dynamic states in both linear and nonlinear systems. It uses a two-stage input estimation process and is shown to be uniquely identifiable if at least one input is known. While full state measurements are recommended, the method can work with fewer measurements (e.g., displacement and velocity), but acceleration-only data is unreliable.
Monitoring the health and behavior of dynamic systems, such as bridges or buildings, is crucial for safety and maintenance. Often, engineers rely on sensors to track various aspects of a system, but direct measurement of all inputs (like wind loads or traffic forces) and internal parameters can be challenging, costly, or even impossible. This is where “output-only” strategies become vital, allowing us to understand a system using only its measured responses, such as vibrations.
Traditional methods for identifying system parameters often assume that the input forces are known, which is rarely the case in real-world scenarios. Other approaches might struggle with nonlinear systems, introduce linearization errors, or are only suitable for offline analysis, meaning they can’t provide real-time insights. There’s a clear need for robust techniques that can simultaneously estimate unknown inputs, system parameters, and dynamic states in real-time, even for complex nonlinear systems.
A Novel Approach to Real-Time System Estimation
Researchers Marios Impraimakis and Andrew W. Smyth from Columbia University have developed a novel unscented Kalman filter (UKF) method designed for real-time input-parameter-state (IPS) estimation. This advanced filtering technique addresses the limitations of previous methods by jointly estimating all dynamic states, system parameters, and the unknown input simultaneously. The core of their method involves a two-stage estimation process for the unknown input within each time step. First, an initial estimate of the input is made using predicted dynamic states and system parameters. Then, a final, more accurate input estimation is achieved by incorporating corrected states and parameters that have been updated with actual measurements.
A significant advantage of this new methodology is that it avoids the need for complex Jacobian derivatives, least squares introductions, or nonphysical propagation processes, which can often complicate and limit other filtering techniques. The method has been shown to work effectively for both linear and nonlinear systems, making it highly versatile for various engineering applications. You can read the full research paper for more technical details here: An unscented Kalman filter method for real time input-parameter-state estimation.
Understanding System Identifiability
A key aspect of any estimation problem is “identifiability” – whether the system’s properties can be uniquely determined from the available data. The researchers demonstrated through a perturbation analysis that a system can potentially be uniquely identified if at least one of its inputs is known, whether that input is zero or non-zero. This is particularly important for multi-degree-of-freedom (MDOF) systems, where knowing the input at just one point can help correctly identify parameters and other unknown inputs. This contrasts with simpler single-degree-of-freedom (SDOF) systems, where identifying unknown inputs and parameters simultaneously can be more challenging without additional information.
Practical Applications and Data Requirements
The effectiveness of the proposed IPS-UKF method was validated through numerical examples involving both linear 3-DOF and nonlinear 2-DOF systems. In both cases, the method successfully estimated the system responses, stiffness, damping, and input forces with satisfactory convergence. This demonstrates its potential for real-world applications in structural health monitoring and other fields.
While the method generally recommends measuring all dynamic states (displacement, velocity, and acceleration) for a comprehensive estimation, the researchers also conducted a sensitivity analysis. This analysis revealed that full state measurement is not always mandatory. For instance, measuring at least two dynamic states at each degree of freedom can still be adequate for identifying nonlinear systems. However, relying solely on acceleration measurements proved to be unreliable, often leading to misleading results and divergence of the estimation process. This insight is crucial for optimizing sensor placement and data acquisition strategies in practical applications.
Also Read:
- Real-DRL: Bridging the Gap for Safe AI in Physical Systems
- Enhancing Time-Series Forecasts with Adaptive Quadratic Training Objectives
Conclusion
This novel unscented Kalman filter offers a powerful and practical solution for real-time input-parameter-state estimation. By effectively handling unknown inputs and nonlinear system dynamics, it provides a better understanding of system behavior compared to classical output-only identification strategies. The ability to jointly estimate dynamic states, parameters, and inputs online makes it a valuable tool for modern dynamical system monitoring, damage detection, and prognosis.
