Introduction
Modeling complex nonlinear systems is a fundamental challenge in control engineering. Most physical systems are inherently nonlinear, and traditional approaches often rely on linearizing governing equations around a specific operating point using techniques like Taylor series expansion. While useful, these linearized models are inherently local and fail when the system shifts to a different operating condition. This paper explores a robust alternative: using a feed-forward Artificial Neural Network (ANN) to synthesize a global dynamic model from data gathered at multiple local, linearized operating points.
Methodology: ANN-Based System Identification
The authors frame system identification as a Multiple-Input-Single-Output (MISO) regression problem. The core of the approach is a multi-layer feed-forward ANN trained with the Levenberg-Marquardt back-propagation algorithm.
To determine the most effective architecture, a parametric study was conducted, varying the number of hidden layers (1 or 2), the number of neurons per layer (ranging from 5 to 25), and the activation functions (hyperbolic tangent and logarithmic sigmoid). To ensure the reliability of the results and account for the sensitivity of gradient-based training to initial weights, each configuration was trained 20 times independently. Performance was measured by both the mean Root Mean Square Error (RMSE) and its standard deviation.
Experimental Datasets
The proposed method was validated on two distinct systems:
- Nuclear Reactor Power Level: Data from Indian Pressurized Heavy Water Reactors (PHWRs) provided by the Nuclear Power Corporation of India Ltd. The model predicts instantaneous reactor power based on control rod position, elapsed time, initial power, and rod drop percentage.
- AC Servo Position Control System: Experimental data from a GE Fanuc Versa Motion Servomotor. The model predicts angular position based on time, acceleration, and velocity.
A critical insight from the study was the inclusion of "time elapsed" as an explicit input feature. This allows the ANN to capture the temporal dynamics of the physical systems without needing a fixed sampling interval in the model structure.
Results and Discussion
The ANN successfully learned global nonlinear models for both systems. For the nuclear reactor, a single hidden layer with 15 neurons using the tansig activation function proved optimal, achieving a mean RMSE of $0.0014$.
As shown in the reactor data, the model must account for various initial power levels and step-back conditions. The ANN effectively "stitched" these local behaviors into a cohesive global model. Interestingly, the study found that single-hidden-layer networks often provided more consistent performance (lower standard deviation of error) than deeper architectures, which were more prone to getting trapped in local minima during training.
For the AC servo motor, the ANN accurately replicated complex "S-shaped" and "truncated ramp" position curves across different acceleration profiles. While a two-layer network achieved the absolute lowest error ($0.0021$ RMSE), single-layer versions were nearly as accurate and demonstrated greater stability across random initializations.
Conclusion
This research demonstrates that standard feed-forward ANNs can effectively model complex nonlinear dynamics by learning from multiple local operating conditions. This approach bypasses the need for complex analytical modeling while providing a robust global representation. While the optimal network architecture remains system-dependent and requires empirical tuning, these ANN-based models offer a powerful foundation for designing advanced nonlinear control systems.
Summary updated