Introduction
In the realm of industrial robotics and real-time control, solving time-dependent linear equations (TDLE) is a critical task. These equations often govern the relationship between a robot's joint configurations and its end-effector position. However, practical applications impose strict physical boundaries—such as joint angle limits—that must be respected to prevent mechanical damage. While continuous-time Zeroing Neural Networks (ZNNs) have proven effective for time-varying problems, their transition to digital hardware requires discretization, which frequently introduces significant approximation errors.
Methodology
This research introduces a novel Discrete-Time Zeroing Neural Network (DTZNN) designed to bridge the gap between theoretical precision and hardware implementation. The approach centers on three key innovations:
- Boundary Integration: Instead of treating physical constraints as post-processing checks, the model incorporates them directly into the system state using a projection-based method. This ensures that the solution remains within the "safe zone" at every iteration.
- High-Order Discretization: To overcome the accuracy limitations of standard Euler or low-order Taylor methods, the authors employ a 7-step Taylor difference formula. By utilizing the current state and six previous iterations, the model achieves a steady-state calculation error of $O(\epsilon^5)$.
- Computational Efficiency: Despite the increased "memory" of the 7-step formula, the algorithm maintains a computational complexity of $O(n^2)$, making it highly suitable for real-time digital controllers.
Experimental Validation
The DTZNN was rigorously tested through both synthetic benchmarks and robotic simulations.
Synthetic Benchmarks
Using time-varying matrices and vectors, the model demonstrated unprecedented accuracy. Under various sampling intervals ($\epsilon$), the proposed DTZNN achieved steady-state errors in the range of $10^{-13}$ to $10^{-14}$. This represents a significant improvement over existing state-of-the-art models, which typically hover around $10^{-10}$ accuracy. The $O(\epsilon^5)$ convergence rate implies that a tenfold reduction in the sampling interval results in a 100,000-fold reduction in error.
Robotic Motion Planning
The model's practical utility was validated using a kinematic simulation of the PUMA560 robotic arm. Tasked with following complex geometric paths (tricuspid, circular, and Rhodonea), the DTZNN enabled the end-effector to track trajectories with a maximum positioning error of just $1.34 \times 10^{-11}$ meters. Crucially, the network successfully managed initial states outside the allowed boundaries, rapidly pulling the system back into compliance without violating physical limits.
Conclusion and Future Directions
The proposed $O(\epsilon^5)$ DTZNN offers a robust and highly accurate solution for constrained TDLE problems in digital environments. While the current model excels in clean simulations, future work will focus on enhancing noise tolerance for real-world sensor data and integrating external constraints such as obstacle avoidance. This advancement paves the way for more precise and reliable autonomous systems in complex industrial settings.
Summary updated