Residual PINN models BLDC motors 118x faster than traditional solvers
Training takes under 2 minutes on a CPU and inference hits microsecond latencies.
Haitham El-Hussieny presents a novel Residual Physics-Informed Neural Network (Res-PINN) for high-fidelity modeling of Brushless DC (BLDC) motors, a core component in robotic joint control. Traditional ODE solvers, while accurate, are too slow for real-time control loops. This paper proposes a continuous-time surrogate that directly maps inputs (simulation time, three-phase voltages, excitation parameters) to all six motor state variables: rotor angle, angular velocity, three-phase currents, and winding temperature. The network uses a ResNet backbone and a composite physics-data loss function with curriculum scheduling to avoid premature convergence.
Performance benchmarks are striking: training takes under two minutes on a standard CPU, and once trained, inference achieves latencies of just 0.1–22 microseconds per query—up to 118x faster than conventional ODE solvers. This speed makes the model suitable for real-time observer and control applications in robotics, where fast, accurate motor state estimation is critical. The work bridges the gap between physics-based modeling and machine learning, offering a practical tool for embedded and edge deployments.
- Residual PINN predicts all six BLDC motor states (rotor angle, velocity, currents, temperature) with a single forward pass.
- Training takes under 2 minutes on a standard CPU, enabling rapid iteration for different motor designs.
- Inference latency of 0.1–22 µs, up to 118× faster than conventional ODE solvers, suitable for real-time robotic control.
Why It Matters
Enables real-time, high-fidelity motor simulation for robotics—no expensive hardware or slow solvers needed.