This repository contains the complete research documentation and resources for developing neural network–based modeling and control systems for a 2-Degree-of-Freedom (2DOF) helicopter. The work focuses on experimental validation, system identification, and advanced control using deep learning–driven approaches.
The 2DOF helicopter represents a highly nonlinear, coupled, and unstable system, making it an excellent benchmark for testing intelligent control algorithms. This project experimentally investigates neural network architectures (FNN, LSTM, NARX) for system modeling and integrates them into control schemes including PID, Fuzzy Logic, and Model Predictive Control (MPC).
- Real-world implementation of data-driven helicopter control
- Comparative study of FNN, LSTM, and NARX architectures
- Experimental system identification via Differential Evolution (DE)
- Development of a NARX–MPC control scheme achieving superior accuracy and speed
- Comprehensive performance evaluation in real time (100 Hz sampling)
The experimental setup allows pitch and yaw motion, with strong aerodynamic coupling between axes.
Specifications
| Parameter | Description |
|---|---|
| Degrees of Freedom | 2 (Pitch and Yaw) |
| Control Inputs | Dual 2200KV BLDC motors |
| Sensors | Dual rotary encoders, voltage & current sensors |
| Actuators | Dual ESC 30A controllers |
| Controller | Arduino Mega 2560 (100 Hz sampling rate) |
| Power Supply | 11.1V LiPo battery |
| Interface | Serial communication for data logging and control |
Key Components
| Component | Function |
|---|---|
| Arduino Mega 2560 | Executes control loops and handles data acquisition |
| Dual ESC 30A | Regulates motor speed |
| BLDC Motors (2200KV) | Provides high-speed propulsion for pitch/yaw axes |
| Sensors | Encoders for position, current and voltage monitoring |
| Power | 11.1V LiPo battery with regulated distribution and grounding |
This research explores three neural network models for helicopter dynamics and control. Each model is developed and tested in its own dedicated repository.
Repository: 2DOF Helicopter FNN
- Architecture: Multi-layer perceptron with backpropagation training
- Application: Static mapping of system states to control actions
- Advantages: Simple implementation, fast computation
Repository: 2DOF Helicopter LSTM
- Architecture: Recurrent neural network with memory cells
- Application: Temporal sequence modeling for dynamic control
- Advantages: Long-term dependency learning, sequence processing
Repository: 2DOF Helicopter NARX
- Architecture: Recurrent network with external input feedback
- Application: System identification and predictive control
- Advantages: Excellent for nonlinear system modeling
| Parameter | Description |
|---|---|
| Excitation Signal | Pseudo-Random Binary Sequence (PRBS) |
| Sampling Rate | 100 Hz |
| Experiment Duration | 30 minutes per axis |
| Preprocessing | Low-pass filtering and normalization |
| Data Split | 70% training / 15% validation / 15% testing |
| Optimization | Differential Evolution for hyperparameter tuning |
| Model | R² | MSE | Inference Time (s) | Observations |
|---|---|---|---|---|
| FNN | 0.57 | 3.2 | 0.43 | Weak in temporal modeling |
| LSTM | 0.887 | 0.883 | 1.76 | Learns temporal dependencies, higher computation cost |
| NARX | 0.9992 | 0.005 | 0.7 | Highest accuracy and real-time feasible performance |
The 2DOF helicopter control performance was evaluated using PID, Fuzzy Logic, and Model Predictive Control (MPC) strategies. Each controller was tested for pitch and yaw responses under identical operating conditions.
| Controller | Overshoot (%) | Settling Time (s) | Steady-State Error (°) | RMSE (°) | Improvement vs PID | Remarks |
|---|---|---|---|---|---|---|
| PID | 5.4 (Pitch) / 6.5 (Yaw) | 9.22 / 18.9 | 0.5 / 0.6 | 13.7 | — | Sensitive to coupling and nonlinearities |
| Fuzzy Logic | 4.6 (Pitch) / 5.1 (Yaw) | 8.4 / 16.9 | 0.3 / 0.4 | 10.9 | +30% | Improved robustness to nonlinearities |
| MPC (NARX) | 0 / 1.25 | 8.21 / 12.56 | 0.0 / 0.28 | 9.4 / 1.4 | +78% | Best stability, minimal error, predictive control |
| Method | Description | Key Result |
|---|---|---|
| FNN | Baseline static neural controller | Limited accuracy for dynamic systems |
| LSTM | Temporal model capturing past state effects | Improved accuracy but slower inference |
| NARX | Dynamic recurrent model with exogenous feedback | R² = 0.9992, MSE = 0.005 |
| PID | Classical control baseline | Highest overshoot and error |
| FLC | Nonlinear rule-based control | 30% RMSE reduction vs PID |
| MPC (NARX) | Predictive control with NARX model | 78% RMSE reduction vs PID, best settling and stability |