Neural‑Network State‑Space Estimators

This repository provides all code and data to reproduce the simulations and experiments from the paper:

Neural‑Network State‑Space Estimators Minxing Sun, Li Miao, Qingyu Shen, Yao Mao*, Qiliang Bao

---

📁 Directory Structure

Matlab/                  # Main MATLAB scripts
  ├─ Test.m              # 1. PROBLEM STATEMENT validation
  ├─ NNSSE_Sin.m         # 2. SIMULATION on noisy sine wave
  ├─ NNSSE_DR.m          # 3. EXPERIMENT 1: dual‑reflection mirror
  ├─ NNSSE_Drone.m       # 4. EXPERIMENT 2: UAV trajectory
  ├─ StateSpaceModel_*.m # NNSSM variants
  └─ Estimator_*.m       # EKF / UKF / PF implementations
  └─ NNSSE_*.mat         # All data to reproduce result in paper

Python/                  # Neural‑network training scripts
  ├─ 01_RNN_train_2025.py# Format: <id>_<net>_<task>_<date>.py
  ├─ …                   # Other network types
  ├─ config.json         # Paths & hyper‑parameters
  └─ NN_*.bat            # One‑click training on Windows

EstimationResult/        # Prediction results for MATLAB comparisons
  ├─ Data_1  # limited‑training nets (Exp‑1)
  ├─ Data_2  # full‑training    nets (Exp‑1)
  ├─ Data_3  # limited‑training nets (Exp‑2)
  ├─ Data_4  # full‑training    nets (Exp‑2)
  └─ Data_5  # NNSSEs for Sin   nets (Sim-0)

---

1. PROBLEM STATEMENT Validation (Matlab/Test.m)

Reproduces Fig. 2. Compares classical Kalman filters on states [p_i], [p_i,v_i], [p_i,v_i,a_i], [p_i,v_i,a_i,ẍ_i] against position-stack estimators (E2P, E3P, E4P, E4PVW, E4PRW, E4PTRW). Gaussian noise is regenerated each run, so individual results vary slightly.

---

2. SIMULATION (Matlab/NNSSE_Sin.m)

Emulates a manoeuvring target with:

• Ground truth: 1 Hz sine, amplitude 10
• Sampling: 200 Hz, latency 0.15 s → 3‑frame prediction
• Noise: zero‑mean Gaussian, unit covariance

The script picks a StateSpaceModel_*.m and an Estimator_*.m to assemble each NNSSE variant.

---

3. EXPERIMENT 1: Dual‑Reflection Mirror (Matlab/NNSSE_DR.m)

Bench setup: laser, CCD camera, fast steering mirror, validation mirror. Camera at 200 Hz with 0.015 s latency → 3‑frame prediction. Baseline NN predictions are loaded from:

EstimationResult/Data_1  # limited‑training
EstimationResult/Data_2  # full‑training

---

4. EXPERIMENT 2: UAV Trajectory (Matlab/NNSSE_Drone.m)

Yaw data from a real UAV, including zenith‑crossing reversals. Scripts load neural baselines from:

EstimationResult/Data_3  # limited‑training
EstimationResult/Data_4  # full‑training

---

5. Neural‑Network Training (Python/)

1. Edit config.json to set paths and parameters (train/test/estimate).

2. Run training scripts, e.g.:

   cd Python
   python 00011_RNN_Train_20241129.py

   or on Windows:

   NN_FullTrain.bat

3. Copy generated files into Matlab/EstimationResult/Data_X for MATLAB.

---

Contact: sunminxing20@gmail.com

© 2025 Minxing Sun et al. Licensed under MIT.