Skip to main content

WP2 — PI-VIB-ResNet Model Design

Lead: Sejong University (SJU) — AINTLab — Prof. Syafrudin
Timeline: M4 – M14
Key milestone: M10 — First trained PI-VIB-ResNet model; M14 — Ablation study complete

Objectives​

  1. Design and implement the PI-VIB encoder with physics consistency loss
  2. Train ResNet-BPNN regressor on WP1 T-EV field data
  3. Conduct systematic ablation studies to validate each architectural component

Architecture Components​

VIB Encoder​

  • Input: 51-point impedance spectrum (real + imaginary parts → 102 features after wavelet denoising)
  • Encoder: 3-layer MLP → (μ, log σ²) → reparameterization → latent z ∈ ℝ^32
  • Physics loss: Randles circuit consistency constraints

ResNet-BPNN​

  • 12–18 residual blocks (2-layer each)
  • Batch normalization + dropout (p=0.2)
  • Output: scalar SOH ∈ [0, 100]%

Deliverables​

  • D2.1: PI-VIB architecture specification document (M6)
  • D2.2: Trained PI-VIB-ResNet model (PyTorch checkpoint) — M10
  • D2.3: Ablation study report — M14
  • D2.4: Open-source code release (GitHub) — M14