RNN Models - CLAUDE.md

This module provides architecture guidance for recurrent neural networks in PHM-Vibench. For available models and configuration, see [@README.md].

Architecture Overview

RNN models implement recurrent architectures designed for capturing sequential dependencies in vibration signals:

Input Signal (L, C)
     ↓
┌─────────────────────────────────────┐
│  RNN Layers (N ×)                     │
│  - LSTM / GRU cells                   │
│  - Bidirectional (optional)           │
│  - Dropout for regularization         │
└─────────────────────────────────────┘
     ↓
┌─────────────────────────────────────┐
│  Attention Layer (optional)          │
│  - Focus on important time steps     │
└─────────────────────────────────────┘
     ↓
Output (Classification/Prediction)

Available Models

| Model | Description | Best For | |-------|-------------|----------| | AttentionLSTM | LSTM with attention mechanism | Long sequences, salient features | | AttentionGRU | GRU with attention | Faster training, good performance | | ConvLSTM | Convolutional LSTM | Spatial-temporal patterns | | ResidualRNN | RNN with residual connections | Deep recurrent architectures |

Design Considerations

LSTM vs GRU

LSTM: More parameters, better for complex patterns
GRU: Fewer parameters, faster training

Bidirectional Processing

Processes sequence in both forward and backward directions
Captures past and future context
Doubles parameter count

Sequence Length

RNNs can struggle with very long sequences (>10000 steps):

Consider downsampling or windowing
Use attention mechanisms to focus on important parts

Configuration Pattern

model:
  type: "RNN"
  name: "AttentionLSTM"

  # Architecture
  hidden_size: 128        # Hidden state dimension
  num_layers: 2           # Number of RNN layers
  bidirectional: true     # Process both directions

  # Regularization
  dropout: 0.2

  # Output
  num_classes: 10

PHM-Vibench CLAUDE.md