Edge-1DCNN-LSTM: Preliminary Experiment

"Edge-1DCNN-LSTM: A Lightweight Hybrid Model for Time-Series Anomaly Detection in Edge Sensing Devices"

Directory Structure

reproduce_paper/
├── README.md              # This file
├── requirements.txt       # Python dependencies
├── models/
│   ├── __init__.py            # Model registry (9 models) + factory functions
│   ├── edge_1dcnn_lstm.py     # Proposed Edge-1DCNN-LSTM model
│   ├── vanilla_lstm.py        # Baseline: Vanilla LSTM
│   ├── vanilla_1dcnn.py       # Baseline: Vanilla 1D-CNN
│   ├── standard_1dcnn_lstm.py # Comparison: Standard 1DCNN-LSTM
│   ├── lightweight_1dcnn_lstm.py # Comparison: Lightweight 1DCNN-LSTM
│   └── end_edge_collaborative.py  # End-edge collaborative framework
├── training/
│   ├── trainer.py             # FocalLoss + train_model (Adam, early stopping)
│   ├── pruner.py              # L1 unstructured pruning + fine-tuning
│   └── quantizer.py           # Dynamic INT8 quantization
├── data/
│   ├── dataset_base.py        # Unified data loader (Simulated + FD001)
│   └── prepare_datasets.py    # Data generation (simulate) & conversion (FD001)
└── results/                   # Experiment output (auto-created)

Setup

# Install dependencies
pip install -r requirements.txt

# Prepare datasets (auto-generates simulated data; FD001 requires manual download)
python -m reproduce_paper.data.prepare_datasets

The simulated dataset (50K samples, 15% anomaly) is generated from a fixed seed (seed=42) and placed at datasets/simulate/dataset.json.

For NASA FD001, the script will attempt an automatic download. If that fails, please:

1. Download C-MAPSS FD001 from NASA Prognostics Repository
2. Place train_FD001.txt in datasets/NASA/
3. Re-run: python -m reproduce_paper.data.prepare_datasets --fd001

Paper Experiments

Experiment B (Section III.B): Detection Performance Comparison

Compares 9 models on two datasets:

ID	Model	Category
1	Pure LSTM	Baseline
2	Pure 1D-CNN	Baseline
3	Standard 1DCNN-LSTM	Comparison
4	Lightweight 1DCNN-LSTM	Comparison
5	Pruned 1DCNN-LSTM (75% sparsity)	Comparison
6	Quantized 1DCNN-LSTM (dynamic INT8)	Comparison
7	End-Edge Collaborative	Framework
8	Ours (Edge-1DCNN-LSTM, FP32)	Proposed
9	Ours (Edge-1DCNN-LSTM, INT8)	Proposed

Expected Results (Simulated Dataset):

Model	F1 Score	AUC
Ours (FP32)	~0.97-0.99*	~0.995
Ours (INT8)	~0.96-0.99*	~0.993

* F1 score varies across runs due to random data splitting (seed=42). AUC is highly stable. The paper reports F1=0.994 from a specific training run with optimal threshold tuning.

Experiment C (Section III.C): Edge Deployment & Quantization

Compares FP32 vs INT8 hybrid quantization on the proposed model:

• INT8 mixed-precision: LSTM layers → INT8, CNN/BatchNorm layers → FP32 (LSTM dominates 99.1% of parameters)
• Expected: ~61% model size reduction, <0.5% F1 degradation
• Note: The 29.5% inference speedup reported in the paper was measured on Raspberry Pi 4B. On x86 CPUs, dynamic quantization may show minimal speedup. Model size reduction is hardware-independent.

Experiment D (Section III.D): End-Edge Collaborative Framework

Evaluates the sentinel-based collaborative framework with threshold sweep (0.5–0.95):

• Sentinel: Vanilla LSTM (lightweight, on edge device)
• Full Model: Edge-1DCNN-LSTM (on cloud/edge server)
• Expected (threshold=0.9): System F1≈0.990, ~72% energy saving*, ~27% wake rate
• *: Energy saving is estimated using E = P × T with power values from Raspberry Pi 4B (sentinel: 0.3W, full: 5.0W). Actual values depend on the target edge hardware.

Model Architecture

Input (B, seq_len, 1)
  → Conv1d(1, 16, k=3, p=1) → BatchNorm → ReLU → MaxPool(k=2)
  → LSTM(input=16, hidden=32, layers=2, dropout=0.2)
  → Take last hidden state → Dropout(0.2) → FC(32, 1) → Sigmoid

Parameters: ~12,700 (with input_dim=1) | ~14,900 (paper-reported with overhead)

Training Configuration

Parameter	Simulated	FD001
Epochs	100	200
Learning Rate	0.001	0.001
Batch Size	64	64
Optimizer	Adam	Adam
Scheduler	ReduceLROnPlateau	ReduceLROnPlateau
Early Stopping	patience=15	patience=25
Loss	BCELoss	FocalLoss (alpha=0.75, gamma=2)
Sequence Length	10	10
Train/Val/Test	70/15/15	70/15/15
: Sentinel confidence =	p - 0.5	× 2; low-confidence samples forwarded to full model.

Project Status

This repository contains the preliminary experimental code corresponding to the submitted manuscript Edge-1DCNN-LSTM: A Lightweight Hybrid Model for Time-Series Anomaly Detection in Edge Sensing Devices.

Planned Updates

-Generalization Validation: Add pre-trained models and testing scripts for additional public industrial time-series anomaly detection datasets (e.g., SMD, SWaT, WADI)

-Hardware Deployment Optimization: Complete the hardware-aware INT8 optimization code for ultra-low-power MCUs (e.g., ESP32, STM32) with dedicated acceleration units

-Documentation Improvement: Add a detailed step-by-step deployment tutorial for terminal-edge hardware setup, and expand the hyperparameter tuning guide

-Code Refactoring: Refactor the codebase for better readability and modularity, and add Python package installation support.