Detecting Inaccurate Sensors on a Large-Scale Sensor Network Using Centralized and Localized Graph Neural Networks

Introduction

An Anomaly Detecting Framework Base on Deep Spatial-Temporal Graph Model to detect inaccurate air quality sensors.

Paper Link in English

Paper Link in Chinese

Quickstart

GNN_reg.py

• Training method is many to one (model separated).

Usage:

$python GNN_reg.py regression 
                --name <model name; you can choose any name> 
                --model <model type: "gwnet" or "STGCN"> 
                --max_epoch <number of epoch>
                --use_gpu <whether use gpu>
                --num_nodes <number of graph node>
                --visual= <show figures during the training process, TRUE or FALSE>

Example:

$python GNN_reg.py regression 
                --name="GraphWaveNet_v1"
                --model="gwnet"
                --max_epoch=10
                --use_gpu=True
                --num_nodes=6
                --visual=False

Global_GNN_reg.py

• Training method is all to all.

Usage:

$python Global_GNN_reg.py regression 
                --name <model name; you can choose any name> 
                --model <model type: "gwnet" or "STGCN"> 
                --max_epoch <number of epoch>
                --use_gpu <whether use gpu>
                --num_nodes <number of graph node>

Example:

$python Global_GNN_reg.py regression 
                --name="Global_GraphWaveNet_v1"
                --model="gwnet"
                --max_epoch=10
                --use_gpu=True
                --num_nodes=144

Dep_GNN_reg.py

• Training method is many to one (model uniform).

Usage:

$python Dep_GNN_reg.py regression 
                --name <model name; you can choose any name> 
                --model <model type: "gwnet" or "STGCN"> 
                --max_epoch <number of epoch>
                --use_gpu <whether use gpu>
                --num_nodes <number of graph node>

Example:

$python Dep_GNN_reg.py regression 
                --name="Dep_GraphWaveNet_v1"
                --model="gwnet"
                --max_epoch=10
                --use_gpu=True
                --num_nodes=6

deep_learning_reg.py

• Training method is many to one (model separated).

Usage:

$python deep_learning_reg.py regression 
                --name <model name; you can choose any name> 
                --model <model type: "TCN" or "LSTM" or "DNN"> 
                --max_epoch <number of epoch>
                --use_gpu <whether use gpu>

Example:

$python deep_learning_reg.py regression 
                --name="TCN_v1"
                --model="TCN"
                --max_epoch=10
                --use_gpu=True

machine_learning_reg.py

• Training method is many to one (model separated).

Usage:

$python machine_learning_reg.py regression 
                --name <model name; you can choose any name> 
                --model <model type: "Lasso" or "Ridge" or "RandomForest"> 

Example:

$python machine_learning_reg.py regression 
                --name="Lasso_v1"
                --model="Lasso"

Model parameters/ description

Other hyperparameter setting please see the config.py file

Datasets

device_ground_truth.csv

csv format:

last_three_number, time, bias, device_ID

• last_three_number: The last three digits of the air quality sensor's type
• time: Inspection date of the Air Quality Sensors
• bias: abnormal or normal, 1 means abnormal, 0 means normal
• device_ID: ID number of the Air Quality Sensors

temporal_spatio_pm_2_5_144.gz

• a gz file that stores temporal spatio pm2.5 series data of 144 devices.

data format:

time , ID_1, ID_2, ID_3, ... , ID_144
00:00, 18.0, 17.0, 15.0, ... , 27.0
00:01, 17.0, 18.0, 10.0, ... , 33.0
00:02, 19.0, 17.0, 16.0, ... , 19.0
               .
               .

normalized_laplacian_144.npy

• a npy file that stores a normalized laplacian graph matrix with 144 nodes.

temporal_spatio_pm_2_5

• a folder containing 144 csv files.
• each csv file contains temporal spatial pm2.5 series data of 6 devices.
• label column mean pm2.5 series data of the center device
• ID_1, ID_2, ... , ID_5 means the five nearest devices around the center device

each csv format:

time , label, ID_1, ID_2, ... , ID_5
00:00, 18.0,  17.0, 15.0, ... , 27.0
00:01, 17.0,  18.0, 10.0, ... , 33.0
00:02, 19.0,  17.0, 16.0, ... , 19.0
               .
               .

normalized_laplacian

• a folder containing 144 npy files.
• each npy file contains a normalized laplacian graph matrix with 6 nodes.
• each npy file is the graph structure corresponding to the csv file in temporal_spatio_pm_2_5 folder

Project experiment environment

• OS：
  • Distributor ID: Ubuntu
  • Description: Ubuntu 18.04.4 LTS
  • Release: 18.04
  • Codename: bionic
• Python 3.6.9
  • numpy: 1.17.0
  • pandas: 1.1.5
  • torch: 1.8.1
  • sklearn: 0.0
  • tqdm: 4.60.0
  • matplotlib: 3.1.2
  • fire: 0.4.0
  • joblib: 1.0.1

Citation

Please cite our work if you find our work useful in your research.

@ARTICLE{wu23detecting,
  author={Wu, Dennis Y. and Lin, Tsu-Heng and Zhang, Xin-Ru and Chen, Chia-Pan and Chen, Jia-Hui and Chen, Hung-Hsuan},
  journal={IEEE Sensors Journal}, 
  title={Detecting Inaccurate Sensors on a Large-Scale Sensor Network Using Centralized and Localized Graph Neural Networks}, 
  year={2023},
  volume={23},
  number={15},
  doi={10.1109/JSEN.2023.3287270}
}

Reference

1. https://github.com/chenyuntc/pytorch-book
2. https://github.com/nnzhan/Graph-WaveNet
3. https://github.com/Aguin/STGCN-PyTorch
4. https://github.com/locuslab/TCN