Codebase for our project

Project title (Kor): 다양한 시나리오로 확장 가능한 준지도 학습 기반 용해공정 불량 탐지 알고리즘 개발

Project title (Eng): Development of a Semi-Supervised Learning-Based Anomaly Detection Algorithm for Raw Material Processing in Melting Tanks, Scalable Across Various Scenarios.

Contact: [email protected]

We won the grand prize at the 2nd K-Artificial Intelligence Manufacturing Data Analysis Competition!

• Hosted by: 중소벤처기업부(Ministry of SMEs and Startups) | 스마트제조혁신추진단(Smart Manufacturing Innovation Promotion Team) | KAIST

• Competition Link: https://www.kamp-ai.kr/main

• Article Link: 인공지능신문, 한국경제, 연합뉴스, 뉴시스, 뉴스1, ITBizNews

Report & PPT slide

Dataset

• Link: https://www.kamp-ai.kr/aidataDetail?AI_SEARCH=%EC%9A%A9%ED%95%B4%ED%83%B1%ED%81%AC&page=1&DATASET_SEQ=8&EQUIP_SEL=&GUBUN_SEL=&FILE_TYPE_SEL=&WDATE_SEL=

Time series Anomaly Detection (DeepSAD)

This directory contains implementations of anomaly detection framework for real-world time-series data. Our framework is based on the deep semi-supervised learning algorithm, DeepSAD.

Code Explanation

Train / Evaluation --argument {default}

cd DeepSAD

python main.py --net_name {LSTM} --data{custom} --root_path {./} --data_path {data.csv} --timeenc {1} --seq_len {50} --n_features {4} --embedding_dim {32} --eta {1.0} --pretrain {True} --ae_n_epochs {10} --train_epochs {10} --gpu {5}

Key Arguments

--root_path: str, Data storage directory

--xp_path: str, Directory containing experimental results and training history

--data_path: str, Full name of data file

--seq_len: int, Input sequence length

--ae_n_epochs: int, Number of epochs to train autoencoder

--train_epochs: int, Number of epochs to train

--lr: float, learning rate

--n_features: int, Number of features in multivariate time series

--embedding_dim: int, Hidden dimensions

--n_layers: int, Number of neural network layers

--gpu: str, Computation device to use, e.g., 5 -> cuda:5

Time series Forecasting (Autoformer)

This directory contains implementations of time series forecasting framework. Our framework is based on the Transformer, Autoformer.

File Directory

.
├─── DeepSAD
│    └──  base
│       ├── base_net.py
│       └──  base_trainer.py
│    └──  data_provider
│       ├── data_factory.py
│       └── data_loader.py
│    └──  networks
│       ├── lstm.py
│       └── main.py
│    └──  optim
│       ├── ae_trainer.py
│       └── DeepSAD_trainer.py
│    └──  main.py
│    └──  DeepSAD.py
│
├─── Autoformer 
│    └──  data_provider
│       ├── data_factory.py
│       └── data_loader.py
│    └──  exp
│       ├── exp_basic.py
│       └── exp_main.py
│    └──  layers
│       ├── AutoCorrelation.py
│       ├── Autoformer_EncDec.py
│       ├── Embed.py
│       ├── SelfAttention_Family.py
│       └── Transformer_EncDec.py
│    └──  models
│       ├── Autoformer.py
│       ├── Informer.py
│       ├── Reformer.py
│       └── Transformer.py
│    └──  utils
│       ├── masking.py
│       ├── metrics.py
│       ├── timefeatures.py
│       └── tools.py
│    └──  run.py