Please refer to '/lib/dataloader_...' for more details of data loading and preprocessing
1.1 SWaT and WADI
- train and test data in '.csv' with the last column as labels (0 for normal and 1 for abnormal)
- we adopt the data version of SWaT with 45 features
1.2 SMD and MSL
- train and test data in '..._train.pkl' and '..._test.pkl' formed as arrays with the shape (samples, features), for example: 'machine-1-1_train.pkl' or 'MSL_train.pkl'
- test labels also in '...test_label.pkl formed as arrays with the shape (samples,)
1.3 Unsupervised Datasets
- the training and test data should be concatenated together and saved in '/.../.npz', where label 'a' are x-samples with the shape (samples, features) and label 'b' are y-samples with the shape (samples,), you can refer to '/data/unsupervised_data/test_data_smd_unsup.npz'
1.4 Your Datasets
- you can use '/lib/dataloader_...' according to your data format or create new dataloaders
- data: dataset name, such as 'SWaT', 'WADI', ...
- temp_methods: for STAMP, keep it as 'SAttn'
2.1 Graph Structure
- nnodes：number of features, which varies across datasets
- top-k：number of neighbor nodes
2.2 Pred Model
- window_size：time window length
- n_pred：prediction step
- temp_kernel：the kernel size of Convolutional Input-Output Layer
- layer_num：number of TLL and SLL Layers
- act_func：activation function
2.3 Attention
- embed_size: embedding size
- num_heads: number of attention heads
- is_conv: while True, the Feed-Foward Network(FFN) adopts Linear layers; else, FFN will adopt Conv1D layers
2.4 AE
- latent_size：dimension of L-space
2.5 Training
- is_down_sample：perform down-sampling to the original samples or not
- down_len：down-sampling ratio
- is_mas：perform a moving average operation by sub-windows to extend channel or not
2.6 Testing
- test_alpha：weight of prediction error
- test_beta：weight of reconstruction error
- test_gamma：weight of adversarial error
2.7 params in get_final_result()
- topk: number of features of calculating the anomaly score
- option: set to 2
- method: ['sum', 'max', 'mean'], types of aggragation operators
3.1 Run STAMP
python run.py --down_len 1 --epoch 5 --data SMD --nnodes 38 --window_size 15 --n_pred 3 or python run.py --down_len 100 --epoch 30 --data SWaT --nnodes 45 --window_size 15 --n_pred 3 or python run.py --down_len 100 --epoch 20 --data WADI --nnodes 127 --window_size 15 --n_pred 3 or python run.py --down_len 1 --epoch 20 --data MSL --nnodes 55 --window_size 15 --n_pred 3
3.2 check the saved model weights in '/expe'
3.3 Evaluation
python test.py --down_len 1 --data SMD --nnodes 38 --window_size 15 --n_pred 3 --test_alpha 0.5 --test_beta 0.1 --test_gamma 0.4
or python test.py --down_len 100 --data SWaT --nnodes 45 --window_size 15 --n_pred 3 --test_alpha 0.8 --test_beta 0.1 --test_gamma 0.1
or python test.py --down_len 100 --data WADI --nnodes 127 --window_size 15 --n_pred 3 --test_alpha 0.1 --test_beta 0.1 --test_gamma 0.8
or python test.py --down_len 1 --data MSL --nnodes 55 --window_size 15 --n_pred 3 --test_alpha 0.1 --test_beta 0.8 --test_gamma 0.1
4.1 Get Model-Derived information
python get_model_information.py --down_len 1 --epoch 5 --data SMD --nnodes 38 --window_size 15 --n_pred 3 --test_alpha 0.5 --test_beta 0.1 --test_gamma 0.4
or python get_model_information.py --down_len 50 --epoch 10 --data SWaT --nnodes 45 --window_size 15 --n_pred 3 --test_alpha 0.8 --test_beta 0.1 --test_gamma 0.1
4.2 check the saved model information in '/weights'
4.3 Screening Based on Model Information
python /unsupervise/Screening.py
4.4 Run STAMP
python run_unsup.py --down_len 1 --epoch 5 --data SMD --nnodes 38 --window_size 15 --n_pred 3
or python run_unsup.py --down_len 50 --epoch 30 --data SWaT --nnodes 45 --window_size 15 --n_pred 3
4.5 Evaluation
python test_unsup.py --down_len 1 --data SMD --nnodes 38 --window_size 15 --n_pred 3 --test_alpha 0.5 --test_beta 0.1 --test_gamma 0.4
or python test_unsup.py --down_len 50 --data SWaT --nnodes 45 --window_size 15 --n_pred 3 --test_alpha 0.8 --test_beta 0.1 --test_gamma 0.1