Predictive Building System Maintenance -- Anomaly Detection

(For simplified demo on Google Colab, save "APS490" folder to your Google Drive or view it here: https://drive.google.com/file/d/1FwafCGFpnnft0bUER9L9gcMgOXJAw_sW/view?usp=sharing).

1. Recommanded environment setting

1.1 install miniconda3

a. Install Miniconda:

(Linux OS; installation in another os could be found in the link above)

   curl 'https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh' > Miniconda.sh
   bash Miniconda.sh
   rm Miniconda.sh

b. Close and re-open your terminal session.

c. Change directories (cd) to where you cloned this repository.

d. Initialize conda

$ conda init

1.2 create an environment inside the conda

$ conda env create -f environment.yml

1.3 activate the environment

$ conda activate aps490

2. Training

• You may use the following command to train the model.

  $ python -m predection.main train

  Keyword arguments:

• (optional)data_root: training dataset directory root. Default: ./prediction/datasets/train/
• (optional)output_root: output directory root. Default: ./output/
• (optional)seed: random seed. Default: 42
• (optional)seq_length: squence length of the sliding window. Default: 48
• (optional)pred_length: prediction length of the sliding window. Default: 24
• (optional)epoch_num: number of epochs. Default: 1000
• (optional)learning_rate: learning rate of the adam optimizer. Default: 0.0008
• (optional)hidden_size: hidden_size of RNN . Default: 7
• (optional)num_layers: num_layers of RNN. Default: 1
• (optional)loss: training loss function. (mae, mse, or huber_<delta>) Default: huber_0.022
• (optional)model: RNN model type (lstm or gru). Default: gru

Example:

$ python -m prediction.main train --output_root=./output/

• UI Support:

  $ streamlit run train_app.py

3. Validate

You may use the following command to train the model.

$ python -m prediction.main validate --output_root=<output directory root> --checkpoint_path=<checkpoint path>

Keyword arguments:

• checkpoint_path: path to the model that will be evaulated
• (optional)output_root: the output directory root. Default: ./output/
• (optional)data_root: validation dataset directory root. Default: ./prediction/datasets/valid/
• (optional)seed: random seed. Default: 42
• (optional)seq_length: squence length of the sliding window. Default: 48
• (optional)pred_length: prediction length of the sliding window. Default: 24

Example:

$ python -m prediction.main validate --output_root=./output/ --checkpoint_path=./output/model.pth

4. UI main page

Visualize the prediction and anomaly analysis

$ streamlit run stream_app.py