Semantic State Prediction with Power Consumption Data

Semantic state prediction with power consumption data collected by smart plugs.

🖥️ Code:: Usage

pip install -r requirements.txt

(TBU)

Directory Structure

The overall structure of the project is as follows:

semanticProject/
    ├── channel/
    ├── data/
    ├── dataloader/
    ├── docs/
    ├── models/
    ├── outputs/
    ├── scripts/
    ├── trainer/
    ├── utils/
    ├── config.py
    ├── README.md
    └── requirements.txt

• It is recommended not to modify the directory structure.
• If you wish to make changes, please ensure to update the paths in the config.py file:
  • self.root_path
  • self.save_path

Configuration

The config.py file allows you to customize various aspects of the project through command-line arguments. Below is a categorized summary of the arguments you can use:

Environment Arguments

• --debug
  Enable debug mode. (default: True)

• --device
  Device to be used for training (e.g., CPU, GPU). (default: 'cuda')

• -rand, --random_state
  Random state for reproducibility. (default: 42)

• --data_dir_path
  Path to the data directory. (default: <root_path>/data)

• --save_data_path
  Path to save output datasets. (default: <save_path>/dataset)

• --save_vis_path
  Path to save visualizations. (default: <save_path>/visualization)

• --save_model_path
  Path to save model checkpoints. (default: <save_path>/checkpoints)

Data Arguments

• -dc, --num_dc
  Number of data collections. (choices: [1, 2, 3, 4, 5], default: 4)

• -s, --scaler_type
  Type of data scaler. (choices: ["standard", "minmax"], default: "standard")

• -eq, --make_equal_dist
  Ensure equal data distribution. (default: True)

• -agg, --aggregate
  Use aggregated dataset. (default: True)

Model Arguments

• -i, --input_size
  Input size for the model. (choices: [2, 7], default: 2)

• -c, --num_classes
  Number of output classes for the model. (default: 343)

• -b, --batch_size
  Mini-batch size for training. (default: 32)

• -st, --stride
  Stride length for data. (default: 1)

• -nepoch, --num_epochs
  Number of training epochs. (default: 500)

• -lr, --learning_rate
  Learning rate for the optimizer. (default: 0.001)

• -p, --patience
  Early stopping patience. (default: 15)

• -vs, --val_size
  Validation set size as a fraction of the dataset. (default: 0.1)

• -ts, --test_size
  Test set size as a fraction of the dataset. (default: 0.2)

• -seq_len, --sequence_length
  Sequence length for training. (default: 50)

• --class_names
  List of class names. (default: [])

• --model_name
  Name of the deep learning model. (choices: ["CNN-RNN", "CNN-LSTM", "DAE-RNN", "DAE-LSTM"])

Embedding Arguments

• --embedding_dim
  Embedding dimension for semantic embeddings. (default: 50)

Channel Arguments

• --snr_db
  Signal-to-noise ratio (SNR) in dB. (default: 20)

• --channel_name
  Channel environment type. (choices: ["AWGN", "Rayleigh", "Rician", "Nakagami"], default: "AWGN")

• --doppler_freq
  Doppler frequency. (default: 0)

• --k_factor
  K-factor of the Rician channel. (default: 1.0)

• --variance
  Variance of the Rician channel. (default: 1.0)

• --m_factor
  M-factor of the Nakagami channel. (default: 1.0)

• --omega
  Average power (omega) of the Nakagami channel. (default: 1.0)

Make sure to use these arguments as needed when running scripts to configure the project behavior.

🖥️ Code:: Explanation

(TBU)

🗃️ Data Collection:: 2nd Test Data Acquisition

🔎 Notes

Behavior Details for Data Collection

• Monitor: Put into sleep mode when not in use.
• Fridge: Simulate real usage by opening and closing at specific intervals repeatedly.
• Camera: Use recordings for labeling purposes.

Additional Considerations

• Collect data for over an hour to verify energy (Wh) values.
• Use average power consumption of each user's monitor and PC as a signature.
• Leverage intermediate data (e.g., door, power) instead of using aggregated power data directly.

Issues Identified

• Jee-in's PC could not enter sleep mode during data collection, so "working w/o PC" and refrigerator or meeting-related tasks were not power-efficient.
  • Improvement: Use another PC for data collection in future tests.
• Hae-won's PC often failed to turn on properly.

Experimental Setup

Type	Content
Date	November 28, 2024
Time	3:15 PM - 4:30 PM
Participants	Kim Jee-in, Lee Hae-won, Kim Yu-rim
Location	GIST Building C10 Room 207
Behavior Labels	6 types
Smart Meters	7 units
Base Sensors	2 units
Reference	Camera

Behavior and Device Mapping

Behavior	Jee-in	Hae-won	Yu-rim	Notes
Working with PC	SMP04 SMP05	SMP03 SMP14	SMP12	Yu-rim: Using a charging laptop.
Working without PC	-	-	-	Ji-in only, PCs set to sleep mode.
Lunch / Coffee Break	SMP01	SMP01	SMP01	Fridge usage simulation: open 30s, close 1 min, repeat. PCs in sleep mode (except Ji-in).
Having Meeting	SMP02	SMP02	SMP02	10 min with projector and laptop. PCs in sleep mode (except Ji-in).
Go to Work	Door Sensor	Door Sensor	Door Sensor	-
Get Off Work	Door Sensor	Door Sensor	Door Sensor	-

Scenario

🗃️ Data Collection:: 4th Test Data Acquisition

🔎 Notes

• 15:25 - 15:35: During Jee-in's "Lunch" activity, movement detected near the front bookshelf.
• 15:35: Hae-won exits during "Absence" activity, with a slight pause.

Experimental Setup

Type	Content
Date	January 8, 2025
Time	2:00 PM - 4:00 PM
Participants	Kim Jee-in, Lee Hae-won
Location	GIST Building C10 Room 207
Behavior Labels	4 types
Smart Meters	5 units
Base Sensors	2 units
Reference	Camera

States

• S0 ~ S11 (12 states)

Index	State
S0	(abs, abs)
S1	(wpc, abs)
S2	(abs, wpc)
S3	(wpc, wpc)
S4	(wopc, abs)
S5	(abs, wopc)
S6	(wopc, wpc)
S7	(wpc, wopc)
S8	(l, abs)
S9	(abs, l)
S10	(l, wpc)
S11	(wpc, l)

State Machine Graph

Scenario

• Duration: 120 minutes (2 hours)
• Data Collection Per State: 10 minutes each (to address data imbalance)

Time (min)	5	5	5	10	10	10	10	10	5	10	10	10	10	5	5
Jee-in	abs	wpc	wpc	wpc	wpc	abs	abs	abs	wpc	wopc	l	l	wopc	wpc	abs
Hae-won	abs	abs	wpc	wopc	l	l	wopc	wpc	wpc	wpc	wpc	abs	abs	abs	abs
State	S0	S1	S3	S7	S11	S9	S5	S2	S3	S6	S10	S8	S4	S1	S0

Behavior Details

• Before "Lunch," PCs are set to sleep mode.
• Working without PC: PCs are in sleep mode.
• Absence: PCs and monitors are powered off.
• Lunch: Fridge and freezer doors are opened, kept open for 1 minute, then closed for 4 minutes.
• Working with PC: Model training is conducted in the background, utilizing the CPU (Jee-In).

Data Visualization

🗃️ Data Collection:: 5th Test Data Acquisition