Matrix Profile Paper

The Matrix Profile has the potential to revolutionize time series data mining because of its generality, versatility, simplicity and scalability. In particular it has implications for time series motif discovery, time series joins, shapelet discovery (classification), density estimation, semantic segmentation, visualization, rule discovery, clustering etc.

Project Organization

• src: Folder containing the contextual matrix profile codebase (i.e., folder cmp) and necessary packages (i.e., distancematrix).
• tests: Contains Python-based test cases to validate source code.
• pyproject.toml: Contains metadata about the project and configurations for additional tools used to format, lint, type-check, and analyze Python code. ( See here for reference)

Getting started

Create virtual environment and activate it and install dependencies:

• Makefile

  make setup

• Linux:

  python3 -m venv .venv
  source .venv/bin/activate
  pip install -r requirements.txt

• Windows:

  python -m venv venv
  venv\Scripts\activate
  pip install -r requirements.txt

Run the main script through the console:

source .venv/bin/activate
python -m src.cmp.main

You should see in the terminal a message output like the following:

CONTEXT 1 : Subsequences of 05:45 h (m = 23) that start in [00:00,01:00) (ctx_from00_00_to01_00_m05_45)
99.997%        0.0 sec

- Cluster 1 (1.181 s)   -> 1 anomalies
- Cluster 2 (0.508 s)   -> 3 anomalies
- Cluster 3 (0.473 s)   -> 4 anomalies
- Cluster 4 (0.658 s)   -> 5 anomalies
- Cluster 5 (-)         -> no anomalies green

At the end of the execution you can find the results in the results folder.

Run using docker

Build the docker image.

• Makefile

  make docker-build

• Linux:

  docker build -t cmp .

Run the docker image.

• Makefile

  make docker-run

• Linux:

  docker run cmp

At the end of the execution you can find the results in the results folder.

Context definition

# 2) User Defined Context
# We want to find all the subsequences that start from 00:00 to 02:00 (2 hours) and covers the whole day
# In order to avoid overlapping we define the window length as the whole day of
# observation minus the context length.

# - Beginning of the context 00:00 AM [hours]
context_start = 17

# - End of the context 02:00 AM [hours]
context_end = 19

# - Context time window length 2 [hours]
m_context = context_end - context_start  # 2

# - Time window length [observations]
# m = 96 [observations] - 4 [observation/hour] * 2 [hours] = 88 [observations] = 22 [hours]
# m = obs_per_day - obs_per_hour * m_context
m = 20  # with guess

# Context Definition:
# example FROM 00:00 to 02:00
# - m_context = 2 [hours]
# - obs_per_hour = 4 [observations/hour]
# - context_start = 0 [hours]
# - context_end = context_start + m_context = 0 [hours] + 2 [hours] = 2 [hours]
contexts = GeneralStaticManager([
    range(
        # FROM  [observations]  = x * 96 [observations] + 0 [hour] * 4 [observation/hour]
        (x * obs_per_day) + context_start * obs_per_hour,
        # TO    [observations]  = x * 96 [observations] + (0 [hour] + 2 [hour]) * 4 [observation/hour]
        (x * obs_per_day) + (context_start + m_context) * obs_per_hour)
    for x in range(len(data) // obs_per_day)
])

Cite

You can cite this work by using the following reference or either though this Bibtex file.

Chiosa, Roberto, et al. "Towards a self-tuned data analytics-based process for an automatic context-aware detection and diagnosis of anomalies in building energy consumption timeseries." Energy and Buildings 270 (2022): 112302.

Contributors

• Roberto Chiosa

License

This code is licensed under the MIT License - see the LICENSE file for details.