Evolving Markov Chains

Introduction

Evolving Markov Chains (EMC) is an online method for mode discovery and recognition. This repository contains the implementation in Python and the necessary utilities to evaluate results.

The experiments contain:

• synthetic data (probability tracking, change point detection, mode discovery and recognition)
• human activity recognition through accelerometer readings
• electric motor condition monitoring through vibration data
• eye-state detection through EEG signals

Installation

To install the emc package, you can use conda:

conda env create -f environment.yaml -n emc
conda activate emc
pip install -e .

Usage

Here is how to run EMC on a sequence:

from emc.estimator.EMC import EMC

emc_ins = EMC(
    alpha=4,
    order=1,
    lambda_=[0.94, 0.98],
    beta=0,
    delta=[0.3, 0.2],
    eta=[0.4, 0.3],
    tau=50,
)
emc_ins.process_sequence(sequence, progress=True)

Synthetic data generation

from emc.generator.markov import MarkovianSwitchingSystem

mss_ins = MarkovianSwitchingSystem(
    alphabet_cardinality=4,
    mode_process_order=1,
    subprocess_order=1,
    number_of_subprocesses=5,
    rng_or_seed=42
)
mss_ins.generate_subprocess_sequence(
    number_of_regimes=10,
    avoid_loops=True
)
symbol_sequence = mss_ins.generate_symbol_sequence(
    regime_length_bounds=[500,1000]
)

Reproducing Experiment Results

Source code for experiments is located in exps directory with the following structure:

• synthetic
  • 01_pt: probability tracking
  • 02_cpd: change point detection
  • 03_mdr: mode discovery and recognition
• har: human activity recognition on RealWorld (HAR) dataset (link)
• cwru: electric motor condition monitoring and anomaly detection on CWRU dataset (link)
• eeg: eye-state detection on EEG signals (link)

Each directory contains the necessary steps to get the results shown in the paper.

Experiments on Synthetic Data

The data is not included in the repo but can be generated using the generate_data.py file.

python exps/synthetic/01_pt/generate_data.py

The second step is to run hyperparameter optimization. The values used in the paper are included in /output/params.json file. Therefore this step can be skipped (optimization takes some time). If needed, it can be run again by:

python exps/synthetic/01_pt/optimize_params.py

Finally, using the data and the parameter values, the test runs are executed as follows:

python exps/synthetic/01_pt/test.py

The resuls will be saved to output/results directory.

Human Activity Recognition Experiments

To execute test runs using the included alphabet and hyperparameter values, simply run:

python exps/har/test.py

The resuls will be saved to output/results directory.

If needed, the alphabet construction and hyperparameter optimization steps can be run as:

python exps/har/build_alphabet.py
python exps/har/optimize_params.py

Electric Motor Condition Monitoring Experiments

To execute test runs using the included alphabet and hyperparameter values, simply run:

python exps/cwru/test.py

The resuls will be saved to output/results directory.

Likewise, the alphabet construction and hyperparameter optimization steps can be run as:

python exps/cwru/build_alphabet.py
python exps/cwru/optimize_params.py

EEG Eye-State Detection Experiments

The microstate sequence and hyperparameters are included in the repo. Hence, to execute the test runs, simpy run:

python exps/eeg/test.py

Similarly, the preprocessing and hyperparameter optimization steps can be run as:

python exps/eeg/preprocess.py
python exps/eeg/optimize_params.py

License

This project is licensed under the MIT License.