This is the implementation of [1], which makes use of simulated data (reproducible from these codes) as well as real data [2]. For ORION-AE data description, check [2] and its Dataverse repository.

GMMSEQ: Gaussian Mixture Model with sequentially appearing clusters

GMMSEQ is a new clustering initially developed to cope with specificities of acoustic emission (AE) signals. But it works with general time-series as soon as they are described by ordered feature vectors with timestamps.

GMMSEQ is a generative model that includes the fact that features are indexed by continuous timestamps. The proportions in the GMM are assumed to vary in time according to a model of evolution based on sigmoid functions to represent cumulative occurrence of clusters.

For structural health monitoring where acoustic emissions are used to detect damages, each sigmoid function allows us to represent:

• The level of activation of a given damage related to the cumulative number of signals generated by this damage;
• The growth rate of the damage driven by the slope of the sigmoid function at the origin and related to the kinetics of the damage;
• The instant of the damage onset.

Therefore, GMMSEQ makes it possible to identify when a damage first occurs (onset), how it grows (kinetics) and how it accumulates (cluster progression).

The current implementation supports Gaussian distribution, but soon other distributions will be added.

Getting Started

Add some folders to the MATLAB path:

$ pathdefmixmod

The implementation is based on three main codes:

1. GMMSEQ_init.m: to initialize a GMMSEQ model.
2. GMMSEQ_train.m: to estimate the parameters of a GMMSEQ model from data.
3. GMMSEQ_test.m: to test a GMMSEQ model on new data.

Their use is illustrated with as many comments as possible in main_test_simulated_data.m. Some illustrations are provided with real data [2] in main_test_real_data.m.

Reproduce the results of the paper

Have a look in repository tests:

• main_test_simulated_data_withCriteres_v2.m allows to focus on Section 1 of the tests.
• test_selection_models_GMMSEQ_via_histTau.m allows to focus on Section 2 of the tests. Note that here you need your own features for this part. For information, you can reproduce results of [1] by using features provided in the Drive's link below.

The set of .mat files available at the repository link represent a set of features obtained from raw series of measurements in ORION-AE [2]. Features have been obtained by a method described in [1], in Section "Signal Processing" relying on [3]. These features matrices (using variables "P3") allow to reproduce the figures of the paper.

Cite

If you make use of this code in your work, please refer to [1]:

@article{1,
      title={Clustering acoustic emission data stream with sequentially appearing clusters using mixture models}, 
      author={Emmanuel Ramasso and Thierry Denoeux and Gael Chevallier},
      year={2021},
      journal={Mechanical Systems and Signal Processing},
      note={Submitted},
      publisher={Elsevier}
}

If you make use of the data, please refer to [2]:

@article{2,
  title={Monitoring a Bolted Vibrating Structure Using Multiple Acoustic Emission Sensors: A Benchmark},
  author={Emmanuel Ramasso and Benoit Verdin and Gael Chevallier},
  journal={Data},
  volume={7},
  number={3},
  pages={31},
  year={2022},
  publisher={Multidisciplinary Digital Publishing Institute}
}

If you make use of the features, please refer to [3]:

@article{3,
  title={A signal processing approach for enhanced Acoustic Emission data analysis in high activity systems: Application to organic matrix composites},
  author={Kharrat, Mohamed and Ramasso, Emmanuel and Placet, Vincent and Boubakar, ML},
  journal={Mechanical Systems and Signal Processing},
  volume={70},
  pages={1038--1055},
  year={2016},
  publisher={Elsevier}
}