Github repo for the research paper titled "Integrating Adaptive Moving Window and Just-in-Time Learning Paradigms for Soft-Sensor Design"

The proposed method, MW_Adp-JITL is implemented in Matlab; both the code and the simulated datasets we used in our experiments are freely available under a GNU GPL 3 open source license from this repo. You can find more details about the algoritm in the manuscript published in Neurocomputing.

If you use the code or the simulated datasets, please cite our corresponding paper: Integrating Adaptive Moving Window and Just-in-Time Learning Paradigms for Soft-Sensor Design

@article{adpsensor20,    
  Author = "Aysun Urhan and Burak Alakent",
  Title = "Integrating Adaptive Moving Window and Just-in-Time Learning Paradigms for Soft-Sensor Design",
  Journal = "Neurocomputing",
  Year = "2020",
  issn = "0925-2312",
  doi = "https://doi.org/10.1016/j.neucom.2020.01.083",
  url = "http://www.sciencedirect.com/science/article/pii/S0925231220301417",
}

• I tried my best to clean up the code. But keep in mind that the code has gone through countless changes during my MS thesis, so you might see a few unnecessary variables defined here and there. Any feedback on how to improve the program and make it run faster is appreciated :)

• Please open up an issue asap if you encounter any errors/bugs!! I'd be more than glad to help debug.

• Our method is based on relevance vector machine (RVM), I used the SparseBayes software version 2.0, provided by Tipping himself. I had to modify the code in SB2_FullStatistics.m at line 105 to make sure that matrix A is positive definite. I used nearestSPD.m to get the nearest PD matrix.

• You can find our implementation of MWAdp and MWAdpJITL methods on debutanizer column data in Demo.m file.

• TODO: Implement MW\sub{Adp}-JITL in Python, to use from the commandline and upload it to conda/pypi as a package.

20 simulations runs were conducted for 8 different concept drift models (CDM). Details of the simulation models can be found in the article. Each CDM is stored as a struct with "X" and "Y" arrays. For a total of 700 observations, 19 predictors and their 2 lagged measurements are included in X (there are 19x3 = 57 predictors/columns in total in X) and the response variables (concentration of product B) is in Y.