ART-2 Neural Network & Machine Monitoring Hybrid System

Matlab tools for Online Clustering

This contribution is a Matlab library implementing the ART-2 Neural Network and the Monitoring Hybrid System which is a complex system of artificial intelligence (ART-2, EMMoG, VBMoG) used to analyze and cluster multidimensional data in real-time. Originally, it was applied to the intelligent monitoring of wind turbines.
ART-2 Neural Network
--------------------

ART-2 network is an unsupervised neural network, based on the adaptive resonance theory (ART). It was introduced by Carpenter and Grossberg in [1].

Art2.m file is a Matlab class representing the model of ART-2 network. The Art2 class members represent network weights and parameters.
The Art2.m depends on an external log4m.m logger class which provides the ability to log network operations and states on a console or into a specified file.
The contribution contains the Art2TestCases.m suite which contains several unit tests based on the examples presented in section 5.3.3 of [2] as well as functional test cases. They confirm the high quality of the implementation and can show an example system running.
The implementation is based on the detailed description of the ART-2 network provided by Fausett in section 5.3 of [2].

An example how to use it is provided here: https://github.com/mjwojcik/Art2MonitoringHybridSystem#how-to-use-it

[1] Carpenter, G. A. & Grossberg S. (1987). ART2: self-organization of stable category recognition codes for analog input pattern. Applied Optics 26, 4919-4930
[2] Fausett L. (1994). Fundamentals of Neural Networks: Architectures, Algorithms, and Applications. Prentice-Hall, Inc., Upper Saddle River, NJ, USA.

Monitoring Hybrid System
------------------------

The Monitoring Hybrid System is a complex system of artificial intelligence (ART-2, EMMoG, VBMoG) used to analyze multidimensional data in real-time. The system was designed to perform an online clustering task as a monitoring of machine states. It can be considered as an early warning tool which is able to notify about potential machine drawbacks. This approach is a data driven algorithm, which decides on a similarity of current data to the data already known by the system. In other words, the data from machine is compared to one of known states and in case when a new state is discovered, a human expert is alarmed.

MonitoringHybridSystem.m file is the Matlab class which is a hybrid of the following components:

* ART-2 neural network – to make general online classification
* stereographic projection and scaling unit – for data preprocessing
* Mixture of Gaussian (MoG) elements – to remember already discovered data areas
* Variational Bayesian Mixture of Gaussian (VBMoG) element – to determine how to set vigilance parameter of the ART-2 network.

I/O: The system input is a stream of multidimensional data points. The system returns numerical identifiers of recognized internal clusters (states) and it generates notifications about new clusters.

The implementation depends on the log4m.m logger class and Michael Chen's libraries of EMMoG and VBMoG algorithms (GaussianMixtureLab package).

The contribution contains the MonitoringHybridSystemTestCases.m suite which contains several unit tests as well as functional test cases. They confirm the high quality of the implementation and can show an example system running.

The core algorithm initializes ART-2 network using VBMoG unit and then manage the running network. If the number of known states is unchanged for specified period of time then it means that ART-2 network is stable and it is recorded as new MoG element. ART-2 network is used only for input points which are not matched by any saved MoG element. The algorithm can also manipulate the ART-2 vigilance parameter. The system inform the user which ART-2 data pattern and which MoG element were used for each input data point. System produces a notification about creating new ART-2 data pattern or new MoG element, what means that a new event was observer in a processing data stream.

An example how to use it is provided here: https://github.com/mjwojcik/Art2MonitoringHybridSystem#how-to-use-it-1

Reference

If you use this tools in some scientific research, please cite the following paper:

Bielecki, A. & Wójcik M. (2017)
Hybrid system of ART and RBF neural networks for online clustering.
Applied Soft Computing, vol. 58, 1-10, ISSN 1568-4946

The research was supported by the National Centre for Research and Development in Poland, grant number WND-DEM-1-153/01.

The system implementation is a part of the Mateusz Wójcik’s PhD thesis - “Hybrid neural system for intelligent monitoring of wind turbine”.