Welcome to the semi-supervised learning for early detection and diagnosis of AHU faults project page!


Motivation of the this work:

This work is motivated by a real world FDD (fault detection and diagnosis) situation when faulty training samples are rarely available for data-driven methods. For supervised data-driven FDD for AHU, chiller or any other components of the HVAC system,adequate training samples for both positive and negative sides are highly demanded. Not enough faulty training samples may cause fatal failure for existing supervised approaches.

The proposed semi-supervised solves the above problem and is able to retrieve confidently selected faulty testing samples into the training pool to enlarge the faulty training sample dataset size. Furthermore, three tradeoffs are investigated to verify the minimal initial faulty training dataset size.


Flowchart of the proposed method:


Experimental results: The number of faulty training samples v.s. the final classification accuracy.


Open-source codes:

Source code of the Semi-supervised Learning for AHU faults.

Minimum system requirements:

1. Matlab 2016a

2. LibSVM-3.2.1 (Chang, Chih Chung, and C. J. Lin. "LIBSVM: A library for support vector machines." 2.3(2011):1-27.)

3. Genetic Algorithm Toolbox (Fleming, Peter, H. Pohlheim, and C. Fonseca. "Genetic algorithm toolbox for use with MATLAB." (1994).)


Back to Yan Ke's Homepage..