Anomaly detection

Anomaly detection on a production line using principal component analysis (PCA) and kernel principal component analysis (KPCA) from scratch.

Quick preview

• Authors: Sylvain Combettes, Houssam L'Ghoul
• Date: Oct. 2018 - June 2019
• Context: For our penultimate-year project at Mines Nancy (half a day per week), we did research for the French company Saint-Gobain, the European or worldwide leader in all of its businesses (mainly construction materials). In 2018, Saint-Gobain had a €41.8 billion turnover, operated in 67 countries and had more than 180,000 employees.
• Topic: Detection of sensor failure in a production line.
• Methods: Principal component analysis (PCA) and kernel principal component analysis (KPCA).
• Programming: MATLAB.
• Result: the algorithm can detect 100% of the failure days observed by Saint-Gobain.
• Links: [report incoming]

How to use this repository

• datav3.mat is a file containing data without anomalies
• dataDefautv3.mat is a file containing data with anomalies
• ACP_lineaire_cstr.m is a MATLAB script detecting anomalies in dataDefautv3.mat with comparison to datav3.mat using a (linear) PCA (principal component analysis)
• ACP_non_lineaire_cstr.m is a MATLAB script detecting anomalies in dataDefautv3.mat with comparison to datav3.mat using a (non-linear) KPCA (kernel principal component analysis)

ACP_lineaire_cstr.m and ACP_non_lineaire_cstr.m can be used independently: there are two methods with the same goal.

To note

• I was only able to publish one fourth of the total project, the rest being confidential.
• The MATLAB scripts ACP_lineaire_cstr.m and ACP_non_lineaire_cstr.m are commented in French.
• The report (uploaded soon) is in French.
• The data datav3.mat and dataDefautv3.mat is from "Seongkyu Yoon and John MacGregor. Fault diagnosis with multivariate statistical models part i : using steady state fault signatures. Journal of Process Control, 11(4) :387 – 400, 2001"