Uncertainty-based fault detection - Revision history

Mohbou: Replaced content with "Removed."

2018-10-17T13:33:12Z

Replaced content with "Removed."

Slawek at 18:02, 14 October 2018

2018-10-14T18:02:07Z

Slawek at 10:30, 26 October 2016

2016-10-26T10:30:07Z

Mohbou: Created page with "{{StudentProjectTemplate |Summary=Introducing uncertainty for deviation detection based on on-board vehicle data |Keywords=Data Mining, Anomaly Detection, Uncertainty, Unsuper..."

2015-10-12T22:27:29Z

Created page with "{{StudentProjectTemplate |Summary=Introducing uncertainty for deviation detection based on on-board vehicle data |Keywords=Data Mining, Anomaly Detection, Uncertainty, Unsuper..."

New page

{{StudentProjectTemplate
|Summary=Introducing uncertainty for deviation detection based on on-board vehicle data
|Keywords=Data Mining, Anomaly Detection, Uncertainty, Unsupervised Learning
|TimeFrame=Spring 2016
|Prerequisites=Learning Systems and Data Mining courses
|Supervisor=Mohamed-Rafik Bouguelia, Sławomir Nowaczyk
|Level=Master
|Status=Open
}}
In a previous project we have collected a big amount of data from 19 Volvo buses operating in Kungsbacka. This data represents on-board signals such as vehicle speed, air pressure etc, as well as a database of Vehicle Service Records (VSR) containing repairs that had been performed on the vehicles during workshops.

Some methods have already been proposed for online unsupervised deviation detection in this data. Basically, a daily signal of the vehicle is represented in some suitable representation (e.g. as a histogram) and compared to the fleet (the other vehicles) in order to check if it is deviating or not. In order to evaluate the deviation detection method, the detected deviations are compared to the repairs preformed on the vehicle in the VSR. A detected deviation which disappears just after a repair indicates that we correctly detected a fault before it occurs.

However, the current deviation detection methods and their evaluation, do not take into account "uncertainty". The goal of the master's project is to introduce uncertainty at different levels:

At the deviation detection level:
- When representing the raw data (signal) with histograms, there is some uncertainty regarding the amount of data which is used to compute the histograms. A smaller amount of data introduces more uncertainty in the computed histogram.
- When comparing a vehicle V against the fleet, we are assuming (or hoping) that the vehicles in the fleet are behaving normally (not deviating). However, some vehicles in the fleet may be deviating. Therefore, we need to introduce different uncertainties when comparing V against each vehicle in the fleet.

At the evaluation level:
The VSR data contain many inaccuracies. For example, the indicated date of repair may be shifted with several days or weeks from the actual repair date. Therefore, uncertainty may be introduced in the evaluation of the method. This can be based on the combination of different criteria such as:
- The vehicle mileage at the time of the repair.
- Changes in the signal before and after the indicated date of repair.

← Older revision		Revision as of 13:33, 17 October 2018
Line 1:		Line 1:
−	~~{{StudentProjectTemplate~~	+	Removed.
−	~~\|Summary=Introducing uncertainty for deviation detection based on on-board vehicle data~~
−	~~\|Keywords=Data Mining, Anomaly Detection, Uncertainty, Unsupervised Learning~~
−	~~\|TimeFrame=Fall 2018~~
−	~~\|Prerequisites=Learning Systems and Data Mining courses~~
−	~~\|Supervisor=Mohamed-Rafik Bouguelia, Sławomir Nowaczyk~~
−	~~\|Level=Master~~
−	~~\|Status=Open~~
−	}}
−	In a previous project we have collected a big amount of data from 19 Volvo buses operating in Kungsbacka. This data represents on-board signals such as vehicle speed, air pressure etc, as well as a database of Vehicle Service Records (VSR) containing repairs that had been performed on the vehicles during workshops.
−
−	Some methods have already been proposed for online unsupervised deviation detection in this data. Basically, a daily signal of the vehicle is represented in some suitable representation (e.g. as a histogram) and compared to the fleet (the other vehicles) in order to check if it is deviating or not. In order to evaluate the deviation detection method, the detected deviations are compared to the repairs preformed on the vehicle in the VSR. A detected deviation which disappears just after a repair indicates that we correctly detected a fault before it occurs.
−
−	~~However, the current deviation detection methods and their evaluation, do not take into account "uncertainty". The goal of the master's project is to introduce uncertainty at different levels:~~
−
−	~~At the deviation detection level:~~
−	- When representing the raw data (signal) with histograms, there is some uncertainty regarding the amount of data which is used to compute the histograms. A smaller amount of data introduces more uncertainty in the computed histogram.
−	- When comparing a vehicle V against the fleet, we are assuming (or hoping) that the vehicles in the fleet are behaving normally (not deviating). However, some vehicles in the fleet may be deviating. Therefore, we need to introduce different uncertainties when comparing V against each vehicle in the fleet.
−
−	~~At the evaluation level:~~
−	The VSR data contain many inaccuracies. For example, the indicated date of repair may be shifted with several days or weeks from the actual repair date. Therefore, uncertainty may be introduced in the evaluation of the method. This can be based on the combination of different criteria such as:
−	~~- The vehicle mileage at the time of the repair.~~
−	~~- Changes in the signal before and after the indicated date of repair~~.

@@ Line 2: / Line 2: @@
 |Summary=Introducing uncertainty for deviation detection based on on-board vehicle data
 |Keywords=Data Mining, Anomaly Detection, Uncertainty, Unsupervised Learning
-|TimeFrame=Spring 2017
+|TimeFrame=Fall 2018
 |Prerequisites=Learning Systems and Data Mining courses
 |Supervisor=Mohamed-Rafik Bouguelia, Sławomir Nowaczyk