A Novel Feature Extraction Method for Monitoring (Vehicular) Fuel Storage System Leaks

Fling (Finn) Tseng, Imad Makki, Dimitar Filev, and Ratna Babu Chinnam
Submission Type: 
Full Paper
phmc_14_062.pdf370.57 KBSeptember 19, 2014 - 9:48am

System state determination with incomplete sensory information set proved to be a technically challenging problem. In this paper, authors tackle a problem of this type associated with vehicle fuel storage systems. Federal and state regulations require fuel storage leak detection mechanism to be conducted periodically and regulate its execution rate and performance to ensure effective emission controls. Being able to robustly determine a fuel storage system’s state in terms of its effectiveness of fuel containment is therefore of great importance to all vehicle OEMs. Prevailing practice in the industry is to utilize a method relevant to natural vacuum phenomenon and is loosely associated with ideal gas law. While highly effective, such monitoring schemes require stringent execution “entry condition” evaluations to ensure ideal test conditions are met such that impact of various noise factors is minimized. Differences in ambient conditions compounded with varying customer drive cycle patterns present great challenge to existing monitor designs for the purpose of leak detection. All this calls for an alternative that is less sensitive to known noise factors yet provide similar detection capability.
Different from common practice of carrying out a system monitoring task with observable sensory information directly, the authors performed data analysis focused on signatures revealed from probability density curve of a signal derivative of in-tank pressure. This is one of the signals typically kept “alive” during leak detection monitoring phase when the engine is off. More specifically, we developed a non-parametric method to continuously extract signatures indicative of a leak in the sealable system.
To reduce the requirement on computing resources, we also propose a simplified method for obtaining signatures from the probability density curve of interest. In conclusion, signal distribution based signatures/features are derived to serve as inputs to a threshold based classification rule model. The approach not only yielded very promising results but also minimized entry conditions, allowing more frequent test execution and greater compliance.

Publication Year: 
Publication Volume: 
Publication Control Number: 
Page Count: 
Submission Keywords: 
feature extraction
Fuel Storage
Leak Detection
Statistical Distribution
Submission Topic Areas: 
Data-driven methods for fault detection, diagnosis, and prognosis
Industrial applications
Submitted by: 

follow us

PHM Society on Facebook Follow PHM Society on Twitter PHM Society on LinkedIn PHM Society RSS News Feed