Classification of the health of diesel engines using sparse linear discriminant analysis

Neha R Chandrachud, Purdue University

Abstract

The aim of this thesis is to generate a classifier to classify the state of a diesel engine as healthy or faulty from steady-state data acquired from the engine. The motivation behind the work is the necessity to develop an on-board diagnostic tool due to increased regulations on different kinds of diesel emissions, such as Nitrogen Oxide and Nitric Oxide (collectively called NOx), Particulate Matter (PM) and Carbon Monoxide (CO). This on-board diagnostic system must detect and diagnose a fault using minimum resources such as sensors to reduce maintenance, troubleshooting costs and downtime. It also must detect the faults accurately having the least number of false positives. Thus, the objective of the work is to build a classification model that classifies the engine state at different steady-state operating points from the data acquired from an optimum number of sensors on the engine. The classifier developed in this thesis is based on Sparse Discriminant Analysis. It is a technique performing Linear Discriminant Analysis with a sparseness criterion imposed such that classification, dimension reduction and feature selection are merged into one step. The data mainly comes from the air handling loop which is a critical loop for the operation of a diesel engine. The fault condition considered in this work is the Exhaust Gas Recirculation (EGR) Cooler fouling. The above mentioned technique was first applied to some data sets that were acquired from the engine using the standard AVL8 test cycle. The classification accuracy was found to be 68%. To improve the classification performance, data at predefined steady-state regions was extracted from the original data. Using the technique mentioned above, the engine state was successfully classified into healthy or faulty with zero false positives and 100% classification accuracy for the fault when the engine was at idle. Apart from the classification, minimum number of sensors required were found to achieve the above mentioned result. It was found that only two variables are needed to classify the engine state accurately at idle. The two signals that were found significant were Charge Temperature and Exhaust Temperature. The classifier was also developed at several other steady-state points that are common operating points for a diesel engine. However, the results at those other points were not as successful as the above mentioned ones. To accommodate all the points, a single classifier was built which had 81% classification accuracy. A possible reason for more misclassifications at other points was suspected to be the insufficient data at those points. This could mean that the engine was not purely in the steady state at those points. To overcome this problem, longer duration data were collected again from all the points and then again a classifier was built. This time the accuracy was almost 100% and the classifier was performing great. To check the performance of the new classifier over an unknown data set, the earlier data set was applied as a test data set. This gave a classification acuracy of 89.5%. Thus the classification accuracy of the earlier data set was improved from 81% to 89.5%. When this classifier was tested on the raw AVL8 test data set, the classification accuracy increased from 68% to 74%.

Degree

M.S.E.C.E.

Advisors

Meckl, Purdue University.

Subject Area

Electrical engineering

COinS