"Numerous attempts have been made in recent years for detecting various faults in an automobile
engine. The aim of developing this technique is to detect the faulty engine more accurately and
reduce the cost of risk for an automobile industry. These developed techniques have potential to
detect and classify the type of fault in the engine automatically. Previously developed techniques
are based on either unsupervised learning or supervised learning. Most of the techniques produced
good results but at the cost of complexity. In the proposed work, an algorithm has been developed
for classifying the healthy and faulty engines and detecting the type of fault (if there is any) using
Support Vector Machines(SVM). SVM has less complexity compared to other methods. SVM has
the ability to classify nonlinear data using a kernel function. A kernel function maps a nonlinear
input data into high dimension space so that this mapped data can be classified linearly.
An acoustic signal is obtained, for a duration of 11 seconds, from a faulty or healthy automobile
engine. Acoustic signals have appropriate information in the frequency domain. This
information is useful for fault detection and classification. Total eighteen statistical features of the
audio signature have been extracted in the frequency domain using Hilbert transform and fed to
the classifier. Classification and detection algorithm is formed by sequential channels of SVM. SVM
works on the principle of Structural Risk Minimization(SRM) which has the ability to reduce confidence
interval and empirical risk. Due to SRM, an over-fitting problem in SVM learning is rarely
seen for a finite data set. Moreover, SVM gives good generalization and performance ability compared
to Artificial Neural Network(ANN). Four faults have been tested on a real time-data. Various
combinations of SVM (using different kernel functions), ANN, and Logistic Regression have been
tested and a maximum accuracy up to 98.17% is achieved by proposed method sequential channels
of SVM, which is better than the previous algorithms on the same data set. The novelty of the proposed
algorithm is in its reduced complexity for the computed features, and the ability to produce
better results."