Comparative Study of MLP and KAN Models for Fault Detection in Automotive Engines

Abstract

Engines are critical components in a wide range of industrial and trans- portation systems. Over time, they are subject to wear and faults that can lead to performance degradation or complete failure if not detected early. Traditional diagnostic methods often rely on physical inspections and spe- cialized sensors, which can be costly, time-consuming, and require expert intervention. In recent years, the integration of machine learning techniques with audio signal processing has opened new possibilities for non-invasive and efficient fault detection. This dissertation focuses on the development of a machine learning-based system for diagnosing engine faults using audio recordings. The approach involves extracting key features from engine sound signals, including Mel- Frequency Cepstral Coefficients (MFCC), Log-Mel Spectrograms, Wavelet Transforms, and Zero-Crossing Rate (ZCR). These features serve as inputs to classification models such as the Multilayer Perceptron (MLP) and the Kolmogorov–Arnold Network (KAN), which are designed to distinguish be- tween normal and abnormal engine behavior. The objective is to build an intelligent diagnostic system capable of an- alyzing engine sounds and supporting predictive maintenance efforts. This work contributes to the growing field of AI-based fault detection and proposes a framework that can be extended to various types of engines and integrated into real-time monitoring solutions.