Audio-Based Engine Fault Diagnosis with Wavelet, Markov Blanket, ROCKET, and Optimized Machine Learning Classifiers

Area of Science:

• Automotive Engineering
• Signal Processing
• Machine Learning

Background:

• Engine fault diagnosis is crucial for automotive aftermarket management.
• Creating labeled datasets is difficult due to signal variations and feature distribution divergence.
• Nonlinearity and divergence in engine data complicate accurate fault identification.

Purpose of the Study:

• To develop a robust hybrid approach for classifying engine fault conditions using audio emission signals.
• To address the challenges of nonlinearity and feature distribution in engine fault diagnosis.
• To enhance decision-making processes in the automotive industry through improved fault classification.

Main Methods:

• Experimental measurement of audio emission signals from compression ignition engines under simulated fault conditions (injector failure, intake hose failure, no failure).
• Application of Wavelet Packet Transform (WPT) for signal decomposition into sub-time series.
• Utilizing Markov blanket feature selection, Random Convolutional Kernel Transform (ROCKET), and Tree-structured Parzen Estimator (TPE) for hyperparameter tuning with ten machine learning classifiers.
• Integration of WPT, feature selection, ROCKET, and TPE-optimized ML classifiers for a hybrid diagnostic system.

Main Results:

• The hybrid approach successfully classified different engine fault conditions based on audio emissions.
• Wavelet Packet Transform effectively processed audio data into informative frequency and resolution sub-time series.
• Markov blanket feature selection identified crucial features, enhancing classification accuracy.
• The ROCKET method, combined with TPE-tuned ML classifiers, demonstrated superior generalization performance compared to standard methods.

Conclusions:

• The proposed hybrid approach offers a powerful and effective solution for engine fault diagnosis using audio signals.
• This method overcomes challenges related to signal nonlinearity and feature distribution in automotive applications.
• The findings support improved planning and decision-making in the automotive industry through reliable engine condition classification.