Sound Engine Based In-Situ Environment Leveraging Neural Network Classification Algorithm

Vehicle fault detection plays a crucial role in enhancing road safety and reducing maintenance costs. Traditionally, fault diagnosis is performed by mechanics based on auditory cues from vehicle components, a process that heavily depends on human expertise and is prone to errors and delays. This study proposes the development and evaluation of a sound-based vehicle fault classification model using a deep learning approach. Mel Frequency Cepstral Coefficients (MFCCs) were employed to extract relevant features from engine audio signals. Two models were implemented: a Multilayer Perceptron (MLP) utilizing hand-crafted features such as Zero Crossing Rate, Spectral Centroid, and MFCCs, and a One-Dimensional Convolutional Neural Network (1D-CNN) trained on log-Mel spectrograms. Research results on the MIMII Fan dataset show that the MLP model achieved a classification accuracy of 95% with a low loss value of 0.1556, outperforming the 1D-CNN model, which reached an accuracy of 81% with a higher loss of 0.4088. The lower loss observed in the MLP model is likely due to its use of compact, manually selected features that allow the model to converge faster and reduce the risk of overfitting. In contrast, the 1D-CNN relies on raw spectrogram data, which introduces greater complexity into the learning process. This work contributes a crucial comparative analysis between these two distinct lightweight strategies, emphasizing their trade-offs in computational efficiency and practical applicability. The developed system offers significant potential for real-time anomaly detection in resource-constrained environments, making it highly applicable in fields such as smart vehicles and automotive Internet of Things (IoT).