Fault detection from automobile engine sound signals using artificial intelligence methods

Traditional methods for diagnosing automobile faults rely on the expertise of skilled professionals in service centers and specialized diagnostic tools. However, these approaches are often time-consuming and costly for vehicle owners. Consequently, the development of rapid and cost-effective alternative methods for fault detection in automobiles is of significant importance. In this thesis, two efficient and effective alternative methods based on machine learning and deep learning are proposed for the detection of specific faults using the engine sound recordings of automobiles. The proposed methods are presented in two separate studies. Using these methods, six distinct vehicle conditions (five different fault states and a fault-free state) were detected from the sound recordings. Considering the limitations of previous studies in this domain, this thesis utilizes a large and diverse dataset encompassing various vehicle models and real-world operating conditions. Thus, significant improvements in the robustness and generalization capability of the fault detection system have been achieved. In the first study conducted within the scope of this thesis, a vehicle fault detection framework utilizing Mel-Frequency Cepstral Coefficients (MFCC), Discrete Wavelet Transform (DWT)-based features, and an Extreme Learning Machine (ELM) classifier was presented. The experimental results demonstrated that using MFCC-based features of engine sound recordings with the ELM classifier achieved satisfactory performance in automobile fault detection. The performance metrics for this setting were 92.22% precision, 92.22% recall, 92.10% F1-score, and 92.14% accuracy. On the other hand, when DWT-based features were used, the performance was observed to be lower compared to MFCC-based features. Additionally, a frequency analysis was conducted in this part of the thesis to identify the most prominent frequencies of different fault types, providing valuable insights for future fault detection studies. In the second study of this thesis, a deep learning-based model named CarFaultNet was proposed for the prediction of automobile faults using skalogram and spectrogram images of engine sound recordings. By employing two different time-frequency representations of sound recordings, the proposed model gains a multi-view approach. In other words, the model leverages both skalogram and spectrogram information simultaneously to establish a decision-making mechanism for fault detection. Structurally, the CarFaultNet model consists of a parallel combination of two MobileNet convolutional neural network architectures. For comparison purposes, this study also includes the application of other convolutional neural network-based models, such as AlexNet, GoogLeNet, ShuffleNet, and SqueezeNet, for fault detection. The results indicate that the CarFaultNet model outperforms both traditional machine learning methods and single-view deep learning models. The performance metrics of the proposed model were 95.32% accuracy, 94.83% recall, 94.99% F1-score, and 95.00% precision. Furthermore, the Class Activation Mapping (CAM) technique was employed to highlight the regions of skalogram and spektrogram images that significantly influence the decision-making process of the CarFaultNet model. Considering the approaches proposed and the results obtained in this thesis, it is anticipated that machine learning and deep learning-based methods could provide reliable and practical solutions for automobile fault detection in the future.