Development of a Convolutional Autoencoder-Based Unsupervised Classification and Visualization Model for Extracting Feature Frequency in C-EPS Fault Diagnosis

In the automotive industry, deep learning models such as convolutional neural networks, autoencoders, and transformers have been actively studied to improve defect detection rates in End-of-Line (EOL) processes. However, applying these approaches to the field of Noise, Vibration, and Harshness (NVH) faces several practical challenges. ① Difficulty in securing sufficient training data and risk of overfitting due to longer evaluation times compared to other data types. ② Label imbalance caused by the relatively small amount of defective data. ③ Reduced labeling accuracy due to human error. ④ Decreased model robustness under domain shifts such as jig variations, evaluation environments, signal-to-noise ratio (SNR), and the occurrence of new abnormal noises. ⑤ Constraints in application due to compatibility issues with existing evaluation equipment. To address these issues, this study proposes an unsupervised learning-based method for classifying operating noise data of Column-type Electric Power Steering (C-EPS) systems using a Convolutional Autoencoder (CAE), along with a visualization technique for feature frequencies corresponding to abnormal noises. This approach minimizes human error and leverages the encoder of the trained model to apply Explainable AI (XAI), thereby effectively extracting and visualizing key feature frequencies that reflect the noise and vibration characteristics of the system. Unlike conventional deep learning models that directly determine abnormality, the proposed method utilizes extracted feature frequency bands as diagnostic evidence, making it independent of overfitting and domain shift issues. Furthermore, the approach enables quantitative data representation and is easily applicable to existing EOL equipment. In addition, post-processing of the extracted feature frequency bands allows for root cause analysis, thereby demonstrating the potential for integration with traditional signal analysis methods.