Enhancing Air Compressor Fault Diagnosis: A Comparative Study of GPT-2 and Traditional Machine Learning Models

This study investigates the application of GPT-2, a transformer-based model, for fault diagnosis in reciprocating air compressors, highlighting its ability to capture complex patterns in acoustic signals. Two approaches are compared: the first involves extracting time, frequency, and complexity-based features and classifying them using traditional machine learning models, with a narrow neural network achieving the best performance. The second approach reformulates these features as sequential data for GPT-2, which, through meticulous hyperparameter optimization, delivered superior diagnostic accuracy. Additionally, SHapley Additive exPlanations analysis was employed to enhance model interpretability by identifying the most influential features, providing valuable insights into the fault diagnosis process. While GPT-2 demonstrated notable performance gains over conventional models, it required a more precise hyperparameter tuning. This study offers valuable insights into the application of large language models for classifying damaged mechanical systems.