Audio-based excavator preoperational checks using few-shot deep learning

Routine maintenance of construction machinery, such as excavators, is essential for ensuring operational efficiency and reliability. A key component of this process is the daily preoperational check, which traditionally relies on workers' auditory assessments to identify faults. However, this manual approach is subjective and error-prone, depending heavily on individual experience. To address these limitations, this study proposes an automated fault detection framework using few-shot learning with a single audio sensor - a noncontact and cost-effective diagnostic solution. Recognizing the challenges posed by limited and imbalanced audio data sets, especially due to the impact of faulty machine operations, the proposed model leverages a Siamese neural network (SNN) with transfer learning (TL) to classify faults from raw, noisy audio data. It supports the identification of multiple fault types, including hydraulic, engine, air conditioning, and mechanical issues. A new domain-specific data set of excavator fault sounds was constructed and used to evaluate the approach. Experimental results show that the best-performing configuration - VGG16 with transfer learning and Mel spectrogram features - achieved a test accuracy of 88.08%, precision of 89.86%, and recall of 85.85%, with out-of-sample recall reaching 81.61%. These findings highlight the model's robustness and generalizability in realistic field conditions. The proposed methodology demonstrates strong potential for application in similar industrial environments, contributing to the advancement of intelligent, audio-based fault diagnosis in construction machinery.