SiamSpecNet: One-Shot Bearing Fault Diagnosis Using Siamese Networks and Gabor Spectrograms

Abstract

Bearing faults are a leading cause of failure in mechanical systems, underscoring the need for efficient and reliable diagnostic methods. Although deep learning has shown promise in this domain, its dependence on large volumes of labeled, machine-specific data limits its industrial applicability. This study introduces SiamSpecNet, a one-shot learning framework that integrates Gabor spectrograms with a lightweight convolutional Siamese Neural Network (SNN) to address data scarcity in bearing fault diagnosis. The twin subnetworks of the SNN employ a previously developed CNN architecture as a precursory feature extractor. Raw vibration signals are converted into two-dimensional time–frequency spectrograms, and a comparative analysis of three transformation techniques—Short-Time Fourier Transform (STFT), Morlet wavelet, and Gabor transform—demonstrates that Gabor spectrograms achieve superior classification precision and faster processing. Evaluations on datasets from a laboratory test rig, the MFPT database, and real industrial equipment show that with only 11–13 training samples per class, the model achieves 97–100% in-domain accuracy and strong cross-domain generalization, with cross-domain accuracies ranging from 85.96% to 100%, and MFPT being the most challenging (85.96%).These results demonstrate that SiamSpecNet—combining one-shot learning, optimized spectrogram generation, and a robust CNN backbone—offers a scalable and data-efficient solution for bearing fault diagnosis in industrial environments.