A One-Shot Learning Approach for Fault Classification of Bearings via Multi-Autoencoder Reconstruction

In industrial scenarios, the scarcity of labeled data for fault diagnosis of rotating machinery poses a significant challenge to the development of reliable data-driven models. This paper proposes an interpretable one-shot fault diagnosis approach based on class-specific autoencoder, where a separate autoencoder is trained to reconstruct the vibration signal pattern of each fault class. The method operates effectively without requiring the extensive meta-training phases required by few-shot learning. During inference, the reconstruction similarity measured via Pearson correlation is used to identify the most likely fault class. To evaluate the method, experiments were conducted on the Case Western Reserve University (CWRU) dataset under two classification schemes: one with 4 classes representing fault types, and another with 10 classes including fault severities. In both cases, models were trained by using only one sample per class, and performance was averaged over 50-fold cross-validation using different samples in each fold. The proposed model achieved superior performance across all tested operating conditions (1797, 1772, 1750, and 1730 RPM), reaching average accuracies of 64% and 67% in the 10-class and 4-class scenarios, respectively. Comparisons with conventional and few-shot learning methods including Long Short-Term Memory (LSTM), Residual Network (ResNet18-1D), Model-Agnostic Meta-Learning (MAML), and Prototypical Networks. These results establish signal reconstruction with class-specific autoencoder as a simple yet powerful alternative to complex meta-learning strategies for fault diagnosis in real-world, data-scarce, industrial environments.