ANS-net: anti-noise Siamese network for bearing fault diagnosis with a few data

Abstract

Fault diagnosis has been limited due to data scarcity. Accordingly, this study focuses on fault diagnosis representation for rolling bearing with few fault data and noisy conditions. Data-driven-based methods have achieved huge success in fault diagnosis, but considerable data are required to maintain promising results and robustness to noises. Therefore, we propose the ANS-net framework in measuring intrinsic difference with a few vibration signals of rolling bearing and build a fault diagnosis model. Few-shot learning tests are employed under a few data and different noisy conditions. The ANS-net is composed of two identical combined networks and a relation layer. The former part extracts feature vectors from a sample pair, and the latter part measures the similarities of the both output vectors. Noisy signals are filtered by the cut-off operation before the input layer and the first convolutional layer with a wide kernel. The scaled exponential linear unit is used in each layer and combined with $$\alpha $$ -dropout layer in self-normalizing layers to map data into a certain distribution to improve the adaptability of the network. Extensive experiments are conducted to validate the proposed method. When tested under a few data, the ANS-net demonstrates better recognition rate than the baselines. When tested under logical noisy conditions, the ANS-net shows better robustness than the baselines. The ANS-net also performs better than the baseline methods with various loads and new categories.