Enformer: A Lightweight Encoder-Based Network for Fault Diagnosis in Rotating Machinery Using Probability Sparse Self-Attention

Fault diagnosis plays a crucial role in the routine maintenance of modern industrial rotating machinery. In this context, the high computational complexity of attention mechanisms in Transformer-based models and the inherent limitations of convolutional neural network (CNN) methods in capturing global dependencies within time series data present significant challenges. To address these issues, we introduce Enformer, a lightweight network tailored for fault diagnosis in rotating machinery. The proposed model integrates numerical, temporal, and positional features from time series signals, while leveraging the Probability Sparse Self-Attention (PSSA) mechanism to effectively mitigate computational complexity. Moreover, the model's encoder and distillation layers progressively extract fine-grained features and capture global dependencies, thereby enabling precise fault detection. Experimental results demonstrate that Enformer achieves exceptional diagnostic accuracy across various time-step configurations, with a peak accuracy of 98.33%, significantly outperforming traditional methods. Additionally, the lightweight architecture of Enformer ensures robust adaptability and resource efficiency, positioning it as a promising approach for accurate and efficient fault diagnosis in industrial applications.