A Novel Deep Parallel Time-Series Relation Network for Fault Diagnosis

Currently, deep learning-based methods are widely used in the fault diagnosis of time-series data for their high precision. However, the application of traditional deep learning fault diagnosis models is limited by their calculational efficiency and poor interpretation ability. To address the problems, a fault diagnosis model named the deep parallel time-series relation network (DPTRN) is proposed in this article. There are three main advantages of the DPTRN. First, our proposed time relationship unit can perform feature extraction on each time node of a time-series sample simultaneously; therefore, the DPTRN performs fault diagnosis in a parallel way and improves the computing efficiency significantly. Second, by improving the absolute position embedding, our novel decoupling position embedding unit can be directly applied for fault diagnosis and can learn contextual information. Third, our proposed DPTRN has an obvious advantage in feature interpretability compared with traditional deep learning-based models. Applying the DPTRN model on four datasets, we achieved higher diagnosis performance with much lower cost, which indicates the effectiveness, efficiency, and interpretability of the proposed DPTRN model.