End-to-end discriminative representation learning for fault diagnosis in safety-critical time series

Dong, F.*; Xiao, Y.*; Chen, S.*; Demachi, Kazuyuki*; Takaya, Shigeru  ; Yoshikawa, Masanori 

Ensuring safe and stable industrial plant operations requires accurate and timely fault diagnosis from multivariate time-series (MTS) sensor data. Conventional methods struggle with the complexity, high dimensionality, and limited feature extraction. To address this, we propose a novel end-to-end fault diagnosis framework that enhances class separability by leveraging instance-wise global and timestamp-wise local views of MTS representations. The dataset is augmented from the dual-view, and a complementary contrastive loss function captures both global and local contextual information. Unlike previous representation learning approaches, the diagnosis model's backbone and classifier are jointly optimized in an end-to-end scheme, ensuring aligned representations with the classification objective. Experimental results on simulated nuclear power plant fault datasets demonstrate the framework's effectiveness and robustness.