DSRJR: A monitoring substitution framework via dual-stream reconstruction and joint representation for fault diagnosis

Abstract

Rotary machinery functions in harsh environments and is susceptible to failure, so dependable fault diagnosis methods are essential to ensure equipment safety. Most existing methods assume that the sensor types of inputs are consistent in the training and testing datasets. However, in practice, sensor failure or the lack of corresponding sensors installed on the equipment can cause the model to fail due to incomplete input information, which ultimately leads to monitoring interruption. Therefore, developing an effective diagnostic knowledge transfer mechanism is essential. To address fault monitoring interruptions caused by the absence of the single sensor, this study proposes a monitoring substitution framework via dual-stream reconstruction and joint representation for fault diagnosis. Dual-stream joint alignment framework (DSJA) designs unique feature extraction models according to the characteristics of each signal, and introduces locality-sensitive latent diffusion space (LSLDS) for feature reconstruction. Building on this foundation, joint representation of feature consistency and domain invariance is developed to effectively map and align feature spaces across different modalities. The monitoring-substitution of the two signals is validated using gradient-based class activation map (Grad-CAM) visualization, and the diagnostic and monitoring-substitution performance of the proposed method is assessed across four scenarios. Compared with existing methods, the proposed approach demonstrates superior fault diagnosis and monitoring-substitution capabilities.