Transformer-Enabled Cross-Domain Diagnostics for Complex Rotating Machinery With Multiple Sensors

Abstract

Unsupervised cross-domain fault diagnosis for rotating machinery is of great practical significance for real-world industrial scenarios; however, most existing methods are developed based on the vibration signal from a single sensor. With the increasing complexity of industrial systems, multisensor collaborative monitoring has played an important role in the health management of complex machinery. However, existing methods cannot provide effective solutions for multisensor cross-domain diagnosis. To tackle this issue, this article proposes a novel transformer-enabled multisensor collaborative cross-domain intelligent diagnosis framework, which can achieve the transformation and fusion of diagnostic knowledge among multiple sensors. More specifically, a transferable transformer network is built to extract discriminative features and conduct fine-grained domain-invariant learning for the single sensor data. Then, a novel weighted voting strategy based on the transferability and the source accuracy rate is proposed to achieve more accurate decision fusion results. Furthermore, an unsupervised prediction consistency loss is established to further improve the prediction performance of the model for all target sensors. We design ten cross-domain fault diagnosis tasks in three industrial scenarios. The experimental results show that the proposed method has significant benefits for the health management of complex rotating machinery.