Enhancing Bearing Fault Diagnosis in Real Damages: A Hybrid Multidomain Generalization Network for Feature Comparison

Recently, domain adaptation methods in the field of fault diagnosis for rotating machinery have gained increasing popularity. However, these methods often require a substantial amount of labeled prior data distribution. In practical diagnostic scenarios, acquiring vast amounts of authentic data proves to be challenging. Additionally, a notable disparity exists between datasets generated through artificially induced damages and those produced from actual real-world damages. Consequently, developing models and applying domain adaptation methods to industrial data with real damages poses a challenging task. Acknowledging these challenges and aiming to maximize the transfer model’s generalization capability, this article proposes a novel mixed domain fusion network (MDFNet) model for fault diagnosis in rotating machinery under unknown real operating conditions. The core innovation of this model lies in capturing invariant representations of different environmental information. Simultaneously, through feature discrimination, the inner product of similar feature representations is averaged higher than other features, enabling the diagnostic model to learn robust information specific to certain working environments and generalize to unseen working environments. Experimental results conducted on the German Paderborn rolling bearing dataset validate the effectiveness of this approach. Furthermore, additional experiments conducted on a planetary parallel-axis gearbox bearing comprehensive fault simulation test bench, involving a substantial amount of diagnostic tasks on collected real-world data, further confirm the improved generalization performance of the proposed method under unknown real operating conditions.