Working-Condition-Indexed Generative Domain Generalization for Intelligent Fault Diagnosis Under Unseen Conditions

Domain generalization techniques have gained widespread application in cross-domain fault diagnosis in recent years. These methods enhance model generalization capabilities without accessing target domain data, thereby improving fault diagnosis performance under unknown operating conditions. However, most existing approaches overlook the fundamental relationships between operating conditions across different domains, resulting in directionless domain generalization. To address this limitation, we propose a working-condition-indexed generative domain generalization (WCIGDG) method for fault diagnosis in data-free conditions using given working-condition parameters. Within this framework, a domain index is employed to model relationships between operating conditions across domains. A domain index predictor is constructed to guide the data generation model, enabling the generated data to better adapt to the target operating condition, while adversarial training is introduced to enhance the effectiveness of data generation. Subsequently, the generated data is combined with source domain data to train the fault diagnosis model, strengthening its adaptability to the target operating condition. Experiments are conducted on two publicly available rotating machinery fault diagnosis datasets, namely the CWRU bearing dataset and the WT planetary gearbox dataset, using rotational speed as the domain index across 68 cross-speed domain generalization tasks. The results demonstrate the effectiveness of WCIGDG for both bearing and gear fault diagnosis. Compared with SDGN, the best-performing baseline among the compared methods, WCIGDG improves the average accuracy/F1-Score by 1.22/2.72 percentage points on CWRU and 1.33/1.93 percentage points on WT.