FAULT DIAGNOSIS OF GEARBOX UNDER VARIABLE WORKING CONDITION BASED ON WEIGHTED SUBDOMAIN ADAPTIVE ADVERSARIAL NETWORK

In practical engineering, gearboxes are subject to complex and variable operating environments, which hinder the ability of a single vibration signal to accurately and effectively represent fault information under different working conditions. To address this issue, a gearbox fault diagnosis method for variable working conditions based on weighted subdomain adaptive adversarial networks was proposed. Initially, a multi-source heterogeneous signal fusion strategy was employed to transform vibration signal spectrograms, current signal Gramian matrices, and infrared thermograms into a multi-channel dataset, offering diverse perspectives on gearbox operational states. Subsequently, a self-calibrated convolutions network (SCNet) incorporating an efficient channel attention (ECA) mechanism acted as a feature extractor, dynamically adjusting the interactions and dependencies between multi-source heterogeneous signals to balance the scale differences between the source and target domain heterogeneous data. Concurrently, during adversarial training of the feature extractor and domain discriminator, maximum mean discrepancy (MMD) and linear discriminant analysis (LDA) were introduced to measure the domain alignment degree of the current cross-domain task feature representation and the diagnostic task decision boundary. A dynamic balancing factor was constructed to real-time adjust domain alignment loss and class discriminability loss, effectively aligning each class space between the source and target domains. Finally, validated by a collected gearbox fault dataset under variable operating conditions. The results show that the proposed method achieves diagnostic accuracy exceeding 95% across different conditions, demonstrating its feasibility and effectiveness.