A meta transfer learning-driven few-shot fault diagnosis method for combine harvester gearboxes

Combine harvester gearboxes operate for extended periods under variable operating conditions, making it costly to gather sufficient fault data. A meta transfer learning-driven fault diagnosis method for combine harvester gearboxes is proposed to solve the complex operating conditions and scarce fault samples. The meta learning is employed to train the model so that the performance of the proposed method is not contingent upon the quantity of training data. The multi-step loss optimization (MSL) method is introduced to improve the inner loop and address the unstable update gradients in training. The enhanced method uses each task to refine the model updating strategy, thus circumventing the gradient explosion and decay. The proposed method employs conditional domain adversarial network to extract deep discriminative features from both domains. The batch feature constraint (BFC) is proposed to balance the features’ transferability and class discriminability. A weight-balancing strategy is employed to reconstruct the training loss function, enabling gearbox fault diagnosis under variable operating conditions with few-shot data. The effectiveness of the proposed method is validated through data collected from the combined harvester gearbox’s fault diagnosis experimental rig.