Wasserstein distance based multi-scale adversarial domain adaptation method for remaining useful life prediction

Accurate remaining useful life (RUL) prediction can formulate timely maintenance strategies for mechanical equipment and reduce the costs of industrial production and maintenance. Although data-driven methods represented by deep learning have been successfully implemented for mechanical equipment RUL prediction, existing methods generally require test data to have a similar distribution to the training data. Due to the domain shift problem caused by the changes in equipment operating conditions and fault types, the previously trained models lose the accuracy of prediction under the new conditions. In response to this problem, we combined a deep learning model with domain adaptation and proposed a Wasserstein distance based multi-scale adversarial domain adaptation (WD-MSADA) method for complex machinery RUL prediction. The proposed WD-MSADA utilizes a discriminator to compute the Wasserstein distance to guide the adversarial domain adaptation process more stably to reduce the distribution differences between the source and target domains. Additionally, a multi-scale convolutional neural network (MSCNN) is proposed as a feature extractor to learn the common multi-scale features between two domains, improving the shortcomings of traditional CNNs and enhancing the domain adaptation capability. Experiments on the RUL prediction of turbofan engines in 12 cross-domain scenarios demonstrate that the proposed WD-MSADA performs reliable RUL prediction in unlabeled target domains, and the prediction results are compared with models without domain adaptation, other advanced domain adaptation methods, and the model without multi-scale, demonstrating the superiority and reliability of the proposed method.

Automated summary

This article introduces a Wasserstein distance based multi-scale adversarial domain adaptation method for accurate remaining useful life (RUL) prediction of complex machinery. Experimental results demonstrate the reliable prediction capability of the proposed method in unlabeled target domains, with better performance compared to other methods and models.