Improved similarity based prognostics method for turbine engine degradation with degradation consistency test

Similarity-based prediction methods have gained increasing attention in data-driven remaining useful life (RUL) technologies, mainly because of their strong generalization capability in multiple industrial scenarios and simple model update processes. The traditional similarity-based prediction methods obtain the referential prediction instances in the referential degradation instance library, which are similar to the degradation trend of the in-service system. The final RUL is the weighted result of each estimated remaining useful life of the referential prediction instances. The key to these methods is the construction of degradation trajectories and the design of similarity matching rules. This study adopts the unsupervised learning of a convolutional neural network autoencoder with an attention mechanism to accurately construct the degradation trajectories of the system. In addition, a new similarity matching rule is proposed to check the degradation consistency of the predicted referential instances, which can eliminate the prediction referential instances that result in marked prediction errors. The experimental results on the turbine engine datasets showed improved prediction performance and lower sensitivity to the size of the training instances. In addition, the proposed method can be easily transplanted to the original similarity-based framework.