Remaining Useful Life Estimation via Cascaded Self-Attention and ResNet Models

Abstract

Prognostics and Health Management occupy an important place in modern industrial maintenance to increase the reliability of systems. Determining the Remaining Useful Life of the system or its parts is vital accurately to maintaining critical parts of the system and successful prognostics and health management. This study proposes a data-based Remaining Useful Life prediction method with a network consisting of a cascade-connected Self-Attention and Residual Network layer. The network is fed by multiple sensor signals to monitor the aero-engines. The proposed model contains four main parts: The Gaussian Noise Layer, the Self-Attention Layer, the Residual Network Layer, and the layer to estimate Remaining Useful Life. The model is created to be more robust and susceptible to noise using the Gaussian Noise Layer. The Self-Attention Layer focuses on crucial points through time. The Residual Network Layer uses feature extraction and makes the model more profound help of the skip connection. Finally, the Remaining Useful Life estimation is made using highly correlated features obtained from the fully connected layer and the output layer. In addition, a new loss function has been offered, similar to the evaluation metrics in the literature. With the proposed model and loss function, 11.017 and 12.629 in root mean square error, 157.19 and 218.6 in score function are obtained in the FD001 and FD003, respectively. The superior performance of these results on the C-MAPSS dataset is demonstrated by comparing the other state-of-the-art methods in the literature.