Implementation of Self-Attention-based 2D CNN Models for Enhancing Remaining Useful Life Predictions Using Multivariate Time Series Data

Abstract

Purpose: This research aims to introduce a novel a methodology for predicting the Remaining Useful Life (RUL) using multivariate time series data.BRMethods: The proposed RUL prediction methodology comprises of the following steps: 1) Reorganizing the multivariate time series data to enhance the correlation between different time series datasets; 2) Streamlining various time series data into a single pixel utilizing 2D convolutional layers; 3) Emphasizing the substantial correlation among different time series using a self-attention layer; 4) Estimating the RUL with Bi-LSTM and fully connected layers.BRResults: In comparison with existing deep learning models utilizing the identical test datasets, the proposed model exhibits greater performance in RUL prediction. A detailed analysis reveals the model’s merits in terms of data reorganization alongside the application of 2D CNN and multi-head self attention layers in the RUL prediction.BRConclusion: The proposed model provides more accurate RUL estimation results relative to pre-existing models using multivariate datasets obtained from multiple sensors, showing promising potential for its use in real-world applications.