A Transformer Approach for Remaining Useful Life Prediction and Fault Diagnosis of Mechanical Equipment

Paper i proceeding, 2026

Accurately estimating remaining useful life (RUL) and performing timely fault diagnosis are critical for ensuring the reliability and safety of industrial equipment. However, these tasks are often treated separately, reducing the effectiveness of maintenance decisions and limiting the use of sensor data. This study extends the Transformer architecture to simultaneously perform RUL prediction and multi-label fault classification within the same framework. Through a shared encoder using self-attention to capture temporal and cross-sensor dependencies, the model learns representations that capture machine health evolution and fault characteristics. Two output heads handle RUL regression and fault classification, while uncertainty-based loss weighting automatically balances the training of both tasks. The model is evaluated on the Microsoft Azure predictive maintenance dataset containing multi-sensor readings from 100 industrial machines. Results show that the proposed approach outperforms machine-learning and deep-learning baselines on both tasks, achieving high accuracy and strong robustness across test machines. By jointly predicting fault types and RUL within a single framework, the proposed method enhances maintenance planning, enables earlier and more reliable interventions, and provides a scalable solution for intelligent monitoring in smart manufacturing environments.