A ResNet-based deep reinforcement learning framework using soft actor-critic for remaining useful life prediction of rolling bearings

Accurately predicting the Remaining Useful Life (RUL) of machinery plays important role for implementing effective predictive maintenance strategies and reducing downtime. However, many existing data-driven approaches rely heavily on supervised learning and treat RUL estimation as a direct regression task from sensor data, lacking the ability to model temporal decision-making or adapt to different domains. To address these limitations, this study proposes a deep reinforcement learning framework that integrates a ResNet-based autoencoder for latent feature extraction from raw vibration signals with the Soft Actor-Critic (SAC) algorithm for dynamic RUL prediction. Unlike conventional methods, our approach allows the model to learn RUL dynamically by interacting with the environment, rather than passively mapping inputs to targets. This interaction enables better adaptability to uncertain degradation patterns. Experimental results on the PHM 2012 dataset demonstrate that the proposed SAC-ResNet framework achieves superior accuracy and generalization performance, highlighting its potential as a promising alternative to traditional RUL estimation models.