Maintenance costs account for a significant portion of operation costs in manufacturing and production plants. An effective equipment maintenance strategy is crucial for ensuring reliable and safe operations and contributes to profitability by minimizing equipment downtime and reducing overall costs. Predictive maintenance has been a popular strategy for optimizing the operational planning of equipment. In contrast to traditional maintenance strategies, where maintenance actions are taken either at regular intervals or in a reactive manner after faults have occurred, predictive maintenance relies on condition monitoring to predict future equipment health and anticipate equipment failures before they occur. This strategy ensures that maintenance can be planned in advance, thereby curing operational downtime costs and improving process efficiency. This thesis introduces a model-free Reinforcement Learning (RL) based Predictive maintenance system for centrifugal pumps to identify any abnormal operating conditions of the pump and monitor processes to detect pump failures related to seal oil leakage. Based on the historical and real-time dataset, RL-based predictive maintenance systems learn optimal policies over time for identifying abnormal behaviour to plan maintenance, thus cutting operational downtime costs and improving process efficiency. An experiment was also conducted to compare the performance of predictive maintenance systems implemented using the RL-based approach with clustering and novelty detection approaches that employ Machine Learning (ML) and Deep Learning (DL) algorithms. Furthermore, the trained RL agent is applied to a pump with different distributions to check the model’s generalizability. The results show that the RL-based approach was able to create an effective predictive maintenance system with a lesser execution time and fewer training samples. The proposed approach showed the highest efficiency in performance compared to the Clustering and Novelty detection approaches.