A hybrid machine learning algorithm approach to predictive maintenance tasks: A comparison with machine learning algorithms

Digital transformation has enabled the industrial sector to digitize all of its production processes, generating a vast quantity of data from sensors and actuators deployed along production lines. This data can provide valuable insights into the behavior of a specific machine, enabling optimization or the prediction of potential malfunctions.Supervised machine learning algorithms are capable of predicting the remaining useful life (RUL) of a machine. However, there is room for improvement in this prediction.In this paper, a hybrid approach that combines supervised learning (multi-layer perceptron MLP) and reinforcement learning (Q-learning) algorithms is proposed with the aim of improving the accuracy and precision of predicting the RUL.To test the efficacy of the proposed hybrid algorithm, a dataset comprising data from C-MAPPS aircraft engines is utilized. To assess the effectiveness of the hybrid algorithm, it was compared with a supervised learning algorithm. The results indicate that the proposed hybrid approach increases accuracy by 15% compared to models that use a single supervised learning algorithm, such as support vector regression (SVR), multi-layer perceptron (MLP), convolutional neural networks (CNN), and long short-term memory (LSTM), and an increase in accuracy of 4% over other hybrid algorithms, such as convolutional neural networks and long short-term memory. Additionally, the hybrid approach enables early predictions of approximately 17 cycles and late predictions of only 2%, while other learning algorithms, such as linear regression, do not make early predictions but 100% of their predictions are late. Implementing a hybrid approach would prevent unexpected machine downtime, enhancing reliability and reducing maintenance times and costs. • A novel method for predicting remaining useful life (RUL) employs a hybrid approach, integrating a supervised learning and a reinforcement learning algorithm. This approach avoids the potential loss or disregard of temporal relationships between samples that may occur in a purely supervised learning process. • The reinforcement learning algorithm's with the environment, states, actions, and rewards for regression tasks allows to increase accuracy in RUL prediction, thereby enabling optimal system or equipment utilization and effective maintenance planning. • A comparison has been made with five algorithms widely used in the field of predictive maintenance, giving the hybrid approach proposed in this work superior qualities which allows having a robust algorithm with a certain capacity for resilience.