LSTM Neural Networks for Remaining Useful Life Estimation of Turbofan Engines

Abstract

Condition-based Maintenance is a maintenance strategy that monitors the actual condition of a system to make predictive decisions whit respect to it. This type of maintenance includes detection, diagnosis, and prediction of system failures. It has become increasingly important because it generates the least losses, reducing total maintenance costs in a business by 5In general, the Remaining Useful Life estimation allows making failure predictions. The complexity of failure prediction in mechanical systems has led to a significant amount of literature. Different solutions have been proposed; however, this still a real problem.Remaining Useful Life estimation can be done from other approaches, for example, using physical models, knowledge-based models, or data-driven models. Extracting relevant features from raw data using physical or knowledge-based techniques alone, in most cases, is not enough due to the complexity of the characteristics present in the data. Literature shows that data-driven approaches are the most used for prediction. In recent years, Deep Learning models for different applications have been used, including failure detection, diagnosis, and prediction. The Deep Learning model’s advantage is that an indepth knowledge of the system is not required, and due to its robustness, complex learning results are satisfactory. For Remaining Useful Life estimation, Long Short Term Memory neural networks are a viable option since they can adequately handle the time series needed for failure predictions using Remaining Useful Life estimation. The three main stages for developing this method based on Long Short Term Memory neural networks were data pre-processing, model training, and model performance evaluation. The methodology uses two datasets of turbofan engines with different operational conditions and faults for its validation. The process evaluates signals obtained from sensors located along with a turbofan engine simulated through a Simulink-based program. This methodology presents a reasonably acceptable performance in terms of Root Mean Squared Error of 2.85 with a standard deviation of 0.39. It means that on average for the engines, the failure prediction will have an error of 3 cycles; and a Score function of 7.26 with a standard deviation of 1.76, which is an asymmetric algorithm where late predictions are more penalized than early predictions, increasing exponentially with the error. The proposed methodology has the advantage of being more straightforward than other methods found in the literature. Besides, the obtained values of the predictions are conservative.