Predictive Analytics Beyond the Hype: A Comprehensive Comparison of LSTM, XGBoost and LightGBM with Emphasis on RMSE and CPU Utilization

Predictive maintenance (PdM) technique leverages data analytics to identify and predict equipment failures. Using data collected from IoT devices such as sensors, PdM applies machine learning techniques, and real-time equipment monitoring to determine exactly when it's the best time to perform equipment maintenance. Anticipating equipment maintenance time before-hand aims to minimize maintenance costs and maximize uptime. Predictive maintenance also aims to predict future adverse events, in order to improvise and plan maintenance schedules in advance. The PdM results are therefore beneficial in reducing downtime and for timely arrangement of essential repair equipment. The maintenance industry was earlier dominated by Preventive and Corrective maintenance practices. The former technique relies on fixed schedule of pre-planned maintenance while the latter requires taking action once a fault occurs. However, in recent times, the increased use of on-board sensors and data-driven algorithms has stimulated the shift to data-driven predictive maintenance approach for aircraft. Leveraging the sensor data, data analytic techniques and Deep Learning models, predictive maintenance approach can calculate the Remaining Useful Lifecycle (RUL) of aircraft engines and can use it to schedule maintenance dynamically. This paper focuses on analyzing the CMAPSS dataset generated by NASA for better understanding of impact created by sensor data. The aim of this paper is to give deeper insight of sensor data, data pre-processing techniques, data featuring, RUL prediction mechanism for maintenance scheduling. The paper presents a comparative analysis of regression algorithms like LSTM, XGBoost and LightGBM. The evaluation criterion is root mean squared error (RMSE) using Regression Analysis and the corresponding CPU utilization. After comparison it is concluded that LSTM outperforms the other deep learning methods for CMAPSS Sensor data.