An Analysis of the Elevator Failure Prediction System with BigDL Time Series Forecasting Models

Predictive maintenance represents a revolutionary approach in industrial operations, aiming to predict failures before they occur, thus ensuring continuity and efficiency. Elevator systems, crucial for daily operations in numerous buildings worldwide, are particularly suited for this technology. Our research introduces an innovative Elevator Failure Prediction System that leverages Chronos, a component of the BigDL library designed for large-scale time series forecasting. In this paper, we significantly contribute to the enhancement of predictive maintenance strategies, offering a robust framework for anticipating and mitigating potential elevator malfunctions. This study mainly focuses on three sophisticated deep learning models tailored for time series prediction: Sequence-to-Sequence (Seq2Seq) Forecaster, Long Short-Term Memory (LSTM) Forecaster, and Temporal Convolutional Network (TCN) Forecaster. Each model is quantitatively evaluated using Root Mean Square Error metrics across various training epochs, with a consistent lookback period of 60 and 120 time steps. This systematic assessment not only delineates the strengths and limitations of each model in predicting elevator failures but also lays the groundwork for future technological advancements in the elevator industry. Through this experimental analysis, we indicate that the Seq2Seq Forecaster performs well in terms of error rate for a look back period of 60 time steps whereas TCN Forecaster performs well in terms of error rate for a look back period of 120 time steps. However, in both the cases, the LSTM Forecaster performs well in terms of time.