Elevator Breakdown Prediction Using LSTM Analysis of Monitoring Data

Elevator malfunctions represent significant challenges in modern building infrastructures due to their resulting inconvenience and downtime, impacting transportation quality and reliability. To address these issues, the current study proposes to utilize Long Short-Term Memory (LSTM) networks to predict safety-related breakdowns that lead to immediate elevator stops. The methodology combines monitoring data analysis with LSTM networks to forecast elevator safety chain breakdowns. In response to the often-encountered problem of unbalanced datasets, a real balanced dataset is constructed using monitoring data from operating elevators worldwide. This dataset is formed using language modeling techniques, identifying behavioral patterns across various time horizons, and mapping these to a classification problem. Further steps involve setting temporal boundaries, embedding data within these boundaries, deploying an LSTM neural network, and subsequently fine-tuning hyperparameters. Experiment results indicate that an LSTM neural network can predict elevator safety chain breakdowns with an F1-score of 85 % on average across multiple time windows.