Elevator condition monitoring : a lightweight receptive field and temporal correlations enhanced deep learning model to tackle imbalanced dataset

Abstract

It has been witnessed with the thriving popularity of elevators in contemporary society during modernization while posing an unavoidable challenge to consistently maintain their safety and reliability in routine operations. The conventional monitoring mechanisms prove insufficient to consistently monitor and evaluate the health condition of elevators in actual deployment. Under a plethora of research utilizing data-driven approaches to measure the degradation status of elevators, there also exist multiple industrial pain points while constraining their broad applications in the elevator industry. For instance, the imbalanced dataset with rare abnormal samples in actual life would lead to incomplete and inadequate representative feature learning during the model training process; the excessive reliance on synthetic laboratory data with proprietary application domains in experiment would impair the model's generalization capability; whilst the complex architecture as stacked in pursuit of functionality comprehensiveness would by no means contribute to effectiveness enhancement at the cost of efficiency decrease. This paper established a lightweight deep learning model with streamlined elegant architecture to resolve the aforementioned issues. The real-life imbalanced data have been collected and utilized directly under the devised algorithmic structure and training strategy without any need for laboratory synthesis on their compositions, yet largely alleviating the impact of imbalanced dataset phenomenon in effect. With the proposed effective feature extraction mechanisms, the receptive field and temporal correlations have been enhanced to further facilitate the monitoring efficiency and accuracy. As validated by the systematic experimental results, the receptive fields have been enlarged by 75% in comparison with the vanilla model structure and the parameter quantity have also been reduced by 21% while boosting the model's training and inferencing efficiency during the actual deployment. Meanwhile, in order to rigorously evaluate the model's robustness, the results have been thoroughly validated after splitting the dataset for multiple times with comprehensive ablation study scenarios to explore the optimal model configuration parameters. The proposed model demonstrated its superior performance by the outstripped average detection precision ratios under asymmetrical data allocations while shedding light on its practical application by industrial practitioners for real-life elevator condition monitoring.