Digital Twin-Assisted Online Monitoring of Elevator Abnormal Vibration Using Improved ShuffleNetV2 and Domain Adaptation

Abstract

Elevator's vibration is closely linked to safe operation, which is crucial for social public safety. In order to achieve the online detection of the horizontal abnormal vibration and solve the problem of scarce labelled abnormal data for training intelligent models, a digital twin-assisted method using an improved lightweight ShuffleNetV2 triplet attention network (STAN) and deep subdomain adaptation network (DSAN) is proposed. According to the actual working conditions of elevator operation, a digital twin system was constructed, enabling the virtual-to-real mapping of the elevator operation states. The simulated data of abnormal horizontal vibration was augmented. The one-dimensional simulated data and measured data were converted into two-dimensional time-frequency map data by continuous wavelet transform (CWT). Feature extraction was performed by the improved lightweight STAN that incorporated grouped mixing blocks and attention mechanism. A DSAN method based on Gaussian kernel function was used for abnormal vibration detection by data fusion and feature alignment. Experiment results showed that the detection model had a size of 1.258 MB and required 0.152 G of floating-point operations (FLOPs), and the response time of the visualization platform was 0.11 s on average. The abnormal vibrations can be classified accurately during elevator operation. The digital twin-assisted online monitoring method combined with an improved lightweight STAN can be easily deployed in edge end to realize the online monitoring elevators with visualization and virtuality to reality interaction.