A dual-stream deep learning framework for lift condition monitoring based on time and frequency domain information

Abstract

This research addresses lift condition monitoring, crucial for ensuring safety and performance. Traditional methods struggle with scalability, especially when dealing with small data volumes or potential fault detection. To overcome this, we propose a deep learning-based approach that integrates time-domain and frequency-domain data to enhance fault detection accuracy, particularly in data-limited scenarios. Our model employs a dual-stream convolutional neural network (CNN) to transform a single vibration signal into inputs for both the time domain and frequency domain. Dynamic modeling of the correlations between time-domain and frequency-domain inputs is performed using a self-attention mechanism, and a multi-layer perceptron is employed for output classification to accomplish fault detection. Experimental results on two datasets (CWRU bearing and lift door vibration data) show that our model outperforms baseline methods, especially in data-limited contexts, significantly improving safety by detecting early faults, such as dust accumulation on lift doors. This study contributes a scalable, robust solution for lift fault detection, improving accuracy in small datasets and providing a foundation for future advancements, including integrating additional sensor data and exploring advanced deep learning models.