Semi-supervised deep learning for recognizing construction activity types from vibration monitoring data

Abstract

Monitoring construction-induced vibrations is crucial for mitigating their adverse impacts on surroundings; hence, contractors have collected a large amount of construction-induced vibration data. Unfortunately, corresponding construction activities were usually not recorded during vibration data acquisition. Consequently, such an unclassified historical vibration database cannot be efficiently utilized to establish empirical models for vibration prediction. Therefore, automatic recognition of the corresponding construction activity types from vibration monitoring data has great potential in practice and research. Previous relevant approaches rely primarily on either manual identification or supervised deep learning algorithms. However, these methods are tedious and require a large amount of labeled data that is usually unavailable in actual scenarios. An advanced approach for the recognition of construction activity types based on semi-supervised deep learning is proposed in this paper to solve this problem. The proposed method comprises a one-dimensional (1D) convolutional neural network (CNN) and a semi-supervised Ladder network. The Ladder-CNN was trained and tested on raw acceleration data measured on real construction sites generated by four different piling operations and one excavation operation. Experiments showed that the Ladder-CNN achieved accuracy up to 98.4% with 10% labeled data and 93.8% with only 1% labeled data. The effects of varying amounts of unlabeled data, different noise levels, and different layer weights were also investigated. Another semi-supervised algorithm (i.e., Pseudo-label) and a supervised 1D CNN model were also tested and compared. The experimental study showed that the Ladder-CNN outperforms the Pseudo-label and is comparable to supervised 1D CNN, demonstrating excellent competitiveness and apparent advantages in practical applications.