Construction activity classification based on vibration monitoring data: A supervised deep-learning approach with time series randaugment

Although vibration monitoring systems have been widely implemented on construction sites, most monitoring data cannot be efficiently used to establish an empirical vibration model because the information of the corresponding construction activities is usually not recorded. Identifying various construction activities from collected vibration data will bring new and unexpected benefits in practical applications. This study aims to fill this knowledge gap by proposing an accurate and efficient construction activity recognition model that combines the deep learning network [i.e., convolutional neural network (CNN)] and state-of-the-art RandAugment algorithm. The optimal number and strength of transformations in RandAugment were obtained through a parametric study. Vibration monitoring data sets, which were collected on various construction sites and generated by five different construction activities, were employed in performance validations. Results show that a well-trained CNN with RandAugment can classify construction activities with extremely high accuracy of 99.21%. Although RandAugment also improves the performance of another machine learning network [i.e., multilayer perceptron (MLP)], the CNN model still outperforms the MLP model in terms of classification accuracy. The proposed CNN with time-series RandAugment provides an accurate and promising tool to classify a tremendous amount of historical construction vibration data, thereby enabling the establishment of an informative database for future research.