Segmenting and Classifying Repetitive Construction Process Time Series Using Small Amount of Labeled Data

Repetitive construction processes, as an essential element of construction industry, still rely intensively on manual execution and on-site decision-making. Within the proposal for integrating Cyber-Physical-System (CPS) in construction, time series analysis of sensor data has great potential to enhance construction project efficiency and support decision-making. However, owing to variable boundary conditions among construction projects, acquiring segmented and labeled training data for time series analysis models requires extensive human effort at the early stages of each construction project, with limited data reusability. We propose a Dynamic Time Warping-based (DTW) ensemble model for segmenting and assigning labels, which are predefined by experts as reference labels, for repetitive construction process through classification, requiring only small amount of labeled training data. The model is validated through a case study involving the Kelly Drilling process in two construction projects, achieving an average accuracy close to 90%. Minor errors occur only at subprocess transition points, in accordance with the error pattern in manual segmentation and labeling efforts. The proposed model addresses the challenge of the large human effort in acquisition of sufficient labeled segmented data in CPS in context construction under flexibility requirements.