Embedded machine learning as an enabler of lean construction

Lean Construction aims to continuously improve construction processes through consistent alignment with customer value. Accordingly, research indicates the necessity of objective and scalable methods for the continuous capture of construction processes. Manual observations enable in-depth contextual analysis but produce discontinuous and sample-based data. Observer-independent measurement approaches for continuous acquisition of time expenditures per construction activity support the provision of scalable data across workers, work packages, shifts, trades, and construction projects.
To address this need, an approach is presented for automated recognition of construction activities using embedded machine learning. In a painting trade case study, a single wrist-worn sensor system classifies main activities at 6 s intervals, achieving an accuracy of 96.4 %. Time-series analysis under open-set site conditions consistently aggregates activity sequences, validated against video-based ground truth. This enables the reconstruction of chronological process sequences and their quantification in terms of time expenditures per activity.
This approach can make production flow observable, support the assessment of performance targets within work packages and takts and the analysis of trade-offs between flow and resource efficiency. Linking activity-based time expenditures with construction outputs may support the derivation of labour consumption rates and thereby contribute to the implementation of Lean Construction in construction management.