Audio-Based Productivity Forecasting of Construction Cyclic Activities

Due to its high cost, project managers must be able to monitor the performance of construction heavy equipment promptly. This cannot be achieved through traditional management techniques, which are based on direct observation or on estimations from historical data. Some manufacturers have started to integrate their proprietary technologies, but construction contractors are unlikely to have a fleet of entirely new and single manufacturer equipment for this to represent a solution. Third party automated approaches include the use of active sensors such as accelerometers and gyroscopes, passive technologies such as computer vision and image processing, and audio signal processing. Hitherto, most studies with these technologies have aimed to activity identification or to identifying active and idle times. Given that most actions performed with construction machinery involve cyclic activities, cycle time estimation is much more relevant. In this study, hardware and software requirements were optimized toward that goal. This approach had three facets: first, signal spectral analysis was performed through the short-time Fourier transform (STFT) and the continuous wavelet transform (CWT) for comparison; second, audio and active sensor data have been submitted to a machine learning framework for activity classification accuracy comparison; and, third, Bayesian statistical models were used to include historical data for cycle time estimation enhancement. As a result, audio signals have been used along with a Markov-chain-based filter to achieve cycle time estimation with an accuracy of over 81% for up to five days of single-machine operation.