Use of smartphone sensors to quantify the productive cycle elements of hand fallers on industrial cable logging operations

Analysis of time and motion study data is central to forest operations, but current methods used to study work cycles are limited in the breadth and depth of available predictor variables. The objective of this research was to evaluate whether activity recognition modeling based on smartphone sensor data could be used to quantify work tasks during motor-manual logging activities. Three productive cycle elements (travel, acquire, fell) and delays were manually timed while three hand fallers worked on industrial cable logging operations in North Idaho. Each faller carried a smartphone that recorded sensor data at 10 Hz using the AndroSensor mobile app. The random forests machine learning algorithm was used to classify cycle elements and delay from the device sensor measurements. Four time domain features (mean, standard deviation, interquartile range, and skewness) were extracted for each of four sensor values (acceleration, linear acceleration, gyroscope, and sound) using 10 sliding window sizes ranging from 1 to 10 seconds. For each window size, calculations were performed with and without gaps between subsequent cycle elements. Models with and without sound were compared. Overall model prediction accuracy ranged from 65.9% to 99.6% and accuracy increased as window size increased. The two calculation methods did not result in noticeable differences in prediction error, but the inclusion of sound decreased error in nearly all models. These results have demonstrated the feasibility of developing activity recognition models to quantify work based on mobile device sensors, which is an important step for advancing real-time analysis of productive cycle times.