Fig 1.
Outline of the general activity recognition model development process.
Steps involved typically include (1) collecting time study data to pair with wearable sensor measurements, (2) preprocessing the data through filtering, (3) extracting time and/or frequency domain features using a sliding window and then selecting relevant features with which to build models, and (4) developing activity recognition models using machine learning or deep learning techniques. Ultimately, models may be programmed into apps on smartphones and smartwatches and subsequently used to characterize work activities in real-time to inform health and safety notifications.
Table 1.
Summary of productive cycle elements for choker setter and chaser work activities.
Fig 2.
Overview of a hypothetical choker setter activity recognition model running on a smartwatch.
The activity recognition model depicted is using a 5-s window with 50% overlap to predict the four work activities. The figure shows filtered acceleration magnitude data, which is colored according to the actual work cycles. Each time a window (shown as rectangles with dashed lines) is used to extract features, the model predicts the work cycle (shown as labels above the windows).
Table 2.
Summary statistics (in seconds) of cycle times for choker setter and chaser work activities.
Fig 3.
Biplots of PCs 1 and 2 for the choker setter and chaser datasets.
The color of points on each plot indicates work cycle element categories.
Table 3.
Summary of choker setter and chaser PCA results.
Fig 4.
Choker setter and chaser random forest model accuracy as a function of the number of trees.
The plots are grouped by worker type (choker setter or chaser) and window size. Line color indicates overall model (OOB) accuracy as well as accuracy for the work cycle elements. Only the 90% overlap of the 1-, 5-, 10- and 15-s windows are shown.
Fig 5.
Choker setter sensitivity, specificity, and precision as a function of window size.
The plots are grouped by metric and work activity. Line color indicates window overlap level.
Fig 6.
Chaser sensitivity, specificity, and precision as a function of window size.
The plots are grouped by metric and work activity. Line color indicates window overlap level.
Fig 7.
Choker setter and chaser multiclass AUC as a function of window size.
The plots are grouped by worker type (choker setter or chaser). Line color indicates window overlap level.
Table 4.
Accuracy metrics for the best choker setter model (created with a 3-s window and 90% overlap).
Table 5.
Confusion matrix for the best choker setter model (created with a 3-s window and 90% overlap).
Table 6.
Accuracy metrics for the best chaser model (created with a 1-s window and 90% overlap).
Table 7.
Confusion matrix for the best chaser model (created with a 1-s window and 90% overlap).