Fig 1.
ACGIH lifting zone system depicting the relative areas collected for analysis.
(Source: NIOSH).
Fig 2.
Plots of the accelerometer and gyroscope sensors for subject 1’s first lift in zone 1 (high risk) from [13].
Not all data collected is shown; the two vertical black lines show the beginning and end of the actual lift.
Table 1.
Mapping of ACGIH lifting zones to relative risk levels.
Fig 3.
Process of ingesting data into the model for training.
All figures not to scale. The resultant matrix could theoretically be any size; a square was selected for highest compatibility with existing CNN research.
Fig 4.
Example of an input image to the network.
The image shown has had a Butterworth bandpass filter of order 2 and bounds 2 and 12 Hz applied to it in addition to a standardizing scaler. The grey block at the bottom represents padding to the model that makes all inputs the same size.
Table 2.
Detailed specification of the layers involved in the proposed model.
All 2D convolution layers contain a ReLU activation layer.
Table 3.
Number of trials for each class of lift.
Fig 5.
Comparison of performance for various hyperparameters set on the proposed model.
A: comparison of RK statistics. B: RK statistic compared with accuracy. Accuracy ranges from 0 to 1 while RK ranges from -1 to 1. Values from -1 to 0 are not shown due to no results in that range.
Fig 6.
Loss gradient descent for α of 0.01 and 0.001.
Each point represents the training categorical cross-entropy at the completion of each epoch.
Table 4.
Summary of classification results for proposed model and alternatives.
Fig 7.
Swarm plot of the testing results by the proposed model.
The x-axis represents the true labeling and the y-axis the model output. The y-axis is divided into three zones that define the resulting class for each value, labeled in the top-left of each box.
Fig 8.
Heatmap plot of the testing results by the proposed model.
Each row has been normalized so that each class has the same color scale.
Table 5.
Confusion matrix of the de-normalized results shown in Fig 8.
The results have been scaled to a single testing set, but all folds were used in these results.
Fig 9.
Saliency plots for final softmax layer of network.
A: low-risk saliency. B: medium-risk saliency. C: high-risk saliency. Bright green/yellow represents the highest weighting; dark purple represents the lowest weighting.
Fig 10.
Saliency plot for high-risk lift trials obtained from the CNN+LSTM model.
Scale ranges from deep blue as the lowest significance and deep red as the highest significance. The x-axis is frames of the input data and y-axis is the sensor data, where A/G is accelerometer or gyroscope and x, y, z are the dimensions for the sensor.