Human gait is systematically deformed by physical loads, this study constructs and evaluates an algorithm for classifying different levels of physical loads. The algorithm uses wearable IMUs data to classify different levels of loads. We aim to evaluate classification as strategy for multi-loads recognition for the control of wearable exoskeletons. 10 adults participated in the experiment. In the experiment, the subjects walked on flat ground carrying a backpack with different weight of loads (0, 15 and 25 kg), and three sensors on the lower limbs collected the subjects' gait data in real time. In this study, a method of classification decision based on multiple bidirectional long short-term memory(multi-BiLSTMs) was proposed which was used to classify the load level of the collected data. The classification accuracy of this method reached 94.1%, and the F-score was 0.935-0.952. Compared with LSTM and BiLSTM, the proposed method has better performance in accuracy of load classification. The results of this study contribute to quantify the load, which has promising applications in the medical and labor protection fields.
Keywords: Physical loads; classification decision; multi-BiLSTMs; wearable IMUs.