Arm motion recognition based on attitude angle sensor

In this paper, the data-driven predictive maintenance technology is applied to the motion recognition of assembly line workers. Specifically, the motion data (acceleration, angular velocity and angle) of workers completing three types of arm movements (two counter-clockwise turns, one turn or one and a half turns, and straight line movement) were collected by the attitude angle measurement sensor. For this data set, an algorithm combining continuous wavelet transform and image deep learning model (pre-trained GoogleNet) has been proposed, and the algorithm is compared with the support vector machine based on the statistical characteristics of discrete wavelet transform and the one-dimensional convolutional network model. It is concluded that the proposed scheme can effectively identify the three kinds of typical movements performed by workers (the average test recognition accuracy is around 93%), which is significantly better than the other comparison models. Through feature analysis, it can be concluded that the scheme combining continuous wavelet transform and image processing deep learning model proposed in this paper has better distinguishable feature space compared with the contrastive models. The scheme proposed in this paper can be extended to predictive maintenance scenarios based on time series data type sensors.