@techreport{oai:ipsj.ixsq.nii.ac.jp:00226074,
 author = {Qingxin, Xia and Takuya, Maekawa and Takahiro, Hara and Hirotomo, Oshima and Yasuo, Namioka and Qingxin, Xia and Takuya, Maekawa and Takahiro, Hara and Hirotomo, Oshima and Yasuo, Namioka},
 issue = {9},
 month = {May},
 note = {Activity recognition using sensor data collected by wearable sensors has been actively studied in the ubicomp community. However, in industrial settings, (1) labeled sensor data collection is costly and (2) the activities performed by workers are more complex than those of daily activities such as walking and running. This study presents a new self-supervised learning approach that effectively utilizes unlabeled data to improve complex activity recognition performance in industrial domain. In this study, we focus on characteristic actions in each work activity (i.e., operation). We first select sensor data motifs corresponding to the characteristic actions that uniquely and consistently occur in an operation. Then, we calculate the similarity series of the motifs that indicate the occurrence of the motifs in every operation. We train a neural network so that it outputs the reconstructed similarity series, which is helpful in identifying the operations containing the motifs. The trained feature extractor is applied to the downstream task that uses limited labeled data to recognize complex work activities. The proposed approach was evaluated on real-world work activity data and achieved state-of-the-art results., Activity recognition using sensor data collected by wearable sensors has been actively studied in the ubicomp community. However, in industrial settings, (1) labeled sensor data collection is costly and (2) the activities performed by workers are more complex than those of daily activities such as walking and running. This study presents a new self-supervised learning approach that effectively utilizes unlabeled data to improve complex activity recognition performance in industrial domain. In this study, we focus on characteristic actions in each work activity (i.e., operation). We first select sensor data motifs corresponding to the characteristic actions that uniquely and consistently occur in an operation. Then, we calculate the similarity series of the motifs that indicate the occurrence of the motifs in every operation. We train a neural network so that it outputs the reconstructed similarity series, which is helpful in identifying the operations containing the motifs. The trained feature extractor is applied to the downstream task that uses limited labeled data to recognize complex work activities. The proposed approach was evaluated on real-world work activity data and achieved state-of-the-art results.},
 title = {Preliminary Investigation of Using SSL for Complex Work Activity Recognition in Industrial Settings},
 year = {2023}
}