@techreport{oai:ipsj.ixsq.nii.ac.jp:00232571,
 author = {Jaime, Morales and Qingxin, Xia and Naoya, Yoshimura and Takuya, Maekawa and Hirotomo, Oshima and Masamitsu, Fukuda and Yasuo, Namioka and Jaime, Morales and Qingxin, Xia and Naoya, Yoshimura and Takuya, Maekawa and Hirotomo, Oshima and Masamitsu, Fukuda and Yasuo, Namioka},
 issue = {11},
 month = {Feb},
 note = {When using human activity recognition (HAR) methods in the industrial domain two main problems exist, complex sensor data and lack of labeled data. The complexity of sensor data has been previously addressed by employing motif-guided attention-based models. However, these models provide user-specific solutions that cannot be applied to the general population. This study presents a knowledge transfer method that allows the application of such reliable methods on a previously unknown target user with minimal labeled data. To accomplish this we propose a new technique for selecting appropriate training data segments from an existing pool of known users by identifying multiple levels of similarity between the source and target users, such as work trend similarity, operation-specific motions, user unique key actions, etc. We call this method motif-guided multilevel knowledge transfer. We evaluate the ability of our framework using 3 separate logistic datasets (one real-life, two simulated)., When using human activity recognition (HAR) methods in the industrial domain two main problems exist, complex sensor data and lack of labeled data. The complexity of sensor data has been previously addressed by employing motif-guided attention-based models. However, these models provide user-specific solutions that cannot be applied to the general population. This study presents a knowledge transfer method that allows the application of such reliable methods on a previously unknown target user with minimal labeled data. To accomplish this we propose a new technique for selecting appropriate training data segments from an existing pool of known users by identifying multiple levels of similarity between the source and target users, such as work trend similarity, operation-specific motions, user unique key actions, etc. We call this method motif-guided multilevel knowledge transfer. We evaluate the ability of our framework using 3 separate logistic datasets (one real-life, two simulated).},
 title = {Preliminary investigation on motif-guided multilevel knowledge transfer for logistic work activity recognition},
 year = {2024}
}