Due to the long-term normal state of electromechanical equipment, the failure time is short, gradually forming the problem of small sample of faults. For this reason, researchers have widely used simulated faults to carry out fault diagnosis research on electromechanical equipment. However, there is still a difference between simulated faults and actual faults, and how to solve this problem has become an industry challenge. To address this problem, this paper proposes a network model based on spatio-temporal signals fusion, and builds a model architecture by combining KNN, GCN and global attention mechanism. The model demonstrates excellent classification ability when tested on public datasets. Furthermore, this paper proposes a wireless sensing system, which is innovatively deployed in industrial production to record the working status of electromechanical equipment at all times. It provides a new solution for the small sample problem which strongly contributes to the development of the fault diagnosis field.