Label scarcity- and class imbalance-aware current-based pipeline for concrete vibration activity recognition

Activity recognition of concrete vibration is significant for the monitoring and control of concrete placement. However, existing vision-, sound-, and kinematic-based sensing approaches prove to be inefficient under the severe conditions at concrete vibration sites. Moreover, activity recognition for concrete vibration suffers from two specific data issues: label scarcity and class imbalance. To overcome these obstacles, we considered a novel current-based pipeline for concrete vibration activity recognition. By using electrical current as its sensing approach, the influence of ambient disturbances from construction sites can be mitigated. Initially, the relationship between concrete vibration activities and their corresponding current signatures was established conceptually using energy consumption and power-flow analysis. Subsequently, we formulated this problem as a time-series classification task, and proposed a deep semisupervised model to address the label scarcity issue. Through the integration of both global temporal attention (GTA) and the inception modules, the model's capability to utilize temporal information in current signals was thus strengthened. Furthermore, to tackle class imbalance among concrete vibration activities, an improved data augmentation method tailored to concrete vibration signals was incorporated based on generative adversarial network (GAN). A case study implemented in southwest China validated the feasibility, superiority, and generalizability of the proposed pipeline.