Sound-Based Intelligent Detection of FOD in the Final Assembly of Rocket Tanks

The traditional method of relying on human hearing to detect foreign object debris (FOD) events during rocket tank assembly processes has the limitation of strong reliance on humans and difficulty in establishing objective detection records. This can lead to undetected FOD entering the engine with the fuel and causing major launch accidents. In this study, we developed an automatic, intelligent FOD detection system for rocket tanks based on sound signals to overcome the drawbacks of manual detection, enabling us to take action to prevent accidents in advance. First, we used log-Mel transformation to reduce the high sampling rate of the sound signal. Furthermore, we proposed a multiscale convolution and temporal convolutional network (MS-CTCN) to overcome the challenges of multi-scale temporal feature extraction to detect suspicious FOD events. Finally, we used the proposed post-processing strategies of label smoothing and threshold discrimination to refine the results of FOD event detection and ultimately determine the presence of FOD. The proposed method was validated through FOD experiments. The results showed that the method had an accuracy rate of 99.16% in detecting FOD and had a better potential to prevent accidents compared to the baseline method.