A Plug-and-Play Sensor System for Machine Fault Diagnosis in Lights-Out Manufacturing

Early awareness of machine faults in lights-out manufacturing allows manufacturers to prevent failures. However, contemporary anomaly detection systems demand significant investments in infrastructure and development. This paper introduces a plug-and-play sensor system to monitor machines in heavy-duty industries. We present a self-adaptive smart sensor system integrating data acquisition, model training, and inference into one system. Featuring a pre-trained feature extractor and a lightweight detector trained on-device after deployment, it offers autonomous machine data analysis to achieve machine fault diagnosis in a plug-and-play manner. We conduct a 25-day experiment with vibration signals and an adaptive anomaly detection algorithm in an actual manufacturing scenario. The evaluation shows that the proposed model occupied only 74KB of disk space and achieved 93.8% accuracy without expert knowledge.