An energy auditing system based on learnable positional embedding and edge computing for in-situ monitoring of 3D printing processes

Abstract

Additive manufacturing (AM), commonly known as 3D printing, along with its sensing, communicating, and computing networks forms a complex cyber-physical system (CPS), which, however, is also vulnerable to anomalies, performance degradation, or even malicious attacks. This paper presents an energy auditing framework that integrates sensors, transformer-based anomaly detection, and edge computing to enable on-site monitoring of 3D printing processes. It only uses the time series data of energy consumption collected from a side channel for enhanced CPS security. A new learnable positional embedding is proposed to capture temporal features of the time series data and hence, accommodate various anomalous scenarios with a single holistic model. The transformer model is trained in a semi-supervised manner to reconstruct patterns of normal data, and therefore, the magnitude of reconstruction errors can be used as the criterion for detecting anomalies. Experiments are conducted to evaluate the energy auditing system. Anomalies in real-world scenes are mimicked by changing nominal 3D printing parameters, such as printing speed, layer thickness, printing bed temperature, and nozzle temperature. Our energy auditing system can achieve detection accuracy and recall values of approximately 98% and 88%, respectively, especially for detecting early terminations caused by the faster printing speed and larger layer thickness, where other benchmark models could easily fail. Deployment of our system on edge computing for reliable real-time monitoring of the 3D printing process is also demonstrated successfully.