02 / Edge AI · Intelligent manufacturing
A side-channel energy-auditing framework for detecting operational anomalies and cyber-physical risks in robotic and additive-manufacturing processes.
Safety and security problem
Networked manufacturing systems can experience process drift, equipment faults, or malicious manipulation. Internal logs alone may be compromised or may not expose subtle physical changes.
Side-channel strategy
An independent energy meter captures power-consumption patterns while semi-supervised models learn normal operation and flag deviations near the machine in real time.
Research presentation / 01
This presentation introduces the 3D-printing testbed, anomalous process settings, learnable positional embedding, and both offline and online validation.
Web-optimized presentation · 2 min 36 sec · 3D printer, energy meter, transformer model, and Jetson edge deployment
System / 02
A PZEM-004T energy meter records the 3D printer's electrical behavior outside the printer-control network.
A Raspberry Pi running OctoPrint and ROS sends only a process-start trigger; monitoring then proceeds independently.
A Jetson AGX Orin reconstructs normal energy patterns and raises an alert when reconstruction error exceeds the learned threshold.
Validated performance / 03
The learnable positional embedding captures both local energy patterns and global process duration, allowing the model to detect early terminations that benchmark methods missed.
Applications / 04
The research advances from robotic-manipulator auditing to real-time monitoring of additive-manufacturing equipment.
GAN and U-TFF models learn normal energy behavior and detect deviations associated with operational anomalies and physical attacks.
A transformer with learnable positional embedding detects changes in bed temperature, printing speed, layer thickness, and process duration.
Method progression / 05
Use normal robotic-manipulator energy traces to train a generative adversarial anomaly detector.
Fuse time- and frequency-domain features through a U-Net architecture for stronger manipulator monitoring.
Encode time instants and sliding-window positions so one transformer can detect local and duration-based 3D-print anomalies.
Publications / 06
Open resource / 07
The public repository contains the energy-auditing dataset and implementation associated with the edge-monitoring study.